diff --git a/.github/ci/libcxx17.imp b/.github/ci/libcxx17.imp
index 7bdcf5bc..d3a262b5 100644
--- a/.github/ci/libcxx17.imp
+++ b/.github/ci/libcxx17.imp
@@ -7,6 +7,7 @@
     { include: [ "<__fwd/sstream.h>", private, "<iosfwd>", public ] },
     { include: [ "<__fwd/streambuf.h>", private, "<iosfwd>", public ] },
     { include: [ "<__fwd/string_view.h>", private, "<string_view>", public ] },
+    { include: [ "<__system_error/errc.h>", private, "<system_error>", public ] },
 
     # Mappings for includes between public headers
     { include: [ "<ios>", public, "<iostream>", public ] },
diff --git a/.github/workflows/arm_compilation.yml b/.github/workflows/arm_compilation.yml
index 3934ac2d..5bf2a93e 100644
--- a/.github/workflows/arm_compilation.yml
+++ b/.github/workflows/arm_compilation.yml
@@ -26,6 +26,7 @@ jobs:
       - uses: actions/checkout@v4
         with:
           fetch-depth: 0
+          persist-credentials: false
 
       - name: Download required linux packages
         if: runner.os == 'Linux'
@@ -91,4 +92,7 @@ jobs:
         uses: actions/upload-artifact@v4
         with:
           name: ${{ matrix.config.simple_name }} ${{ matrix.binaries }}
-          path: .
+          path: |
+            .
+            !.git
+            !.output
diff --git a/.github/workflows/clang-format.yml b/.github/workflows/clang-format.yml
index e20e0d5d..452c2f2a 100644
--- a/.github/workflows/clang-format.yml
+++ b/.github/workflows/clang-format.yml
@@ -11,6 +11,10 @@ on:
     paths:
       - "**.cpp"
       - "**.h"
+
+permissions:
+  pull-requests: write
+
 jobs:
   Clang-Format:
     name: Clang-Format
@@ -25,7 +29,7 @@ jobs:
         id: clang-format
         continue-on-error: true
         with:
-          clang-format-version: "17"
+          clang-format-version: "18"
           exclude-regex: "incbin"
 
       - name: Comment on PR
@@ -33,12 +37,13 @@ jobs:
         uses: thollander/actions-comment-pull-request@fabd468d3a1a0b97feee5f6b9e499eab0dd903f6 # @v2.5.0
         with:
           message: |
-            clang-format 17 needs to be run on this PR.
+            clang-format 18 needs to be run on this PR.
             If you do not have clang-format installed, the maintainer will run it when merging.
-            For the exact version please see https://packages.ubuntu.com/mantic/clang-format-17.
+            For the exact version please see https://packages.ubuntu.com/noble/clang-format-18.
 
             _(execution **${{ github.run_id }}** / attempt **${{ github.run_attempt }}**)_
           comment_tag: execution
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Comment on PR
         if: steps.clang-format.outcome != 'failure'
@@ -49,3 +54,4 @@ jobs:
           create_if_not_exists: false
           comment_tag: execution
           mode: delete
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
diff --git a/.github/workflows/codeql.yml b/.github/workflows/codeql.yml
index d949a5a7..d01ed41f 100644
--- a/.github/workflows/codeql.yml
+++ b/.github/workflows/codeql.yml
@@ -30,6 +30,8 @@ jobs:
     steps:
       - name: Checkout repository
         uses: actions/checkout@v4
+        with:
+          persist-credentials: false
 
       # Initializes the CodeQL tools for scanning.
       - name: Initialize CodeQL
diff --git a/.github/workflows/compilation.yml b/.github/workflows/compilation.yml
index 3524d5e9..5878adec 100644
--- a/.github/workflows/compilation.yml
+++ b/.github/workflows/compilation.yml
@@ -25,6 +25,8 @@ jobs:
         shell: ${{ matrix.config.shell }}
     steps:
       - uses: actions/checkout@v4
+        with:
+          persist-credentials: false
 
       - name: Install fixed GCC on Linux
         if: runner.os == 'Linux'
@@ -86,4 +88,7 @@ jobs:
         uses: actions/upload-artifact@v4
         with:
           name: ${{ matrix.config.simple_name }} ${{ matrix.binaries }}
-          path: .
+          path: |
+             .
+             !.git
+             !.output
diff --git a/.github/workflows/games.yml b/.github/workflows/games.yml
new file mode 100644
index 00000000..f0bca442
--- /dev/null
+++ b/.github/workflows/games.yml
@@ -0,0 +1,43 @@
+# This workflow will play games with a debug enabled SF using the PR
+
+name: Games
+on:
+  workflow_call:
+jobs:
+  Matetrack:
+    name: Games
+    runs-on: ubuntu-22.04
+    steps:
+      - name: Checkout SF repo 
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event.pull_request.head.sha }}
+          path: Stockfish
+          persist-credentials: false
+
+      - name: build debug enabled version of SF
+        working-directory: Stockfish/src
+        run: make -j build debug=yes
+
+      - name: Checkout fast-chess repo
+        uses: actions/checkout@v4
+        with:
+          repository: Disservin/fast-chess
+          path: fast-chess
+          ref: d54af1910d5479c669dc731f1f54f9108a251951
+          persist-credentials: false
+
+      - name: fast-chess build
+        working-directory: fast-chess
+        run: make -j
+
+      - name: Run games
+        working-directory: fast-chess
+        run: |
+          ./fast-chess  -rounds 4 -games 2 -repeat -concurrency 4 -openings file=app/tests/data/openings.epd format=epd order=random -srand $RANDOM\
+               -engine name=sf1 cmd=/home/runner/work/Stockfish/Stockfish/Stockfish/src/stockfish\
+               -engine name=sf2 cmd=/home/runner/work/Stockfish/Stockfish/Stockfish/src/stockfish\
+               -ratinginterval 1 -report penta=true -each proto=uci tc=4+0.04 -log file=fast.log | tee fast.out
+          cat fast.log
+          ! grep "Assertion" fast.log > /dev/null
+          ! grep "disconnect" fast.out > /dev/null
diff --git a/.github/workflows/iwyu.yml b/.github/workflows/iwyu.yml
index 0552a598..f8898b1c 100644
--- a/.github/workflows/iwyu.yml
+++ b/.github/workflows/iwyu.yml
@@ -14,6 +14,7 @@ jobs:
         uses: actions/checkout@v4
         with:
           path: Stockfish
+          persist-credentials: false
 
       - name: Checkout include-what-you-use
         uses: actions/checkout@v4
@@ -21,6 +22,7 @@ jobs:
           repository: include-what-you-use/include-what-you-use
           ref: f25caa280dc3277c4086ec345ad279a2463fea0f
           path: include-what-you-use
+          persist-credentials: false
 
       - name: Download required linux packages
         run: |
diff --git a/.github/workflows/matetrack.yml b/.github/workflows/matetrack.yml
new file mode 100644
index 00000000..dc8dff8d
--- /dev/null
+++ b/.github/workflows/matetrack.yml
@@ -0,0 +1,54 @@
+# This workflow will run matetrack on the PR
+
+name: Matetrack
+on:
+  workflow_call:
+jobs:
+  Matetrack:
+    name: Matetrack
+    runs-on: ubuntu-22.04
+    steps:
+      - name: Checkout SF repo 
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event.pull_request.head.sha }}
+          path: Stockfish
+          persist-credentials: false
+
+      - name: build SF
+        working-directory: Stockfish/src
+        run: make -j profile-build
+
+      - name: Checkout matetrack repo
+        uses: actions/checkout@v4
+        with:
+          repository: vondele/matetrack
+          path: matetrack
+          ref: 814160f82e6428ed2f6522dc06c2a6fa539cd413
+          persist-credentials: false
+
+      - name: matetrack install deps
+        working-directory: matetrack
+        run: pip install -r requirements.txt
+
+      - name: cache syzygy
+        id: cache-syzygy
+        uses: actions/cache@v4
+        with:
+           path: |
+              matetrack/3-4-5-wdl/
+              matetrack/3-4-5-dtz/
+           key: key-syzygy
+
+      - name: download syzygy 3-4-5 if needed
+        working-directory: matetrack
+        if: steps.cache-syzygy.outputs.cache-hit != 'true'
+        run: |
+          wget --no-verbose -r -nH --cut-dirs=2 --no-parent --reject="index.html*" -e robots=off https://tablebase.lichess.ovh/tables/standard/3-4-5-wdl/
+          wget --no-verbose -r -nH --cut-dirs=2 --no-parent --reject="index.html*" -e robots=off https://tablebase.lichess.ovh/tables/standard/3-4-5-dtz/
+
+      - name: Run matetrack
+        working-directory: matetrack
+        run: |
+          python matecheck.py  --syzygyPath 3-4-5-wdl/:3-4-5-dtz/ --engine /home/runner/work/Stockfish/Stockfish/Stockfish/src/stockfish --epdFile mates2000.epd --nodes 100000 | tee matecheckout.out
+          ! grep "issues were detected" matecheckout.out > /dev/null
diff --git a/.github/workflows/sanitizers.yml b/.github/workflows/sanitizers.yml
index b75c06cf..946a81ce 100644
--- a/.github/workflows/sanitizers.yml
+++ b/.github/workflows/sanitizers.yml
@@ -40,6 +40,8 @@ jobs:
         shell: ${{ matrix.config.shell }}
     steps:
       - uses: actions/checkout@v4
+        with:
+          persist-credentials: false
 
       - name: Download required linux packages
         run: |
@@ -73,4 +75,4 @@ jobs:
           export CXXFLAGS="-O1 -fno-inline"
           make clean
           make -j4 ARCH=x86-64-sse41-popcnt ${{ matrix.sanitizers.make_option }} debug=yes optimize=no build > /dev/null
-          ../tests/instrumented.sh --${{ matrix.sanitizers.instrumented_option }}
+          python3 ../tests/instrumented.py --${{ matrix.sanitizers.instrumented_option }} ./stockfish
diff --git a/.github/workflows/stockfish.yml b/.github/workflows/stockfish.yml
index 13d57f9e..1f87e061 100644
--- a/.github/workflows/stockfish.yml
+++ b/.github/workflows/stockfish.yml
@@ -15,8 +15,12 @@ jobs:
   Prerelease:
     if: github.repository == 'official-stockfish/Stockfish' && (github.ref == 'refs/heads/master' || (startsWith(github.ref_name, 'sf_') && github.ref_type == 'tag'))
     runs-on: ubuntu-latest
+    permissions:
+      contents: write # For deleting/creating a prerelease
     steps:
       - uses: actions/checkout@v4
+        with:
+          persist-credentials: false
 
       # returns null if no pre-release exists
       - name: Get Commit SHA of Latest Pre-release
@@ -66,6 +70,8 @@ jobs:
       arm_matrix: ${{ steps.set-arm-matrix.outputs.arm_matrix }}
     steps:
       - uses: actions/checkout@v4
+        with:
+          persist-credentials: false
       - id: set-matrix
         run: |
           TASKS=$(echo $(cat .github/ci/matrix.json) )
@@ -90,15 +96,27 @@ jobs:
     uses: ./.github/workflows/sanitizers.yml
   Tests:
     uses: ./.github/workflows/tests.yml
+  Matetrack:
+    uses: ./.github/workflows/matetrack.yml
+  Games:
+    uses: ./.github/workflows/games.yml
   Binaries:
     if: github.repository == 'official-stockfish/Stockfish'
     needs: [Matrix, Prerelease, Compilation]
     uses: ./.github/workflows/upload_binaries.yml
     with:
       matrix: ${{ needs.Matrix.outputs.matrix }}
+    permissions:
+      contents: write # For deleting/creating a (pre)release
+    secrets:
+      token: ${{ secrets.GITHUB_TOKEN }}
   ARM_Binaries:
     if: github.repository == 'official-stockfish/Stockfish'
     needs: [Matrix, Prerelease, ARMCompilation]
     uses: ./.github/workflows/upload_binaries.yml
     with:
       matrix: ${{ needs.Matrix.outputs.arm_matrix }}
+    permissions:
+      contents: write # For deleting/creating a (pre)release
+    secrets:
+      token: ${{ secrets.GITHUB_TOKEN }}
diff --git a/.github/workflows/tests.yml b/.github/workflows/tests.yml
index 328c9cf9..b97aaa29 100644
--- a/.github/workflows/tests.yml
+++ b/.github/workflows/tests.yml
@@ -106,6 +106,7 @@ jobs:
       - uses: actions/checkout@v4
         with:
           fetch-depth: 0
+          persist-credentials: false
 
       - name: Download required linux packages
         if: runner.os == 'Linux'
@@ -138,16 +139,16 @@ jobs:
       - name: Build Docker container
         if: matrix.config.base_image
         run: |
-          docker buildx build --load -t sf_builder - << EOF
+          docker buildx build --platform ${{ matrix.config.platform }} --load -t sf_builder - << EOF
           FROM ${{ matrix.config.base_image }}
           WORKDIR /app
           RUN apk update && apk add make g++
-          CMD ["sh", "script.sh"]
+          CMD ["sh", "src/script.sh"]
           EOF
 
       - name: Download required macOS packages
         if: runner.os == 'macOS'
-        run: brew install coreutils
+        run: brew install coreutils gcc@11
 
       - name: Setup msys and install required packages
         if: runner.os == 'Windows'
@@ -175,7 +176,7 @@ jobs:
             $COMPCXX -v
           else
             echo "$COMPCXX -v" > script.sh
-            docker run --rm --platform ${{ matrix.config.platform }} -v ${{ github.workspace }}/src:/app sf_builder
+            docker run --rm --platform ${{ matrix.config.platform }} -v ${{ github.workspace }}:/app sf_builder
           fi
 
       - name: Test help target
@@ -341,8 +342,8 @@ jobs:
       - name: Test riscv64 build
         if: matrix.config.run_riscv64_tests
         run: |
-          echo "export LDFLAGS='-static' && make clean && make -j4 ARCH=riscv64 build" > script.sh
-          docker run --rm --platform ${{ matrix.config.platform }} -v ${{ github.workspace }}/src:/app sf_builder
+          echo "cd src && export LDFLAGS='-static' && make clean && make -j4 ARCH=riscv64 build" > script.sh
+          docker run --rm --platform ${{ matrix.config.platform }} -v ${{ github.workspace }}:/app sf_builder
           ../tests/signature.sh $benchref
 
       # ppc64 tests
@@ -350,8 +351,8 @@ jobs:
       - name: Test ppc64 build
         if: matrix.config.run_ppc64_tests
         run: |
-          echo "export LDFLAGS='-static' && make clean && make -j4 ARCH=ppc-64 build" > script.sh
-          docker run --rm --platform ${{ matrix.config.platform }} -v ${{ github.workspace }}/src:/app sf_builder
+          echo "cd src && export LDFLAGS='-static' && make clean && make -j4 ARCH=ppc-64 build" > script.sh
+          docker run --rm --platform ${{ matrix.config.platform }} -v ${{ github.workspace }}:/app sf_builder
           ../tests/signature.sh $benchref
 
       # Other tests
diff --git a/.github/workflows/upload_binaries.yml b/.github/workflows/upload_binaries.yml
index acf91a8f..1067f6e7 100644
--- a/.github/workflows/upload_binaries.yml
+++ b/.github/workflows/upload_binaries.yml
@@ -5,6 +5,9 @@ on:
       matrix:
         type: string
         required: true
+    secrets:
+      token:
+        required: true
 
 jobs:
   Artifacts:
@@ -25,6 +28,8 @@ jobs:
         shell: ${{ matrix.config.shell }}
     steps:
       - uses: actions/checkout@v4
+        with:
+          persist-credentials: false
 
       - name: Download artifact from compilation
         uses: actions/download-artifact@v4
@@ -54,6 +59,7 @@ jobs:
           mv "${{ matrix.config.simple_name }} ${{ matrix.binaries }}" stockfish-workflow
           cd stockfish-workflow
           cp -r src ../stockfish/
+          cp -r scripts ../stockfish/
           cp stockfish-$NAME-$BINARY$EXT ../stockfish/
           cp "Top CPU Contributors.txt" ../stockfish/
           cp Copying.txt ../stockfish/
@@ -78,6 +84,7 @@ jobs:
         uses: softprops/action-gh-release@4634c16e79c963813287e889244c50009e7f0981
         with:
           files: stockfish-${{ matrix.config.simple_name }}-${{ matrix.binaries }}.${{ matrix.config.archive_ext }}
+          token: ${{ secrets.token }}
 
       - name: Get last commit sha
         id: last_commit
@@ -104,3 +111,4 @@ jobs:
           tag_name: stockfish-dev-${{ env.COMMIT_DATE }}-${{ env.COMMIT_SHA }}
           prerelease: true
           files: stockfish-${{ matrix.config.simple_name }}-${{ matrix.binaries }}.${{ matrix.config.archive_ext }}
+          token: ${{ secrets.token }}
diff --git a/.gitignore b/.gitignore
index 8981efca..2fc80d48 100644
--- a/.gitignore
+++ b/.gitignore
@@ -10,3 +10,8 @@ src/-lstdc++.res
 # Neural network for the NNUE evaluation
 **/*.nnue
 
+# Files generated by the instrumented tests
+tsan.supp
+__pycache__/
+tests/syzygy
+tests/bench_tmp.epd
\ No newline at end of file
diff --git a/AUTHORS b/AUTHORS
index 36b2b6f7..ddc53ec0 100644
--- a/AUTHORS
+++ b/AUTHORS
@@ -20,6 +20,7 @@ Alexander Kure
 Alexander Pagel (Lolligerhans)
 Alfredo Menezes (lonfom169)
 Ali AlZhrani (Cooffe)
+Andreas Jan van der Meulen (Andyson007)
 Andreas Matthies (Matthies)
 Andrei Vetrov (proukornew)
 Andrew Grant (AndyGrant)
@@ -44,6 +45,7 @@ Bruno de Melo Costa (BM123499)
 Bruno Pellanda (pellanda)
 Bryan Cross (crossbr)
 candirufish
+Carlos Esparza Sánchez (ces42)
 Chess13234
 Chris Cain (ceebo)
 Ciekce
@@ -68,9 +70,11 @@ Douglas Matos Gomes (dsmsgms)
 Dubslow
 Eduardo Cáceres (eduherminio)
 Eelco de Groot (KingDefender)
+Ehsan Rashid (erashid)
 Elvin Liu (solarlight2)
 erbsenzaehler
 Ernesto Gatti
+evqsx
 Fabian Beuke (madnight)
 Fabian Fichter (ianfab)
 Fanael Linithien (Fanael)
@@ -127,6 +131,7 @@ Kojirion
 Krystian Kuzniarek (kuzkry)
 Leonardo Ljubičić (ICCF World Champion)
 Leonid Pechenik (lp--)
+Li Ying (yl25946)
 Liam Keegan (lkeegan)
 Linmiao Xu (linrock)
 Linus Arver (listx)
@@ -139,6 +144,7 @@ Maciej Żenczykowski (zenczykowski)
 Malcolm Campbell (xoto10)
 Mark Tenzer (31m059)
 marotear
+Mathias Parnaudeau (mparnaudeau)
 Matt Ginsberg (mattginsberg)
 Matthew Lai (matthewlai)
 Matthew Sullivan (Matt14916)
@@ -167,10 +173,12 @@ Niklas Fiekas (niklasf)
 Nikolay Kostov (NikolayIT)
 Norman Schmidt (FireFather)
 notruck
+Nour Berakdar (Nonlinear)
 Ofek Shochat (OfekShochat, ghostway)
 Ondrej Mosnáček (WOnder93)
 Ondřej Mišina (AndrovT)
 Oskar Werkelin Ahlin
+Ömer Faruk Tutkun (OmerFarukTutkun)
 Pablo Vazquez
 Panthee
 Pascal Romaret
@@ -232,6 +240,7 @@ Unai Corzo (unaiic)
 Uri Blass (uriblass)
 Vince Negri (cuddlestmonkey)
 Viren
+Wencey Wang
 windfishballad
 xefoci7612
 Xiang Wang (KatyushaScarlet)
diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
index cf9cecda..caffc916 100644
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -59,7 +59,7 @@ discussion._
 
 Changes to Stockfish C++ code should respect our coding style defined by
 [.clang-format](.clang-format). You can format your changes by running
-`make format`. This requires clang-format version 17 to be installed on your system.
+`make format`. This requires clang-format version 18 to be installed on your system.
 
 ## Navigate
 
diff --git a/README.md b/README.md
index 1530264b..0a5e8eb4 100644
--- a/README.md
+++ b/README.md
@@ -99,7 +99,7 @@ descriptions. An example suitable for most Intel and AMD chips:
 
 ```
 cd src
-make -j profile-build ARCH=x86-64-avx2
+make -j profile-build
 ```
 
 Detailed compilation instructions for all platforms can be found in our
@@ -122,6 +122,11 @@ where the source code can be found) to generate the exact binary you are
 distributing. If you make any changes to the source code, these changes must
 also be made available under GPL v3.
 
+## Acknowledgements
+
+Stockfish uses neural networks trained on [data provided by the Leela Chess Zero
+project][lc0-data-link], which is made available under the [Open Database License][odbl-link] (ODbL).
+
 
 [authors-link]:       https://github.com/official-stockfish/Stockfish/blob/master/AUTHORS
 [build-link]:         https://github.com/official-stockfish/Stockfish/actions/workflows/stockfish.yml
@@ -146,6 +151,8 @@ also be made available under GPL v3.
 [wiki-uci-link]:      https://github.com/official-stockfish/Stockfish/wiki/UCI-&-Commands
 [wiki-usage-link]:    https://github.com/official-stockfish/Stockfish/wiki/Download-and-usage
 [worker-link]:        https://github.com/official-stockfish/fishtest/wiki/Running-the-worker
+[lc0-data-link]:      https://storage.lczero.org/files/training_data
+[odbl-link]:          https://opendatacommons.org/licenses/odbl/odbl-10.txt
 
 [build-badge]:        https://img.shields.io/github/actions/workflow/status/official-stockfish/Stockfish/stockfish.yml?branch=master&style=for-the-badge&label=stockfish&logo=github
 [commits-badge]:      https://img.shields.io/github/commits-since/official-stockfish/Stockfish/latest?style=for-the-badge
diff --git a/Top CPU Contributors.txt b/Top CPU Contributors.txt
index 11636e84..3d8c5236 100644
--- a/Top CPU Contributors.txt	
+++ b/Top CPU Contributors.txt	
@@ -1,106 +1,109 @@
-Contributors to Fishtest with >10,000 CPU hours, as of 2024-02-24.
+Contributors to Fishtest with >10,000 CPU hours, as of 2024-08-31.
 Thank you!
 
 Username                                CPU Hours     Games played
 ------------------------------------------------------------------
-noobpwnftw                               39302472       3055513453
-technologov                              20845762        994893444
-linrock                                   8616428        560281417
+noobpwnftw                               40428649       3164740143
+technologov                              23581394       1076895482
+vdv                                      19425375        718302718
+linrock                                  10034115        643194527
 mlang                                     3026000        200065824
-okrout                                    2332151        222639518
-pemo                                      1800019         60274069
+okrout                                    2572676        237511408
+pemo                                      1836785         62226157
 dew                                       1689162        100033738
-TueRens                                   1474943         75121774
-grandphish2                               1463002         91616949
-JojoM                                     1109702         72927902
-olafm                                      978631         71037944
-sebastronomy                               939955         44920556
+TueRens                                   1648780         77891164
+sebastronomy                              1468328         60859092
+grandphish2                               1466110         91776075
+JojoM                                     1130625         73666098
+olafm                                     1067009         74807270
 tvijlbrief                                 796125         51897690
-gvreuls                                    711320         49142318
+oz                                         781847         53910686
+rpngn                                      768460         49812975
+gvreuls                                    751085         52177668
 mibere                                     703840         46867607
-oz                                         646268         46293638
-rpngn                                      572571         38928563
-leszek                                     531858         39316505
-cw                                         518116         34894291
+leszek                                     566598         42024615
+cw                                         519601         34988161
 fastgm                                     503862         30260818
 CSU_Dynasty                                468784         31385034
-ctoks                                      434591         28520597
-maximmasiutin                              429983         27066286
+maximmasiutin                              439192         27893522
+ctoks                                      435148         28541909
 crunchy                                    427414         27371625
 bcross                                     415724         29061187
+robal                                      371112         24642270
+mgrabiak                                   367963         26464704
 velislav                                   342588         22140902
-mgrabiak                                   338763         23999170
+ncfish1                                    329039         20624527
 Fisherman                                  327231         21829379
-robal                                      299836         20213182
 Dantist                                    296386         18031762
-ncfish1                                    267604         17881149
+tolkki963                                  262050         22049676
+Sylvain27                                  255595          8864404
 nordlandia                                 249322         16420192
+Fifis                                      237657         13065577
 marrco                                     234581         17714473
-tolkki963                                  233490         19773930
+Calis007                                   217537         14450582
 glinscott                                  208125         13277240
 drabel                                     204167         13930674
 mhoram                                     202894         12601997
 bking_US                                   198894         11876016
-Calis007                                   188631         12795784
 Thanar                                     179852         12365359
-Fifis                                      176209         10638245
-vdv                                        175544          9904472
+javran                                     169679         13481966
+armo9494                                   162863         10937118
 spams                                      157128         10319326
-DesolatedDodo                              156659         10210328
-armo9494                                   155355         10566898
+DesolatedDodo                              156683         10211206
+Wencey                                     152308          8375444
 sqrt2                                      147963          9724586
+vdbergh                                    140311          9225125
 jcAEie                                     140086         10603658
-vdbergh                                    139746          9172061
 CoffeeOne                                  137100          5024116
 malala                                     136182          8002293
 xoto                                       133759          9159372
+Dubslow                                    129614          8519312
 davar                                      129023          8376525
 DMBK                                       122960          8980062
 dsmith                                     122059          7570238
-javran                                     121564         10144656
+CypressChess                               120784          8672620
+sschnee                                    120526          7547722
+maposora                                   119734         10749710
 amicic                                     119661          7938029
-sschnee                                    118107          7389266
-Wolfgang                                   114616          8070494
+Wolfgang                                   115713          8159062
 Data                                       113305          8220352
 BrunoBanani                                112960          7436849
-Wencey                                     111502          5991676
-cuistot                                    108503          7006992
-CypressChess                               108331          7759788
+markkulix                                  112897          9133168
+cuistot                                    109802          7121030
 skiminki                                   107583          7218170
+sterni1971                                 104431          5938282
 MaZePallas                                 102823          6633619
-sterni1971                                 100532          5880772
 sunu                                       100167          7040199
 zeryl                                       99331          6221261
 thirdlife                                   99156          2245320
 ElbertoOne                                  99028          7023771
-Dubslow                                     98600          6903242
-markkulix                                   97010          7643900
-bigpen0r                                    94809          6529203
+megaman7de                                  98456          6675076
+Goatminola                                  96765          8257832
+bigpen0r                                    94825          6529241
 brabos                                      92118          6186135
 Maxim                                       90818          3283364
 psk                                         89957          5984901
-megaman7de                                  88822          6052132
 racerschmacer                               85805          6122790
-maposora                                    85710          7778146
 Vizvezdenec                                 83761          5344740
 0x3C33                                      82614          5271253
+szupaw                                      82495          7151686
 BRAVONE                                     81239          5054681
 nssy                                        76497          5259388
+cody                                        76126          4492126
 jromang                                     76106          5236025
+MarcusTullius                               76103          5061991
+woutboat                                    76072          6022922
+Spprtr                                      75977          5252287
 teddybaer                                   75125          5407666
 Pking_cda                                   73776          5293873
-yurikvelo                                   73516          5036928
-MarcusTullius                               71053          4803477
+yurikvelo                                   73611          5046822
+Mineta                                      71130          4711422
 Bobo1239                                    70579          4794999
 solarlight                                  70517          5028306
 dv8silencer                                 70287          3883992
-Spprtr                                      69646          4806763
-Mineta                                      66325          4537742
 manap                                       66273          4121774
-szupaw                                      65468          5669742
 tinker                                      64333          4268790
 qurashee                                    61208          3429862
-woutboat                                    59496          4906352
 AGI                                         58195          4329580
 robnjr                                      57262          4053117
 Freja                                       56938          3733019
@@ -108,39 +111,45 @@ MaxKlaxxMiner                               56879          3423958
 ttruscott                                   56010          3680085
 rkl                                         55132          4164467
 jmdana                                      54697          4012593
+notchris                                    53936          4184018
 renouve                                     53811          3501516
-notchris                                    52433          4044590
 finfish                                     51360          3370515
 eva42                                       51272          3599691
 eastorwest                                  51117          3454811
-Goatminola                                  51004          4432492
 rap                                         49985          3219146
 pb00067                                     49733          3298934
 GPUex                                       48686          3684998
 OuaisBla                                    48626          3445134
 ronaldjerum                                 47654          3240695
 biffhero                                    46564          3111352
-oryx                                        45533          3539290
+oryx                                        45639          3546530
 VoyagerOne                                  45476          3452465
 speedycpu                                   43842          3003273
 jbwiebe                                     43305          2805433
 Antihistamine                               41788          2761312
 mhunt                                       41735          2691355
+jibarbosa                                   41640          4145702
 homyur                                      39893          2850481
 gri                                         39871          2515779
+DeepnessFulled                              39020          3323102
 Garf                                        37741          2999686
 SC                                          37299          2731694
-Sylvain27                                   36520          1467082
+Gaster319                                   37118          3279678
+naclosagc                                   36562          1279618
 csnodgrass                                  36207          2688994
-Gaster319                                   35655          3149442
 strelock                                    34716          2074055
+gopeto                                      33717          2245606
 EthanOConnor                                33370          2090311
 slakovv                                     32915          2021889
-gopeto                                      31884          2076712
+jojo2357                                    32890          2826662
+shawnxu                                     32019          2802552
 Gelma                                       31771          1551204
+vidar808                                    31560          1351810
 kdave                                       31157          2198362
 manapbk                                     30987          1810399
-ZacHFX                                      30551          2238078
+ZacHFX                                      30966          2272416
+TataneSan                                   30713          1513402
+votoanthuan                                 30691          2460856
 Prcuvu                                      30377          2170122
 anst                                        30301          2190091
 jkiiski                                     30136          1904470
@@ -149,14 +158,15 @@ hyperbolic.tom                              29840          2017394
 chuckstablers                               29659          2093438
 Pyafue                                      29650          1902349
 belzedar94                                  28846          1811530
-votoanthuan                                 27978          2285818
-shawnxu                                     27438          2465810
+mecevdimitar                                27610          1721382
 chriswk                                     26902          1868317
 xwziegtm                                    26897          2124586
 achambord                                   26582          1767323
+somethingintheshadows                       26496          2186404
 Patrick_G                                   26276          1801617
 yorkman                                     26193          1992080
-Ulysses                                     25397          1701264
+srowen                                      25743          1490684
+Ulysses                                     25413          1702830
 Jopo12321                                   25227          1652482
 SFTUser                                     25182          1675689
 nabildanial                                 25068          1531665
@@ -164,66 +174,69 @@ Sharaf_DG                                   24765          1786697
 rodneyc                                     24376          1416402
 jsys14                                      24297          1721230
 agg177                                      23890          1395014
-srowen                                      23842          1342508
+AndreasKrug                                 23754          1890115
 Ente                                        23752          1678188
-jojo2357                                    23479          2061238
 JanErik                                     23408          1703875
 Isidor                                      23388          1680691
 Norabor                                     23371          1603244
+WoodMan777                                  23253          2023048
+Nullvalue                                   23155          2022752
 cisco2015                                   22920          1763301
 Zirie                                       22542          1472937
-Nullvalue                                   22490          1970374
-AndreasKrug                                 22485          1769491
 team-oh                                     22272          1636708
 Roady                                       22220          1465606
 MazeOfGalious                               21978          1629593
-sg4032                                      21947          1643353
+sg4032                                      21950          1643373
+tsim67                                      21747          1330880
 ianh2105                                    21725          1632562
+Skiff84                                     21711          1014212
 xor12                                       21628          1680365
 dex                                         21612          1467203
 nesoneg                                     21494          1463031
 user213718                                  21454          1404128
+Serpensin                                   21452          1790510
 sphinx                                      21211          1384728
-qoo_charly_cai                              21135          1514907
+qoo_charly_cai                              21136          1514927
+IslandLambda                                21062          1220838
 jjoshua2                                    21001          1423089
 Zake9298                                    20938          1565848
 horst.prack                                 20878          1465656
+fishtester                                  20729          1348888
 0xB00B1ES                                   20590          1208666
-Serpensin                                   20487          1729674
-Dinde                                       20440          1292390
+ols                                         20477          1195945
+Dinde                                       20459          1292774
 j3corre                                     20405           941444
 Adrian.Schmidt123                           20316          1281436
 wei                                         19973          1745989
-fishtester                                  19617          1257388
+teenychess                                  19819          1762006
 rstoesser                                   19569          1293588
 eudhan                                      19274          1283717
 vulcan                                      18871          1729392
+wizardassassin                              18795          1376884
 Karpovbot                                   18766          1053178
-WoodMan777                                  18556          1628264
 jundery                                     18445          1115855
+mkstockfishtester                           18350          1690676
 ville                                       17883          1384026
 chris                                       17698          1487385
 purplefishies                               17595          1092533
 dju                                         17414           981289
-ols                                         17291          1042003
 iisiraider                                  17275          1049015
-Skiff84                                     17111           950248
 DragonLord                                  17014          1162790
+Karby                                       17008          1013160
+pirt                                        16965          1271519
 redstone59                                  16842          1461780
-Karby                                       16839          1010124
 Alb11747                                    16787          1213990
-pirt                                        16493          1237199
 Naven94                                     16414           951718
-wizardassassin                              16392          1148672
+scuzzi                                      16115           994341
 IgorLeMasson                                16064          1147232
-scuzzi                                      15757           968735
 ako027ako                                   15671          1173203
+infinigon                                   15285           965966
 Nikolay.IT                                  15154          1068349
 Andrew Grant                                15114           895539
 OssumOpossum                                14857          1007129
 LunaticBFF57                                14525          1190310
 enedene                                     14476           905279
-IslandLambda                                14393           958196
+Hjax                                        14394          1005013
 bpfliegel                                   14233           882523
 YELNAMRON                                   14230          1128094
 mpx86                                       14019           759568
@@ -233,54 +246,56 @@ Nesa92                                      13806          1116101
 crocogoat                                   13803          1117422
 joster                                      13710           946160
 mbeier                                      13650          1044928
-Hjax                                        13535           915487
+Pablohn26                                   13552          1088532
+wxt9861                                     13550          1312306
 Dark_wizzie                                 13422          1007152
 Rudolphous                                  13244           883140
 Machariel                                   13010           863104
-infinigon                                   12991           943216
+nalanzeyu                                   12996           232590
 mabichito                                   12903           749391
+Jackfish                                    12895           868928
 thijsk                                      12886           722107
 AdrianSA                                    12860           804972
 Flopzee                                     12698           894821
+whelanh                                     12682           266404
 mschmidt                                    12644           863193
 korposzczur                                 12606           838168
-tsim67                                      12570           890180
-Jackfish                                    12553           836958
 fatmurphy                                   12547           853210
-Oakwen                                      12503           853105
+Oakwen                                      12532           855759
+icewulf                                     12447           854878
 SapphireBrand                               12416           969604
 deflectooor                                 12386           579392
 modolief                                    12386           896470
-TataneSan                                   12358           609332
 Farseer                                     12249           694108
+Hongildong                                  12201           648712
 pgontarz                                    12151           848794
 dbernier                                    12103           860824
-FormazChar                                  11989           907809
+szczur90                                    12035           942376
+FormazChar                                  12019           910409
+rensonthemove                               11999           971993
 stocky                                      11954           699440
-somethingintheshadows                       11940           989472
-MooTheCow                                   11892           776126
+MooTheCow                                   11923           779432
 3cho                                        11842          1036786
-whelanh                                     11557           245188
+ckaz                                        11792           732276
 infinity                                    11470           727027
 aga                                         11412           695127
 torbjo                                      11395           729145
 Thomas A. Anderson                          11372           732094
 savage84                                    11358           670860
+Def9Infinity                                11345           696552
 d64                                         11263           789184
 ali-al-zhrani                               11245           779246
-ckaz                                        11170           680866
+ImperiumAeternum                            11155           952000
 snicolet                                    11106           869170
 dapper                                      11032           771402
 Ethnikoi                                    10993           945906
 Snuuka                                      10938           435504
-Karmatron                                   10859           678058
+Karmatron                                   10871           678306
 basepi                                      10637           744851
-jibarbosa                                   10628           857100
 Cubox                                       10621           826448
-mecevdimitar                                10609           787318
+gerbil                                      10519           971688
 michaelrpg                                  10509           739239
-Def9Infinity                                10427           686978
 OIVAS7572                                   10420           995586
-wxt9861                                     10412          1013864
 Garruk                                      10365           706465
 dzjp                                        10343           732529
+RickGroszkiewicz                            10263           990798
diff --git a/scripts/get_native_properties.sh b/scripts/get_native_properties.sh
index fb124021..ed5fc9af 100755
--- a/scripts/get_native_properties.sh
+++ b/scripts/get_native_properties.sh
@@ -26,6 +26,17 @@ check_znver_1_2() {
   [ "$vendor_id" = "AuthenticAMD" ] && [ "$cpu_family" = "23" ] && znver_1_2=true
 }
 
+# Set the file CPU loongarch64 architecture
+set_arch_loongarch64() {
+  if check_flags 'lasx'; then
+    true_arch='loongarch64-lasx'
+  elif check_flags 'lsx'; then
+    true_arch='lonngarch64-lsx'
+  else
+    true_arch='loongarch64'
+  fi
+}
+
 # Set the file CPU x86_64 architecture
 set_arch_x86_64() {
   if check_flags 'avx512vnni' 'avx512dq' 'avx512f' 'avx512bw' 'avx512vl'; then
@@ -43,6 +54,20 @@ set_arch_x86_64() {
   fi
 }
 
+set_arch_ppc_64() {
+  if $(grep -q -w "altivec" /proc/cpuinfo); then
+    power=$(grep -oP -m 1 'cpu\t+: POWER\K\d+' /proc/cpuinfo)
+    if [ "0$power" -gt 7 ]; then
+      # VSX started with POWER8
+      true_arch='ppc-64-vsx'
+    else
+      true_arch='ppc-64-altivec'
+    fi
+  else
+    true_arch='ppc-64'
+  fi
+}
+
 # Check the system type
 uname_s=$(uname -s)
 uname_m=$(uname -m)
@@ -76,6 +101,10 @@ case $uname_s in
         file_os='ubuntu'
         true_arch='x86-32'
         ;;
+      'ppc64'*)
+        file_os='ubuntu'
+        set_arch_ppc_64
+        ;;
       'aarch64')
         file_os='android'
         true_arch='armv8'
@@ -90,6 +119,10 @@ case $uname_s in
           true_arch="$true_arch-neon"
         fi
         ;;
+      'loongarch64'*)
+        file_os='linux'
+        set_arch_loongarch64
+        ;;
       *) # Unsupported machine type, exit with error
         printf 'Unsupported machine type: %s\n' "$uname_m"
         exit 1
diff --git a/scripts/net.sh b/scripts/net.sh
new file mode 100755
index 00000000..0bc57a19
--- /dev/null
+++ b/scripts/net.sh
@@ -0,0 +1,75 @@
+#!/bin/sh
+
+wget_or_curl=$( (command -v wget > /dev/null 2>&1 && echo "wget -qO-") || \
+                (command -v curl > /dev/null 2>&1 && echo "curl -skL"))
+
+if [ -z "$wget_or_curl" ]; then
+  >&2 printf "%s\n" "Neither wget or curl is installed." \
+	         "Install one of these tools to download NNUE files automatically."
+  exit 1
+fi
+
+sha256sum=$( (command -v shasum > /dev/null 2>&1 && echo "shasum -a 256") || \
+             (command -v sha256sum > /dev/null 2>&1 && echo "sha256sum"))
+
+if [ -z "$sha256sum" ]; then
+  >&2 echo "sha256sum not found, NNUE files will be assumed valid."
+fi
+
+get_nnue_filename() {
+  grep "$1" evaluate.h | grep "#define" | sed "s/.*\(nn-[a-z0-9]\{12\}.nnue\).*/\1/"
+}
+
+validate_network() {
+  # If no sha256sum command is available, assume the file is always valid.
+  if [ -n "$sha256sum" ] && [ -f "$1" ]; then
+    if [ "$1" != "nn-$($sha256sum "$1" | cut -c 1-12).nnue" ]; then
+      rm -f "$1"
+      return 1
+    fi
+  fi
+}
+
+fetch_network() {
+  _filename="$(get_nnue_filename "$1")"
+
+  if [ -z "$_filename" ]; then
+    >&2 echo "NNUE file name not found for: $1"
+    return 1
+  fi
+
+  if [ -f "$_filename" ]; then
+    if validate_network "$_filename"; then
+      echo "Existing $_filename validated, skipping download"
+      return
+    else
+      echo "Removing invalid NNUE file: $_filename"
+    fi
+  fi
+
+  for url in \
+    "https://tests.stockfishchess.org/api/nn/$_filename" \
+    "https://github.com/official-stockfish/networks/raw/master/$_filename"; do
+    echo "Downloading from $url ..."
+    if $wget_or_curl "$url" > "$_filename"; then
+      if validate_network "$_filename"; then
+        echo "Successfully validated $_filename"
+      else
+        echo "Downloaded $_filename is invalid"
+        continue
+      fi
+    else
+      echo "Failed to download from $url"
+    fi
+    if [ -f "$_filename" ]; then
+      return
+    fi
+  done
+
+  # Download was not successful in the loop, return false.
+  >&2 echo "Failed to download $_filename"
+  return 1
+}
+
+fetch_network EvalFileDefaultNameBig && \
+fetch_network EvalFileDefaultNameSmall
diff --git a/src/Makefile b/src/Makefile
index 45f38b01..e7f8ce55 100644
--- a/src/Makefile
+++ b/src/Makefile
@@ -55,15 +55,15 @@ PGOBENCH = $(WINE_PATH) ./$(EXE) bench
 SRCS = benchmark.cpp bitboard.cpp evaluate.cpp main.cpp \
 	misc.cpp movegen.cpp movepick.cpp position.cpp \
 	search.cpp thread.cpp timeman.cpp tt.cpp uci.cpp ucioption.cpp tune.cpp syzygy/tbprobe.cpp \
-	nnue/nnue_misc.cpp nnue/features/half_ka_v2_hm.cpp nnue/network.cpp engine.cpp score.cpp
+	nnue/nnue_misc.cpp nnue/features/half_ka_v2_hm.cpp nnue/network.cpp engine.cpp score.cpp memory.cpp
 
-HEADERS = benchmark.h bitboard.h evaluate.h misc.h movegen.h movepick.h \
+HEADERS = benchmark.h bitboard.h evaluate.h misc.h movegen.h movepick.h history.h \
 		nnue/nnue_misc.h nnue/features/half_ka_v2_hm.h nnue/layers/affine_transform.h \
 		nnue/layers/affine_transform_sparse_input.h nnue/layers/clipped_relu.h nnue/layers/simd.h \
 		nnue/layers/sqr_clipped_relu.h nnue/nnue_accumulator.h nnue/nnue_architecture.h \
 		nnue/nnue_common.h nnue/nnue_feature_transformer.h position.h \
 		search.h syzygy/tbprobe.h thread.h thread_win32_osx.h timeman.h \
-		tt.h tune.h types.h uci.h ucioption.h perft.h nnue/network.h engine.h score.h
+		tt.h tune.h types.h uci.h ucioption.h perft.h nnue/network.h engine.h score.h numa.h memory.h
 
 OBJS = $(notdir $(SRCS:.cpp=.o))
 
@@ -98,8 +98,12 @@ VPATH = syzygy:nnue:nnue/features
 # avx512 = yes/no     --- -mavx512bw         --- Use Intel Advanced Vector Extensions 512
 # vnni256 = yes/no    --- -mavx256vnni       --- Use Intel Vector Neural Network Instructions 512 with 256bit operands
 # vnni512 = yes/no    --- -mavx512vnni       --- Use Intel Vector Neural Network Instructions 512
+# altivec = yes/no    --- -maltivec          --- Use PowerPC Altivec SIMD extension
+# vsx = yes/no        --- -mvsx              --- Use POWER VSX SIMD extension
 # neon = yes/no       --- -DUSE_NEON         --- Use ARM SIMD architecture
 # dotprod = yes/no    --- -DUSE_NEON_DOTPROD --- Use ARM advanced SIMD Int8 dot product instructions
+# lsx = yes/no        --- -mlsx              --- Use Loongson SIMD eXtension
+# lasx = yes/no       --- -mlasx             --- use Loongson Advanced SIMD eXtension
 #
 # Note that Makefile is space sensitive, so when adding new architectures
 # or modifying existing flags, you have to make sure there are no extra spaces
@@ -124,8 +128,9 @@ endif
 ifeq ($(ARCH), $(filter $(ARCH), \
                  x86-64-vnni512 x86-64-vnni256 x86-64-avx512 x86-64-avxvnni x86-64-bmi2 \
                  x86-64-avx2 x86-64-sse41-popcnt x86-64-modern x86-64-ssse3 x86-64-sse3-popcnt \
-                 x86-64 x86-32-sse41-popcnt x86-32-sse2 x86-32 ppc-64 ppc-32 e2k \
-                 armv7 armv7-neon armv8 armv8-dotprod apple-silicon general-64 general-32 riscv64 loongarch64))
+                 x86-64 x86-32-sse41-popcnt x86-32-sse2 x86-32 ppc-64 ppc-64-altivec ppc-64-vsx ppc-32 e2k \
+                 armv7 armv7-neon armv8 armv8-dotprod apple-silicon general-64 general-32 riscv64 \
+                 loongarch64 loongarch64-lsx loongarch64-lasx))
    SUPPORTED_ARCH=true
 else
    SUPPORTED_ARCH=false
@@ -148,13 +153,17 @@ avxvnni = no
 avx512 = no
 vnni256 = no
 vnni512 = no
+altivec = no
+vsx = no
 neon = no
 dotprod = no
 arm_version = 0
+lsx = no
+lasx = no
 STRIP = strip
 
-ifneq ($(shell which clang-format-17 2> /dev/null),)
-	CLANG-FORMAT = clang-format-17
+ifneq ($(shell which clang-format-18 2> /dev/null),)
+	CLANG-FORMAT = clang-format-18
 else
 	CLANG-FORMAT = clang-format
 endif
@@ -355,6 +364,20 @@ ifeq ($(ARCH),ppc-64)
 	prefetch = yes
 endif
 
+ifeq ($(ARCH),ppc-64-altivec)
+	arch = ppc64
+	popcnt = yes
+	prefetch = yes
+	altivec = yes
+endif
+
+ifeq ($(ARCH),ppc-64-vsx)
+	arch = ppc64
+	popcnt = yes
+	prefetch = yes
+	vsx = yes
+endif
+
 ifeq ($(findstring e2k,$(ARCH)),e2k)
 	arch = e2k
 	mmx = yes
@@ -370,8 +393,19 @@ ifeq ($(ARCH),riscv64)
 	arch = riscv64
 endif
 
-ifeq ($(ARCH),loongarch64)
+ifeq ($(findstring loongarch64,$(ARCH)),loongarch64)
 	arch = loongarch64
+	prefetch = yes
+
+ifeq ($(findstring -lasx,$(ARCH)),-lasx)
+	lsx = yes
+	lasx = yes
+endif
+
+ifeq ($(findstring -lsx,$(ARCH)),-lsx)
+	lsx = yes
+endif
+
 endif
 endif
 
@@ -408,7 +442,7 @@ ifeq ($(COMP),gcc)
 		ifeq ($(ARCH),riscv64)
 			CXXFLAGS += -latomic
 		endif
-	else ifeq ($(ARCH),loongarch64)
+	else ifeq ($(arch),loongarch64)
 		CXXFLAGS += -latomic
 	else
 		CXXFLAGS += -m$(bits)
@@ -480,7 +514,7 @@ ifeq ($(COMP),clang)
 		ifeq ($(ARCH),riscv64)
 			CXXFLAGS += -latomic
 		endif
-	else ifeq ($(ARCH),loongarch64)
+	else ifeq ($(arch),loongarch64)
 		CXXFLAGS += -latomic
 	else
 		CXXFLAGS += -m$(bits)
@@ -489,8 +523,8 @@ ifeq ($(COMP),clang)
 endif
 
 ifeq ($(KERNEL),Darwin)
-	CXXFLAGS += -mmacosx-version-min=10.14
-	LDFLAGS += -mmacosx-version-min=10.14
+	CXXFLAGS += -mmacosx-version-min=10.15
+	LDFLAGS += -mmacosx-version-min=10.15
 	ifneq ($(arch),any)
 		CXXFLAGS += -arch $(arch)
 		LDFLAGS += -arch $(arch)
@@ -634,7 +668,7 @@ else
 endif
 
 ifeq ($(popcnt),yes)
-	ifeq ($(arch),$(filter $(arch),ppc64 armv7 armv8 arm64))
+	ifeq ($(arch),$(filter $(arch),ppc64 ppc64-altivec ppc64-vsx armv7 armv8 arm64))
 		CXXFLAGS += -DUSE_POPCNT
 	else
 		CXXFLAGS += -msse3 -mpopcnt -DUSE_POPCNT
@@ -704,6 +738,20 @@ ifeq ($(mmx),yes)
 	endif
 endif
 
+ifeq ($(altivec),yes)
+	CXXFLAGS += -maltivec
+	ifeq ($(COMP),gcc)
+		CXXFLAGS += -mabi=altivec
+	endif
+endif
+
+ifeq ($(vsx),yes)
+	CXXFLAGS += -mvsx
+	ifeq ($(COMP),gcc)
+		CXXFLAGS += -DNO_WARN_X86_INTRINSICS -DUSE_SSE2
+	endif
+endif
+
 ifeq ($(neon),yes)
 	CXXFLAGS += -DUSE_NEON=$(arm_version)
 	ifeq ($(KERNEL),Linux)
@@ -719,6 +767,18 @@ ifeq ($(dotprod),yes)
 	CXXFLAGS += -march=armv8.2-a+dotprod -DUSE_NEON_DOTPROD
 endif
 
+ifeq ($(lasx),yes)
+	ifeq ($(comp),$(filter $(comp),gcc clang mingw icx))
+		CXXFLAGS += -mlasx
+	endif
+endif
+
+ifeq ($(lsx),yes)
+	ifeq ($(comp),$(filter $(comp),gcc clang mingw icx))
+		CXXFLAGS += -mlsx
+	endif
+endif
+
 ### 3.7 pext
 ifeq ($(pext),yes)
 	CXXFLAGS += -DUSE_PEXT
@@ -791,71 +851,75 @@ endif
 ### ==========================================================================
 
 help:
-	@echo ""
-	@echo "To compile stockfish, type: "
-	@echo ""
-	@echo "make -j target [ARCH=arch] [COMP=compiler] [COMPCXX=cxx]"
-	@echo ""
-	@echo "Supported targets:"
-	@echo ""
-	@echo "help                    > Display architecture details"
-	@echo "profile-build           > standard build with profile-guided optimization"
-	@echo "build                   > skip profile-guided optimization"
-	@echo "net                     > Download the default nnue nets"
-	@echo "strip                   > Strip executable"
-	@echo "install                 > Install executable"
-	@echo "clean                   > Clean up"
-	@echo ""
-	@echo "Supported archs:"
-	@echo ""
-	@echo "native                  > select the best architecture for the host processor (default)"
-	@echo "x86-64-vnni512          > x86 64-bit with vnni 512bit support"
-	@echo "x86-64-vnni256          > x86 64-bit with vnni 512bit support, limit operands to 256bit wide"
-	@echo "x86-64-avx512           > x86 64-bit with avx512 support"
-	@echo "x86-64-avxvnni          > x86 64-bit with vnni 256bit support"
-	@echo "x86-64-bmi2             > x86 64-bit with bmi2 support"
-	@echo "x86-64-avx2             > x86 64-bit with avx2 support"
-	@echo "x86-64-sse41-popcnt     > x86 64-bit with sse41 and popcnt support"
-	@echo "x86-64-modern           > deprecated, currently x86-64-sse41-popcnt"
-	@echo "x86-64-ssse3            > x86 64-bit with ssse3 support"
-	@echo "x86-64-sse3-popcnt      > x86 64-bit with sse3 compile and popcnt support"
-	@echo "x86-64                  > x86 64-bit generic (with sse2 support)"
-	@echo "x86-32-sse41-popcnt     > x86 32-bit with sse41 and popcnt support"
-	@echo "x86-32-sse2             > x86 32-bit with sse2 support"
-	@echo "x86-32                  > x86 32-bit generic (with mmx compile support)"
-	@echo "ppc-64                  > PPC 64-bit"
-	@echo "ppc-32                  > PPC 32-bit"
-	@echo "armv7                   > ARMv7 32-bit"
-	@echo "armv7-neon              > ARMv7 32-bit with popcnt and neon"
-	@echo "armv8                   > ARMv8 64-bit with popcnt and neon"
-	@echo "armv8-dotprod           > ARMv8 64-bit with popcnt, neon and dot product support"
-	@echo "e2k                     > Elbrus 2000"
-	@echo "apple-silicon           > Apple silicon ARM64"
-	@echo "general-64              > unspecified 64-bit"
-	@echo "general-32              > unspecified 32-bit"
-	@echo "riscv64                 > RISC-V 64-bit"
-	@echo "loongarch64             > LoongArch 64-bit"
-	@echo ""
-	@echo "Supported compilers:"
-	@echo ""
-	@echo "gcc                     > GNU compiler (default)"
-	@echo "mingw                   > GNU compiler with MinGW under Windows"
-	@echo "clang                   > LLVM Clang compiler"
-	@echo "icx                     > Intel oneAPI DPC++/C++ Compiler"
-	@echo "ndk                     > Google NDK to cross-compile for Android"
-	@echo ""
-	@echo "Simple examples. If you don't know what to do, you likely want to run one of: "
-	@echo ""
-	@echo "make -j profile-build ARCH=x86-64-avx2    # typically a fast compile for common systems "
-	@echo "make -j profile-build ARCH=x86-64-sse41-popcnt  # A more portable compile for 64-bit systems "
-	@echo "make -j profile-build ARCH=x86-64         # A portable compile for 64-bit systems "
-	@echo ""
-	@echo "Advanced examples, for experienced users: "
-	@echo ""
-	@echo "make -j profile-build ARCH=x86-64-avxvnni"
-	@echo "make -j profile-build ARCH=x86-64-avxvnni COMP=gcc COMPCXX=g++-12.0"
-	@echo "make -j build ARCH=x86-64-ssse3 COMP=clang"
-	@echo ""
+	@echo "" && \
+	echo "To compile stockfish, type: " && \
+	echo "" && \
+	echo "make -j target [ARCH=arch] [COMP=compiler] [COMPCXX=cxx]" && \
+	echo "" && \
+	echo "Supported targets:" && \
+	echo "" && \
+	echo "help                    > Display architecture details" && \
+	echo "profile-build           > standard build with profile-guided optimization" && \
+	echo "build                   > skip profile-guided optimization" && \
+	echo "net                     > Download the default nnue nets" && \
+	echo "strip                   > Strip executable" && \
+	echo "install                 > Install executable" && \
+	echo "clean                   > Clean up" && \
+	echo "" && \
+	echo "Supported archs:" && \
+	echo "" && \
+	echo "native                  > select the best architecture for the host processor (default)" && \
+	echo "x86-64-vnni512          > x86 64-bit with vnni 512bit support" && \
+	echo "x86-64-vnni256          > x86 64-bit with vnni 512bit support, limit operands to 256bit wide" && \
+	echo "x86-64-avx512           > x86 64-bit with avx512 support" && \
+	echo "x86-64-avxvnni          > x86 64-bit with vnni 256bit support" && \
+	echo "x86-64-bmi2             > x86 64-bit with bmi2 support" && \
+	echo "x86-64-avx2             > x86 64-bit with avx2 support" && \
+	echo "x86-64-sse41-popcnt     > x86 64-bit with sse41 and popcnt support" && \
+	echo "x86-64-modern           > deprecated, currently x86-64-sse41-popcnt" && \
+	echo "x86-64-ssse3            > x86 64-bit with ssse3 support" && \
+	echo "x86-64-sse3-popcnt      > x86 64-bit with sse3 compile and popcnt support" && \
+	echo "x86-64                  > x86 64-bit generic (with sse2 support)" && \
+	echo "x86-32-sse41-popcnt     > x86 32-bit with sse41 and popcnt support" && \
+	echo "x86-32-sse2             > x86 32-bit with sse2 support" && \
+	echo "x86-32                  > x86 32-bit generic (with mmx compile support)" && \
+	echo "ppc-64                  > PPC 64-bit" && \
+	echo "ppc-64-altivec          > PPC 64-bit with altivec support" && \
+	echo "ppc-64-vsx              > PPC 64-bit with vsx support" && \
+	echo "ppc-32                  > PPC 32-bit" && \
+	echo "armv7                   > ARMv7 32-bit" && \
+	echo "armv7-neon              > ARMv7 32-bit with popcnt and neon" && \
+	echo "armv8                   > ARMv8 64-bit with popcnt and neon" && \
+	echo "armv8-dotprod           > ARMv8 64-bit with popcnt, neon and dot product support" && \
+	echo "e2k                     > Elbrus 2000" && \
+	echo "apple-silicon           > Apple silicon ARM64" && \
+	echo "general-64              > unspecified 64-bit" && \
+	echo "general-32              > unspecified 32-bit" && \
+	echo "riscv64                 > RISC-V 64-bit" && \
+	echo "loongarch64             > LoongArch 64-bit" && \
+	echo "loongarch64-lsx         > LoongArch 64-bit with SIMD eXtension" && \
+	echo "loongarch64-lasx        > LoongArch 64-bit with Advanced SIMD eXtension" && \
+	echo "" && \
+	echo "Supported compilers:" && \
+	echo "" && \
+	echo "gcc                     > GNU compiler (default)" && \
+	echo "mingw                   > GNU compiler with MinGW under Windows" && \
+	echo "clang                   > LLVM Clang compiler" && \
+	echo "icx                     > Intel oneAPI DPC++/C++ Compiler" && \
+	echo "ndk                     > Google NDK to cross-compile for Android" && \
+	echo "" && \
+	echo "Simple examples. If you don't know what to do, you likely want to run one of: " && \
+	echo "" && \
+	echo "make -j profile-build ARCH=x86-64-avx2    # typically a fast compile for common systems " && \
+	echo "make -j profile-build ARCH=x86-64-sse41-popcnt  # A more portable compile for 64-bit systems " && \
+	echo "make -j profile-build ARCH=x86-64         # A portable compile for 64-bit systems " && \
+	echo "" && \
+	echo "Advanced examples, for experienced users: " && \
+	echo "" && \
+	echo "make -j profile-build ARCH=x86-64-avxvnni" && \
+	echo "make -j profile-build ARCH=x86-64-avxvnni COMP=gcc COMPCXX=g++-12.0" && \
+	echo "make -j build ARCH=x86-64-ssse3 COMP=clang" && \
+	echo ""
 ifneq ($(SUPPORTED_ARCH), true)
 	@echo "Specify a supported architecture with the ARCH option for more details"
 	@echo ""
@@ -917,59 +981,9 @@ profileclean:
 	@rm -f stockfish.res
 	@rm -f ./-lstdc++.res
 
-define fetch_network
-	@echo "Default net: $(nnuenet)"
-	@if [ "x$(curl_or_wget)" = "x" ]; then \
-		echo "Neither curl nor wget is installed. Install one of these tools unless the net has been downloaded manually"; \
-	fi
-	@if [ "x$(shasum_command)" = "x" ]; then \
-		echo "shasum / sha256sum not found, skipping net validation"; \
-	elif test -f "$(nnuenet)"; then \
-		if [ "$(nnuenet)" != "nn-"`$(shasum_command) $(nnuenet) | cut -c1-12`".nnue" ]; then \
-			echo "Removing invalid network"; rm -f $(nnuenet); \
-		fi; \
-	fi;
-	@for nnuedownloadurl in "$(nnuedownloadurl1)" "$(nnuedownloadurl2)"; do \
-		if test -f "$(nnuenet)"; then \
-			echo "$(nnuenet) available : OK"; break; \
-		else \
-			if [ "x$(curl_or_wget)" != "x" ]; then \
-				echo "Downloading $${nnuedownloadurl}"; $(curl_or_wget) $${nnuedownloadurl} > $(nnuenet);\
-			else \
-				echo "No net found and download not possible"; exit 1;\
-			fi; \
-		fi; \
-		if [ "x$(shasum_command)" != "x" ]; then \
-			if [ "$(nnuenet)" != "nn-"`$(shasum_command) $(nnuenet) | cut -c1-12`".nnue" ]; then \
-				echo "Removing failed download"; rm -f $(nnuenet); \
-			fi; \
-		fi; \
-	done
-	@if ! test -f "$(nnuenet)"; then \
-		echo "Failed to download $(nnuenet)."; \
-	fi;
-	@if [ "x$(shasum_command)" != "x" ]; then \
-		if [ "$(nnuenet)" = "nn-"`$(shasum_command) $(nnuenet) | cut -c1-12`".nnue" ]; then \
-			echo "Network validated"; break; \
-		fi; \
-	fi;
-endef
-
-# set up shell variables for the net stuff
-define netvariables
-$(eval nnuenet := $(shell grep $(1) evaluate.h | grep define | sed 's/.*\(nn-[a-z0-9]\{12\}.nnue\).*/\1/'))
-$(eval nnuedownloadurl1 := https://tests.stockfishchess.org/api/nn/$(nnuenet))
-$(eval nnuedownloadurl2 := https://github.com/official-stockfish/networks/raw/master/$(nnuenet))
-$(eval curl_or_wget := $(shell if hash curl 2>/dev/null; then echo "curl -skL"; elif hash wget 2>/dev/null; then echo "wget -qO-"; fi))
-$(eval shasum_command := $(shell if hash shasum 2>/dev/null; then echo "shasum -a 256 "; elif hash sha256sum 2>/dev/null; then echo "sha256sum "; fi))
-endef
-
 # evaluation network (nnue)
 net:
-	$(call netvariables, EvalFileDefaultNameBig)
-	$(call fetch_network)
-	$(call netvariables, EvalFileDefaultNameSmall)
-	$(call fetch_network)
+	@$(SHELL) ../scripts/net.sh
 
 format:
 	$(CLANG-FORMAT) -i $(SRCS) $(HEADERS) -style=file
@@ -986,61 +1000,71 @@ all: $(EXE) .depend
 
 config-sanity: net
 	@echo ""
-	@echo "Config:"
-	@echo "debug: '$(debug)'"
-	@echo "sanitize: '$(sanitize)'"
-	@echo "optimize: '$(optimize)'"
-	@echo "arch: '$(arch)'"
-	@echo "bits: '$(bits)'"
-	@echo "kernel: '$(KERNEL)'"
-	@echo "os: '$(OS)'"
-	@echo "prefetch: '$(prefetch)'"
-	@echo "popcnt: '$(popcnt)'"
-	@echo "pext: '$(pext)'"
-	@echo "sse: '$(sse)'"
-	@echo "mmx: '$(mmx)'"
-	@echo "sse2: '$(sse2)'"
-	@echo "ssse3: '$(ssse3)'"
-	@echo "sse41: '$(sse41)'"
-	@echo "avx2: '$(avx2)'"
-	@echo "avxvnni: '$(avxvnni)'"
-	@echo "avx512: '$(avx512)'"
-	@echo "vnni256: '$(vnni256)'"
-	@echo "vnni512: '$(vnni512)'"
-	@echo "neon: '$(neon)'"
-	@echo "dotprod: '$(dotprod)'"
-	@echo "arm_version: '$(arm_version)'"
-	@echo "target_windows: '$(target_windows)'"
-	@echo ""
-	@echo "Flags:"
-	@echo "CXX: $(CXX)"
-	@echo "CXXFLAGS: $(CXXFLAGS)"
-	@echo "LDFLAGS: $(LDFLAGS)"
-	@echo ""
-	@echo "Testing config sanity. If this fails, try 'make help' ..."
-	@echo ""
-	@test "$(debug)" = "yes" || test "$(debug)" = "no"
-	@test "$(optimize)" = "yes" || test "$(optimize)" = "no"
-	@test "$(SUPPORTED_ARCH)" = "true"
-	@test "$(arch)" = "any" || test "$(arch)" = "x86_64" || test "$(arch)" = "i386" || \
+	@echo "Config:" && \
+	echo "debug: '$(debug)'" && \
+	echo "sanitize: '$(sanitize)'" && \
+	echo "optimize: '$(optimize)'" && \
+	echo "arch: '$(arch)'" && \
+	echo "bits: '$(bits)'" && \
+	echo "kernel: '$(KERNEL)'" && \
+	echo "os: '$(OS)'" && \
+	echo "prefetch: '$(prefetch)'" && \
+	echo "popcnt: '$(popcnt)'" && \
+	echo "pext: '$(pext)'" && \
+	echo "sse: '$(sse)'" && \
+	echo "mmx: '$(mmx)'" && \
+	echo "sse2: '$(sse2)'" && \
+	echo "ssse3: '$(ssse3)'" && \
+	echo "sse41: '$(sse41)'" && \
+	echo "avx2: '$(avx2)'" && \
+	echo "avxvnni: '$(avxvnni)'" && \
+	echo "avx512: '$(avx512)'" && \
+	echo "vnni256: '$(vnni256)'" && \
+	echo "vnni512: '$(vnni512)'" && \
+	echo "altivec: '$(altivec)'" && \
+	echo "vsx: '$(vsx)'" && \
+	echo "neon: '$(neon)'" && \
+	echo "dotprod: '$(dotprod)'" && \
+	echo "arm_version: '$(arm_version)'" && \
+	echo "lsx: '$(lsx)'" && \
+	echo "lasx: '$(lasx)'" && \
+	echo "target_windows: '$(target_windows)'" && \
+	echo "" && \
+	echo "Flags:" && \
+	echo "CXX: $(CXX)" && \
+	echo "CXXFLAGS: $(CXXFLAGS)" && \
+	echo "LDFLAGS: $(LDFLAGS)" && \
+	echo "" && \
+	echo "Testing config sanity. If this fails, try 'make help' ..." && \
+	echo "" && \
+	(test "$(debug)" = "yes" || test "$(debug)" = "no") && \
+	(test "$(optimize)" = "yes" || test "$(optimize)" = "no") && \
+	(test "$(SUPPORTED_ARCH)" = "true") && \
+	(test "$(arch)" = "any" || test "$(arch)" = "x86_64" || test "$(arch)" = "i386" || \
 	 test "$(arch)" = "ppc64" || test "$(arch)" = "ppc" || test "$(arch)" = "e2k" || \
-	 test "$(arch)" = "armv7" || test "$(arch)" = "armv8" || test "$(arch)" = "arm64" || test "$(arch)" = "riscv64" || test "$(arch)" = "loongarch64"
-	@test "$(bits)" = "32" || test "$(bits)" = "64"
-	@test "$(prefetch)" = "yes" || test "$(prefetch)" = "no"
-	@test "$(popcnt)" = "yes" || test "$(popcnt)" = "no"
-	@test "$(pext)" = "yes" || test "$(pext)" = "no"
-	@test "$(sse)" = "yes" || test "$(sse)" = "no"
-	@test "$(mmx)" = "yes" || test "$(mmx)" = "no"
-	@test "$(sse2)" = "yes" || test "$(sse2)" = "no"
-	@test "$(ssse3)" = "yes" || test "$(ssse3)" = "no"
-	@test "$(sse41)" = "yes" || test "$(sse41)" = "no"
-	@test "$(avx2)" = "yes" || test "$(avx2)" = "no"
-	@test "$(avx512)" = "yes" || test "$(avx512)" = "no"
-	@test "$(vnni256)" = "yes" || test "$(vnni256)" = "no"
-	@test "$(vnni512)" = "yes" || test "$(vnni512)" = "no"
-	@test "$(neon)" = "yes" || test "$(neon)" = "no"
-	@test "$(comp)" = "gcc" || test "$(comp)" = "icx" || test "$(comp)" = "mingw" || test "$(comp)" = "clang" \
-	|| test "$(comp)" = "armv7a-linux-androideabi16-clang"  || test "$(comp)" = "aarch64-linux-android21-clang"
+	 test "$(arch)" = "armv7" || test "$(arch)" = "armv8" || test "$(arch)" = "arm64" || \
+	 test "$(arch)" = "riscv64" || test "$(arch)" = "loongarch64") && \
+	(test "$(bits)" = "32" || test "$(bits)" = "64") && \
+	(test "$(prefetch)" = "yes" || test "$(prefetch)" = "no") && \
+	(test "$(popcnt)" = "yes" || test "$(popcnt)" = "no") && \
+	(test "$(pext)" = "yes" || test "$(pext)" = "no") && \
+	(test "$(sse)" = "yes" || test "$(sse)" = "no") && \
+	(test "$(mmx)" = "yes" || test "$(mmx)" = "no") && \
+	(test "$(sse2)" = "yes" || test "$(sse2)" = "no") && \
+	(test "$(ssse3)" = "yes" || test "$(ssse3)" = "no") && \
+	(test "$(sse41)" = "yes" || test "$(sse41)" = "no") && \
+	(test "$(avx2)" = "yes" || test "$(avx2)" = "no") && \
+	(test "$(avx512)" = "yes" || test "$(avx512)" = "no") && \
+	(test "$(vnni256)" = "yes" || test "$(vnni256)" = "no") && \
+	(test "$(vnni512)" = "yes" || test "$(vnni512)" = "no") && \
+	(test "$(altivec)" = "yes" || test "$(altivec)" = "no") && \
+	(test "$(vsx)" = "yes" || test "$(vsx)" = "no") && \
+	(test "$(neon)" = "yes" || test "$(neon)" = "no") && \
+	(test "$(lsx)" = "yes" || test "$(lsx)" = "no") && \
+	(test "$(lasx)" = "yes" || test "$(lasx)" = "no") && \
+	(test "$(comp)" = "gcc" || test "$(comp)" = "icx" || test "$(comp)" = "mingw" || \
+	 test "$(comp)" = "clang" || test "$(comp)" = "armv7a-linux-androideabi16-clang" || \
+	 test "$(comp)" = "aarch64-linux-android21-clang")
 
 $(EXE): $(OBJS)
 	+$(CXX) -o $@ $(OBJS) $(LDFLAGS)
@@ -1051,14 +1075,14 @@ FORCE:
 
 clang-profile-make:
 	$(MAKE) ARCH=$(ARCH) COMP=$(COMP) \
-	EXTRACXXFLAGS='-fprofile-instr-generate ' \
-	EXTRALDFLAGS=' -fprofile-instr-generate' \
+	EXTRACXXFLAGS='-fprofile-generate ' \
+	EXTRALDFLAGS=' -fprofile-generate' \
 	all
 
 clang-profile-use:
 	$(XCRUN) llvm-profdata merge -output=stockfish.profdata *.profraw
 	$(MAKE) ARCH=$(ARCH) COMP=$(COMP) \
-	EXTRACXXFLAGS='-fprofile-instr-use=stockfish.profdata' \
+	EXTRACXXFLAGS='-fprofile-use=stockfish.profdata' \
 	EXTRALDFLAGS='-fprofile-use ' \
 	all
 
diff --git a/src/benchmark.cpp b/src/benchmark.cpp
index 3622ac8a..35ad3c18 100644
--- a/src/benchmark.cpp
+++ b/src/benchmark.cpp
@@ -17,6 +17,7 @@
 */
 
 #include "benchmark.h"
+#include "numa.h"
 
 #include <cstdlib>
 #include <fstream>
@@ -91,6 +92,282 @@ const std::vector<std::string> Defaults = {
 };
 // clang-format on
 
+// clang-format off
+// human-randomly picked 5 games with <60 moves from
+// https://tests.stockfishchess.org/tests/view/665c71f9fd45fb0f907c21e0
+// only moves for one side
+const std::vector<std::vector<std::string>> BenchmarkPositions = {
+    {
+        "rnbq1k1r/ppp1bppp/4pn2/8/2B5/2NP1N2/PPP2PPP/R1BQR1K1 b - - 2 8",
+        "rnbq1k1r/pp2bppp/4pn2/2p5/2B2B2/2NP1N2/PPP2PPP/R2QR1K1 b - - 1 9",
+        "r1bq1k1r/pp2bppp/2n1pn2/2p5/2B1NB2/3P1N2/PPP2PPP/R2QR1K1 b - - 3 10",
+        "r1bq1k1r/pp2bppp/2n1p3/2p5/2B1PB2/5N2/PPP2PPP/R2QR1K1 b - - 0 11",
+        "r1b2k1r/pp2bppp/2n1p3/2p5/2B1PB2/5N2/PPP2PPP/3RR1K1 b - - 0 12",
+        "r1b1k2r/pp2bppp/2n1p3/2p5/2B1PB2/2P2N2/PP3PPP/3RR1K1 b - - 0 13",
+        "r1b1k2r/1p2bppp/p1n1p3/2p5/4PB2/2P2N2/PP2BPPP/3RR1K1 b - - 1 14",
+        "r1b1k2r/4bppp/p1n1p3/1pp5/P3PB2/2P2N2/1P2BPPP/3RR1K1 b - - 0 15",
+        "r1b1k2r/4bppp/p1n1p3/1P6/2p1PB2/2P2N2/1P2BPPP/3RR1K1 b - - 0 16",
+        "r1b1k2r/4bppp/2n1p3/1p6/2p1PB2/1PP2N2/4BPPP/3RR1K1 b - - 0 17",
+        "r3k2r/3bbppp/2n1p3/1p6/2P1PB2/2P2N2/4BPPP/3RR1K1 b - - 0 18",
+        "r3k2r/3bbppp/2n1p3/8/1pP1P3/2P2N2/3BBPPP/3RR1K1 b - - 1 19",
+        "1r2k2r/3bbppp/2n1p3/8/1pPNP3/2P5/3BBPPP/3RR1K1 b - - 3 20",
+        "1r2k2r/3bbppp/2n1p3/8/2PNP3/2B5/4BPPP/3RR1K1 b - - 0 21",
+        "1r2k2r/3bb1pp/2n1pp2/1N6/2P1P3/2B5/4BPPP/3RR1K1 b - - 1 22",
+        "1r2k2r/3b2pp/2n1pp2/1N6/1BP1P3/8/4BPPP/3RR1K1 b - - 0 23",
+        "1r2k2r/3b2pp/4pp2/1N6/1nP1P3/8/3RBPPP/4R1K1 b - - 1 24",
+        "1r5r/3bk1pp/4pp2/1N6/1nP1PP2/8/3RB1PP/4R1K1 b - - 0 25",
+        "1r5r/3bk1pp/2n1pp2/1N6/2P1PP2/8/3RBKPP/4R3 b - - 2 26",
+        "1r5r/3bk1pp/2n2p2/1N2p3/2P1PP2/6P1/3RBK1P/4R3 b - - 0 27",
+        "1r1r4/3bk1pp/2n2p2/1N2p3/2P1PP2/6P1/3RBK1P/R7 b - - 2 28",
+        "1r1r4/N3k1pp/2n1bp2/4p3/2P1PP2/6P1/3RBK1P/R7 b - - 4 29",
+        "1r1r4/3bk1pp/2N2p2/4p3/2P1PP2/6P1/3RBK1P/R7 b - - 0 30",
+        "1r1R4/4k1pp/2b2p2/4p3/2P1PP2/6P1/4BK1P/R7 b - - 0 31",
+        "3r4/4k1pp/2b2p2/4P3/2P1P3/6P1/4BK1P/R7 b - - 0 32",
+        "3r4/R3k1pp/2b5/4p3/2P1P3/6P1/4BK1P/8 b - - 1 33",
+        "8/3rk1pp/2b5/R3p3/2P1P3/6P1/4BK1P/8 b - - 3 34",
+        "8/3r2pp/2bk4/R1P1p3/4P3/6P1/4BK1P/8 b - - 0 35",
+        "8/2kr2pp/2b5/R1P1p3/4P3/4K1P1/4B2P/8 b - - 2 36",
+        "1k6/3r2pp/2b5/RBP1p3/4P3/4K1P1/7P/8 b - - 4 37",
+        "8/1k1r2pp/2b5/R1P1p3/4P3/3BK1P1/7P/8 b - - 6 38",
+        "1k6/3r2pp/2b5/2P1p3/4P3/3BK1P1/7P/R7 b - - 8 39",
+        "1k6/r5pp/2b5/2P1p3/4P3/3BK1P1/7P/5R2 b - - 10 40",
+        "1k3R2/6pp/2b5/2P1p3/4P3/r2BK1P1/7P/8 b - - 12 41",
+        "5R2/2k3pp/2b5/2P1p3/4P3/r2B2P1/3K3P/8 b - - 14 42",
+        "5R2/2k3pp/2b5/2P1p3/4P3/3BK1P1/r6P/8 b - - 16 43",
+        "5R2/2k3pp/2b5/2P1p3/4P3/r2B2P1/4K2P/8 b - - 18 44",
+        "5R2/2k3pp/2b5/2P1p3/4P3/3B1KP1/r6P/8 b - - 20 45",
+        "8/2k2Rpp/2b5/2P1p3/4P3/r2B1KP1/7P/8 b - - 22 46",
+        "3k4/5Rpp/2b5/2P1p3/4P3/r2B2P1/4K2P/8 b - - 24 47",
+        "3k4/5Rpp/2b5/2P1p3/4P3/3B1KP1/r6P/8 b - - 26 48",
+        "3k4/5Rpp/2b5/2P1p3/4P3/r2B2P1/4K2P/8 b - - 28 49",
+        "3k4/5Rpp/2b5/2P1p3/4P3/3BK1P1/r6P/8 b - - 30 50",
+        "3k4/5Rpp/2b5/2P1p3/4P3/r2B2P1/3K3P/8 b - - 32 51",
+        "3k4/5Rpp/2b5/2P1p3/4P3/2KB2P1/r6P/8 b - - 34 52",
+        "3k4/5Rpp/2b5/2P1p3/4P3/r2B2P1/2K4P/8 b - - 36 53",
+        "3k4/5Rpp/2b5/2P1p3/4P3/1K1B2P1/r6P/8 b - - 38 54",
+        "3k4/6Rp/2b5/2P1p3/4P3/1K1B2P1/7r/8 b - - 0 55",
+        "3k4/8/2b3Rp/2P1p3/4P3/1K1B2P1/7r/8 b - - 1 56",
+        "8/2k3R1/2b4p/2P1p3/4P3/1K1B2P1/7r/8 b - - 3 57",
+        "3k4/8/2b3Rp/2P1p3/4P3/1K1B2P1/7r/8 b - - 5 58",
+        "8/2k5/2b3Rp/2P1p3/1K2P3/3B2P1/7r/8 b - - 7 59",
+        "8/2k5/2b3Rp/2P1p3/4P3/2KB2P1/3r4/8 b - - 9 60",
+        "8/2k5/2b3Rp/2P1p3/1K2P3/3B2P1/6r1/8 b - - 11 61",
+        "8/2k5/2b3Rp/2P1p3/4P3/2KB2P1/3r4/8 b - - 13 62",
+        "8/2k5/2b3Rp/2P1p3/2K1P3/3B2P1/6r1/8 b - - 15 63",
+        "4b3/2k3R1/7p/2P1p3/2K1P3/3B2P1/6r1/8 b - - 17 64",
+    },
+    {
+        "r1bqkbnr/npp1pppp/p7/3P4/4pB2/2N5/PPP2PPP/R2QKBNR w KQkq - 1 6",
+        "r1bqkb1r/npp1pppp/p4n2/3P4/4pB2/2N5/PPP1QPPP/R3KBNR w KQkq - 3 7",
+        "r2qkb1r/npp1pppp/p4n2/3P1b2/4pB2/2N5/PPP1QPPP/2KR1BNR w kq - 5 8",
+        "r2qkb1r/1pp1pppp/p4n2/1n1P1b2/4pB2/2N4P/PPP1QPP1/2KR1BNR w kq - 1 9",
+        "r2qkb1r/1pp1pppp/5n2/1p1P1b2/4pB2/7P/PPP1QPP1/2KR1BNR w kq - 0 10",
+        "r2qkb1r/1ppbpppp/5n2/1Q1P4/4pB2/7P/PPP2PP1/2KR1BNR w kq - 1 11",
+        "3qkb1r/1Qpbpppp/5n2/3P4/4pB2/7P/rPP2PP1/2KR1BNR w k - 0 12",
+        "q3kb1r/1Qpbpppp/5n2/3P4/4pB2/7P/rPP2PP1/1K1R1BNR w k - 2 13",
+        "r3kb1r/2pbpppp/5n2/3P4/4pB2/7P/1PP2PP1/1K1R1BNR w k - 0 14",
+        "r3kb1r/2Bb1ppp/4pn2/3P4/4p3/7P/1PP2PP1/1K1R1BNR w k - 0 15",
+        "r3kb1r/2Bb2pp/4pn2/8/4p3/7P/1PP2PP1/1K1R1BNR w k - 0 16",
+        "r3k2r/2Bb2pp/4pn2/2b5/4p3/7P/1PP1NPP1/1K1R1B1R w k - 2 17",
+        "r6r/2Bbk1pp/4pn2/2b5/3Np3/7P/1PP2PP1/1K1R1B1R w - - 4 18",
+        "r6r/b2bk1pp/4pn2/4B3/3Np3/7P/1PP2PP1/1K1R1B1R w - - 6 19",
+        "r1r5/b2bk1pp/4pn2/4B3/2BNp3/7P/1PP2PP1/1K1R3R w - - 8 20",
+        "r7/b2bk1pp/4pn2/2r1B3/2BNp3/1P5P/2P2PP1/1K1R3R w - - 1 21",
+        "rb6/3bk1pp/4pn2/2r1B3/2BNpP2/1P5P/2P3P1/1K1R3R w - - 1 22",
+        "1r6/3bk1pp/4pn2/2r5/2BNpP2/1P5P/2P3P1/1K1R3R w - - 0 23",
+        "1r6/3bk1p1/4pn1p/2r5/2BNpP2/1P5P/2P3P1/2KR3R w - - 0 24",
+        "8/3bk1p1/1r2pn1p/2r5/2BNpP1P/1P6/2P3P1/2KR3R w - - 1 25",
+        "8/3bk3/1r2pnpp/2r5/2BNpP1P/1P6/2P3P1/2K1R2R w - - 0 26",
+        "2b5/4k3/1r2pnpp/2r5/2BNpP1P/1P4P1/2P5/2K1R2R w - - 1 27",
+        "8/1b2k3/1r2pnpp/2r5/2BNpP1P/1P4P1/2P5/2K1R1R1 w - - 3 28",
+        "8/1b1nk3/1r2p1pp/2r5/2BNpPPP/1P6/2P5/2K1R1R1 w - - 1 29",
+        "8/1b2k3/1r2p1pp/2r1nP2/2BNp1PP/1P6/2P5/2K1R1R1 w - - 1 30",
+        "8/1b2k3/1r2p1p1/2r1nPp1/2BNp2P/1P6/2P5/2K1R1R1 w - - 0 31",
+        "8/1b2k3/1r2p1n1/2r3p1/2BNp2P/1P6/2P5/2K1R1R1 w - - 0 32",
+        "8/1b2k3/1r2p1n1/6r1/2BNp2P/1P6/2P5/2K1R3 w - - 0 33",
+        "8/1b2k3/1r2p3/4n1P1/2BNp3/1P6/2P5/2K1R3 w - - 1 34",
+        "8/1b2k3/1r2p3/4n1P1/2BN4/1P2p3/2P5/2K4R w - - 0 35",
+        "8/1b2k3/1r2p2R/6P1/2nN4/1P2p3/2P5/2K5 w - - 0 36",
+        "8/1b2k3/3rp2R/6P1/2PN4/4p3/2P5/2K5 w - - 1 37",
+        "8/4k3/3rp2R/6P1/2PN4/2P1p3/6b1/2K5 w - - 1 38",
+        "8/4k3/r3p2R/2P3P1/3N4/2P1p3/6b1/2K5 w - - 1 39",
+        "8/3k4/r3p2R/2P2NP1/8/2P1p3/6b1/2K5 w - - 3 40",
+        "8/3k4/4p2R/2P3P1/8/2P1N3/6b1/r1K5 w - - 1 41",
+        "8/3k4/4p2R/2P3P1/8/2P1N3/3K2b1/6r1 w - - 3 42",
+        "8/3k4/4p2R/2P3P1/8/2PKNb2/8/6r1 w - - 5 43",
+        "8/4k3/4p1R1/2P3P1/8/2PKNb2/8/6r1 w - - 7 44",
+        "8/4k3/4p1R1/2P3P1/3K4/2P1N3/8/6rb w - - 9 45",
+        "8/3k4/4p1R1/2P1K1P1/8/2P1N3/8/6rb w - - 11 46",
+        "8/3k4/4p1R1/2P3P1/5K2/2P1N3/8/4r2b w - - 13 47",
+        "8/3k4/2b1p2R/2P3P1/5K2/2P1N3/8/4r3 w - - 15 48",
+        "8/3k4/2b1p3/2P3P1/5K2/2P1N2R/8/6r1 w - - 17 49",
+        "2k5/7R/2b1p3/2P3P1/5K2/2P1N3/8/6r1 w - - 19 50",
+        "2k5/7R/4p3/2P3P1/b1P2K2/4N3/8/6r1 w - - 1 51",
+        "2k5/3bR3/4p3/2P3P1/2P2K2/4N3/8/6r1 w - - 3 52",
+        "3k4/3b2R1/4p3/2P3P1/2P2K2/4N3/8/6r1 w - - 5 53",
+        "3kb3/6R1/4p1P1/2P5/2P2K2/4N3/8/6r1 w - - 1 54",
+        "3kb3/6R1/4p1P1/2P5/2P2KN1/8/8/2r5 w - - 3 55",
+        "3kb3/6R1/4p1P1/2P1N3/2P2K2/8/8/5r2 w - - 5 56",
+        "3kb3/6R1/4p1P1/2P1N3/2P5/4K3/8/4r3 w - - 7 57",
+    },
+    {
+        "rnbq1rk1/ppp1npb1/4p1p1/3P3p/3PP3/2N2N2/PP2BPPP/R1BQ1RK1 b - - 0 8",
+        "rnbq1rk1/ppp1npb1/6p1/3pP2p/3P4/2N2N2/PP2BPPP/R1BQ1RK1 b - - 0 9",
+        "rn1q1rk1/ppp1npb1/6p1/3pP2p/3P2b1/2N2N2/PP2BPPP/R1BQR1K1 b - - 2 10",
+        "r2q1rk1/ppp1npb1/2n3p1/3pP2p/3P2bN/2N5/PP2BPPP/R1BQR1K1 b - - 4 11",
+        "r4rk1/pppqnpb1/2n3p1/3pP2p/3P2bN/2N4P/PP2BPP1/R1BQR1K1 b - - 0 12",
+        "r4rk1/pppqnpb1/2n3p1/3pP2p/3P3N/7P/PP2NPP1/R1BQR1K1 b - - 0 13",
+        "r4rk1/pppq1pb1/2n3p1/3pPN1p/3P4/7P/PP2NPP1/R1BQR1K1 b - - 0 14",
+        "r4rk1/ppp2pb1/2n3p1/3pPq1p/3P1N2/7P/PP3PP1/R1BQR1K1 b - - 1 15",
+        "r4rk1/pppq1pb1/2n3p1/3pP2p/P2P1N2/7P/1P3PP1/R1BQR1K1 b - - 0 16",
+        "r2n1rk1/pppq1pb1/6p1/3pP2p/P2P1N2/R6P/1P3PP1/2BQR1K1 b - - 2 17",
+        "r4rk1/pppq1pb1/4N1p1/3pP2p/P2P4/R6P/1P3PP1/2BQR1K1 b - - 0 18",
+        "r4rk1/ppp2pb1/4q1p1/3pP1Bp/P2P4/R6P/1P3PP1/3QR1K1 b - - 1 19",
+        "r3r1k1/ppp2pb1/4q1p1/3pP1Bp/P2P1P2/R6P/1P4P1/3QR1K1 b - - 0 20",
+        "r3r1k1/ppp3b1/4qpp1/3pP2p/P2P1P1B/R6P/1P4P1/3QR1K1 b - - 1 21",
+        "r3r1k1/ppp3b1/4q1p1/3pP2p/P4P1B/R6P/1P4P1/3QR1K1 b - - 0 22",
+        "r4rk1/ppp3b1/4q1p1/3pP1Bp/P4P2/R6P/1P4P1/3QR1K1 b - - 2 23",
+        "r4rk1/pp4b1/4q1p1/2ppP1Bp/P4P2/3R3P/1P4P1/3QR1K1 b - - 1 24",
+        "r4rk1/pp4b1/4q1p1/2p1P1Bp/P2p1PP1/3R3P/1P6/3QR1K1 b - - 0 25",
+        "r4rk1/pp4b1/4q1p1/2p1P1B1/P2p1PP1/3R4/1P6/3QR1K1 b - - 0 26",
+        "r5k1/pp3rb1/4q1p1/2p1P1B1/P2p1PP1/6R1/1P6/3QR1K1 b - - 2 27",
+        "5rk1/pp3rb1/4q1p1/2p1P1B1/P2pRPP1/6R1/1P6/3Q2K1 b - - 4 28",
+        "5rk1/1p3rb1/p3q1p1/P1p1P1B1/3pRPP1/6R1/1P6/3Q2K1 b - - 0 29",
+        "4r1k1/1p3rb1/p3q1p1/P1p1P1B1/3pRPP1/1P4R1/8/3Q2K1 b - - 0 30",
+        "4r1k1/5rb1/pP2q1p1/2p1P1B1/3pRPP1/1P4R1/8/3Q2K1 b - - 0 31",
+        "4r1k1/5rb1/pq4p1/2p1P1B1/3pRPP1/1P4R1/4Q3/6K1 b - - 1 32",
+        "4r1k1/1r4b1/pq4p1/2p1P1B1/3pRPP1/1P4R1/2Q5/6K1 b - - 3 33",
+        "4r1k1/1r4b1/1q4p1/p1p1P1B1/3p1PP1/1P4R1/2Q5/4R1K1 b - - 1 34",
+        "4r1k1/3r2b1/1q4p1/p1p1P1B1/2Qp1PP1/1P4R1/8/4R1K1 b - - 3 35",
+        "4r1k1/3r2b1/4q1p1/p1p1P1B1/2Qp1PP1/1P4R1/5K2/4R3 b - - 5 36",
+        "4r1k1/3r2b1/6p1/p1p1P1B1/2Pp1PP1/6R1/5K2/4R3 b - - 0 37",
+        "4r1k1/3r2b1/6p1/p1p1P1B1/2P2PP1/3p2R1/5K2/3R4 b - - 1 38",
+        "5rk1/3r2b1/6p1/p1p1P1B1/2P2PP1/3p2R1/8/3RK3 b - - 3 39",
+        "5rk1/6b1/6p1/p1p1P1B1/2Pr1PP1/3R4/8/3RK3 b - - 0 40",
+        "5rk1/3R2b1/6p1/p1p1P1B1/2r2PP1/8/8/3RK3 b - - 1 41",
+        "5rk1/3R2b1/6p1/p1p1P1B1/4rPP1/8/3K4/3R4 b - - 3 42",
+        "1r4k1/3R2b1/6p1/p1p1P1B1/4rPP1/2K5/8/3R4 b - - 5 43",
+        "1r4k1/3R2b1/6p1/p1p1P1B1/2K2PP1/4r3/8/3R4 b - - 7 44",
+        "1r3bk1/8/3R2p1/p1p1P1B1/2K2PP1/4r3/8/3R4 b - - 9 45",
+        "1r3bk1/8/6R1/2p1P1B1/p1K2PP1/4r3/8/3R4 b - - 0 46",
+        "1r3b2/5k2/R7/2p1P1B1/p1K2PP1/4r3/8/3R4 b - - 2 47",
+        "5b2/1r3k2/R7/2p1P1B1/p1K2PP1/4r3/8/7R b - - 4 48",
+        "5b2/5k2/R7/2pKP1B1/pr3PP1/4r3/8/7R b - - 6 49",
+        "5b2/5k2/R1K5/2p1P1B1/p2r1PP1/4r3/8/7R b - - 8 50",
+        "8/R4kb1/2K5/2p1P1B1/p2r1PP1/4r3/8/7R b - - 10 51",
+        "8/R5b1/2K3k1/2p1PPB1/p2r2P1/4r3/8/7R b - - 0 52",
+        "8/6R1/2K5/2p1PPk1/p2r2P1/4r3/8/7R b - - 0 53",
+        "8/6R1/2K5/2p1PP2/p2r1kP1/4r3/8/5R2 b - - 2 54",
+        "8/6R1/2K2P2/2p1P3/p2r2P1/4r1k1/8/5R2 b - - 0 55",
+        "8/5PR1/2K5/2p1P3/p2r2P1/4r3/6k1/5R2 b - - 0 56",
+    },
+    {
+        "rn1qkb1r/p1pbpppp/5n2/8/2pP4/2N5/1PQ1PPPP/R1B1KBNR w KQkq - 0 7",
+        "r2qkb1r/p1pbpppp/2n2n2/8/2pP4/2N2N2/1PQ1PPPP/R1B1KB1R w KQkq - 2 8",
+        "r2qkb1r/p1pbpppp/5n2/8/1npPP3/2N2N2/1PQ2PPP/R1B1KB1R w KQkq - 1 9",
+        "r2qkb1r/p1pb1ppp/4pn2/8/1npPP3/2N2N2/1P3PPP/R1BQKB1R w KQkq - 0 10",
+        "r2qk2r/p1pbbppp/4pn2/8/1nBPP3/2N2N2/1P3PPP/R1BQK2R w KQkq - 1 11",
+        "r2q1rk1/p1pbbppp/4pn2/8/1nBPP3/2N2N2/1P3PPP/R1BQ1RK1 w - - 3 12",
+        "r2q1rk1/2pbbppp/p3pn2/8/1nBPPB2/2N2N2/1P3PPP/R2Q1RK1 w - - 0 13",
+        "r2q1rk1/2p1bppp/p3pn2/1b6/1nBPPB2/2N2N2/1P3PPP/R2QR1K1 w - - 2 14",
+        "r2q1rk1/4bppp/p1p1pn2/1b6/1nBPPB2/1PN2N2/5PPP/R2QR1K1 w - - 0 15",
+        "r4rk1/3qbppp/p1p1pn2/1b6/1nBPPB2/1PN2N2/3Q1PPP/R3R1K1 w - - 2 16",
+        "r4rk1/1q2bppp/p1p1pn2/1b6/1nBPPB2/1PN2N1P/3Q1PP1/R3R1K1 w - - 1 17",
+        "r3r1k1/1q2bppp/p1p1pn2/1b6/1nBPPB2/1PN2N1P/4QPP1/R3R1K1 w - - 3 18",
+        "r3r1k1/1q1nbppp/p1p1p3/1b6/1nBPPB2/1PN2N1P/4QPP1/3RR1K1 w - - 5 19",
+        "r3rbk1/1q1n1ppp/p1p1p3/1b6/1nBPPB2/1PN2N1P/3RQPP1/4R1K1 w - - 7 20",
+        "r3rbk1/1q3ppp/pnp1p3/1b6/1nBPPB2/1PN2N1P/3RQPP1/4R2K w - - 9 21",
+        "2r1rbk1/1q3ppp/pnp1p3/1b6/1nBPPB2/1PN2N1P/3RQPP1/1R5K w - - 11 22",
+        "2r1rbk1/1q4pp/pnp1pp2/1b6/1nBPPB2/1PN2N1P/4QPP1/1R1R3K w - - 0 23",
+        "2r1rbk1/5qpp/pnp1pp2/1b6/1nBPP3/1PN1BN1P/4QPP1/1R1R3K w - - 2 24",
+        "2r1rbk1/5qp1/pnp1pp1p/1b6/1nBPP3/1PN1BN1P/4QPP1/1R1R2K1 w - - 0 25",
+        "2r1rbk1/5qp1/pnp1pp1p/1b6/2BPP3/1P2BN1P/n3QPP1/1R1R2K1 w - - 0 26",
+        "r3rbk1/5qp1/pnp1pp1p/1b6/2BPP3/1P2BN1P/Q4PP1/1R1R2K1 w - - 1 27",
+        "rr3bk1/5qp1/pnp1pp1p/1b6/2BPP3/1P2BN1P/Q4PP1/R2R2K1 w - - 3 28",
+        "rr2qbk1/6p1/pnp1pp1p/1b6/2BPP3/1P2BN1P/4QPP1/R2R2K1 w - - 5 29",
+        "rr2qbk1/6p1/1np1pp1p/pb6/2BPP3/1P1QBN1P/5PP1/R2R2K1 w - - 0 30",
+        "rr2qbk1/6p1/1n2pp1p/pp6/3PP3/1P1QBN1P/5PP1/R2R2K1 w - - 0 31",
+        "rr2qbk1/6p1/1n2pp1p/1p1P4/p3P3/1P1QBN1P/5PP1/R2R2K1 w - - 0 32",
+        "rr2qbk1/3n2p1/3Ppp1p/1p6/p3P3/1P1QBN1P/5PP1/R2R2K1 w - - 1 33",
+        "rr3bk1/3n2p1/3Ppp1p/1p5q/pP2P3/3QBN1P/5PP1/R2R2K1 w - - 1 34",
+        "rr3bk1/3n2p1/3Ppp1p/1p5q/1P2P3/p2QBN1P/5PP1/2RR2K1 w - - 0 35",
+        "1r3bk1/3n2p1/r2Ppp1p/1p5q/1P2P3/pQ2BN1P/5PP1/2RR2K1 w - - 2 36",
+        "1r2qbk1/2Rn2p1/r2Ppp1p/1p6/1P2P3/pQ2BN1P/5PP1/3R2K1 w - - 4 37",
+        "1r2qbk1/2Rn2p1/r2Ppp1p/1pB5/1P2P3/1Q3N1P/p4PP1/3R2K1 w - - 0 38",
+        "1r2q1k1/2Rn2p1/r2bpp1p/1pB5/1P2P3/1Q3N1P/p4PP1/R5K1 w - - 0 39",
+        "1r2q1k1/2Rn2p1/3rpp1p/1p6/1P2P3/1Q3N1P/p4PP1/R5K1 w - - 0 40",
+        "2r1q1k1/2Rn2p1/3rpp1p/1p6/1P2P3/5N1P/Q4PP1/R5K1 w - - 1 41",
+        "1r2q1k1/1R1n2p1/3rpp1p/1p6/1P2P3/5N1P/Q4PP1/R5K1 w - - 3 42",
+        "2r1q1k1/2Rn2p1/3rpp1p/1p6/1P2P3/5N1P/Q4PP1/R5K1 w - - 5 43",
+        "1r2q1k1/1R1n2p1/3rpp1p/1p6/1P2P3/5N1P/Q4PP1/R5K1 w - - 7 44",
+        "1rq3k1/R2n2p1/3rpp1p/1p6/1P2P3/5N1P/Q4PP1/R5K1 w - - 9 45",
+        "2q3k1/Rr1n2p1/3rpp1p/1p6/1P2P3/5N1P/4QPP1/R5K1 w - - 11 46",
+        "Rrq3k1/3n2p1/3rpp1p/1p6/1P2P3/5N1P/4QPP1/R5K1 w - - 13 47",
+    },
+    {
+        "rn1qkb1r/1pp2ppp/p4p2/3p1b2/5P2/1P2PN2/P1PP2PP/RN1QKB1R b KQkq - 1 6",
+        "r2qkb1r/1pp2ppp/p1n2p2/3p1b2/3P1P2/1P2PN2/P1P3PP/RN1QKB1R b KQkq - 0 7",
+        "r2qkb1r/1pp2ppp/p4p2/3p1b2/1n1P1P2/1P1BPN2/P1P3PP/RN1QK2R b KQkq - 2 8",
+        "r2qkb1r/1pp2ppp/p4p2/3p1b2/3P1P2/1P1PPN2/P5PP/RN1QK2R b KQkq - 0 9",
+        "r2qk2r/1pp2ppp/p2b1p2/3p1b2/3P1P2/1PNPPN2/P5PP/R2QK2R b KQkq - 2 10",
+        "r2qk2r/1p3ppp/p1pb1p2/3p1b2/3P1P2/1PNPPN2/P5PP/R2Q1RK1 b kq - 1 11",
+        "r2q1rk1/1p3ppp/p1pb1p2/3p1b2/3P1P2/1PNPPN2/P2Q2PP/R4RK1 b - - 3 12",
+        "r2qr1k1/1p3ppp/p1pb1p2/3p1b2/3P1P2/1P1PPN2/P2QN1PP/R4RK1 b - - 5 13",
+        "r3r1k1/1p3ppp/pqpb1p2/3p1b2/3P1P2/1P1PPNN1/P2Q2PP/R4RK1 b - - 7 14",
+        "r3r1k1/1p3ppp/pqp2p2/3p1b2/1b1P1P2/1P1PPNN1/P1Q3PP/R4RK1 b - - 9 15",
+        "r3r1k1/1p1b1ppp/pqp2p2/3p4/1b1P1P2/1P1PPNN1/P4QPP/R4RK1 b - - 11 16",
+        "2r1r1k1/1p1b1ppp/pqp2p2/3p4/1b1PPP2/1P1P1NN1/P4QPP/R4RK1 b - - 0 17",
+        "2r1r1k1/1p1b1ppp/pq3p2/2pp4/1b1PPP2/PP1P1NN1/5QPP/R4RK1 b - - 0 18",
+        "2r1r1k1/1p1b1ppp/pq3p2/2Pp4/4PP2/PPbP1NN1/5QPP/R4RK1 b - - 0 19",
+        "2r1r1k1/1p1b1ppp/p4p2/2Pp4/4PP2/PqbP1NN1/5QPP/RR4K1 b - - 1 20",
+        "2r1r1k1/1p1b1ppp/p4p2/2Pp4/q3PP2/P1bP1NN1/R4QPP/1R4K1 b - - 3 21",
+        "2r1r1k1/1p3ppp/p4p2/1bPP4/q4P2/P1bP1NN1/R4QPP/1R4K1 b - - 0 22",
+        "2r1r1k1/1p3ppp/p4p2/2PP4/q4P2/P1bb1NN1/R4QPP/2R3K1 b - - 1 23",
+        "2r1r1k1/1p3ppp/p2P1p2/2P5/2q2P2/P1bb1NN1/R4QPP/2R3K1 b - - 0 24",
+        "2rr2k1/1p3ppp/p2P1p2/2P5/2q2P2/P1bb1NN1/R4QPP/2R4K b - - 2 25",
+        "2rr2k1/1p3ppp/p2P1p2/2Q5/5P2/P1bb1NN1/R5PP/2R4K b - - 0 26",
+        "3r2k1/1p3ppp/p2P1p2/2r5/5P2/P1bb1N2/R3N1PP/2R4K b - - 1 27",
+        "3r2k1/1p3ppp/p2P1p2/2r5/5P2/P1b2N2/4R1PP/2R4K b - - 0 28",
+        "3r2k1/1p3ppp/p2P1p2/2r5/1b3P2/P4N2/4R1PP/3R3K b - - 2 29",
+        "3r2k1/1p2Rppp/p2P1p2/b1r5/5P2/P4N2/6PP/3R3K b - - 4 30",
+        "3r2k1/1R3ppp/p1rP1p2/b7/5P2/P4N2/6PP/3R3K b - - 0 31",
+        "3r2k1/1R3ppp/p2R1p2/b7/5P2/P4N2/6PP/7K b - - 0 32",
+        "6k1/1R3ppp/p2r1p2/b7/5P2/P4NP1/7P/7K b - - 0 33",
+        "6k1/1R3p1p/p2r1pp1/b7/5P1P/P4NP1/8/7K b - - 0 34",
+        "6k1/3R1p1p/pr3pp1/b7/5P1P/P4NP1/8/7K b - - 2 35",
+        "6k1/5p2/pr3pp1/b2R3p/5P1P/P4NP1/8/7K b - - 1 36",
+        "6k1/5p2/pr3pp1/7p/5P1P/P1bR1NP1/8/7K b - - 3 37",
+        "6k1/5p2/p1r2pp1/7p/5P1P/P1bR1NP1/6K1/8 b - - 5 38",
+        "6k1/5p2/p1r2pp1/b2R3p/5P1P/P4NP1/6K1/8 b - - 7 39",
+        "6k1/5p2/p4pp1/b2R3p/5P1P/P4NPK/2r5/8 b - - 9 40",
+        "6k1/2b2p2/p4pp1/7p/5P1P/P2R1NPK/2r5/8 b - - 11 41",
+        "6k1/2b2p2/5pp1/p6p/3N1P1P/P2R2PK/2r5/8 b - - 1 42",
+        "6k1/2b2p2/5pp1/p6p/3N1P1P/P1R3PK/r7/8 b - - 3 43",
+        "6k1/5p2/1b3pp1/p6p/5P1P/P1R3PK/r1N5/8 b - - 5 44",
+        "8/5pk1/1bR2pp1/p6p/5P1P/P5PK/r1N5/8 b - - 7 45",
+        "3b4/5pk1/2R2pp1/p4P1p/7P/P5PK/r1N5/8 b - - 0 46",
+        "8/4bpk1/2R2pp1/p4P1p/6PP/P6K/r1N5/8 b - - 0 47",
+        "8/5pk1/2R2pP1/p6p/6PP/b6K/r1N5/8 b - - 0 48",
+        "8/6k1/2R2pp1/p6P/7P/b6K/r1N5/8 b - - 0 49",
+        "8/6k1/2R2p2/p6p/7P/b5K1/r1N5/8 b - - 1 50",
+        "8/8/2R2pk1/p6p/7P/b4K2/r1N5/8 b - - 3 51",
+        "8/8/2R2pk1/p6p/7P/4NK2/rb6/8 b - - 5 52",
+        "2R5/8/5pk1/7p/p6P/4NK2/rb6/8 b - - 1 53",
+        "6R1/8/5pk1/7p/p6P/4NK2/1b6/r7 b - - 3 54",
+        "R7/5k2/5p2/7p/p6P/4NK2/1b6/r7 b - - 5 55",
+        "R7/5k2/5p2/7p/7P/p3N3/1b2K3/r7 b - - 1 56",
+        "8/R4k2/5p2/7p/7P/p3N3/1b2K3/7r b - - 3 57",
+        "8/8/5pk1/7p/R6P/p3N3/1b2K3/7r b - - 5 58",
+        "8/8/5pk1/7p/R6P/p7/4K3/2bN3r b - - 7 59",
+        "8/8/5pk1/7p/R6P/p7/4KN1r/2b5 b - - 9 60",
+        "8/8/5pk1/7p/R6P/p3K3/1b3N1r/8 b - - 11 61",
+        "8/8/R4pk1/7p/7P/p1b1K3/5N1r/8 b - - 13 62",
+        "8/8/5pk1/7p/7P/2b1K3/R4N1r/8 b - - 0 63",
+        "8/8/5pk1/7p/3K3P/8/R4N1r/4b3 b - - 2 64",
+    }
+};
+// clang-format on
+
 }  // namespace
 
 namespace Stockfish::Benchmark {
@@ -160,4 +437,76 @@ std::vector<std::string> setup_bench(const std::string& currentFen, std::istream
     return list;
 }
 
+BenchmarkSetup setup_benchmark(std::istream& is) {
+    // TT_SIZE_PER_THREAD is chosen such that roughly half of the hash is used all positions
+    // for the current sequence have been searched.
+    static constexpr int TT_SIZE_PER_THREAD = 128;
+
+    static constexpr int DEFAULT_DURATION_S = 150;
+
+    BenchmarkSetup setup{};
+
+    // Assign default values to missing arguments
+    int desiredTimeS;
+
+    if (!(is >> setup.threads))
+        setup.threads = get_hardware_concurrency();
+    else
+        setup.originalInvocation += std::to_string(setup.threads);
+
+    if (!(is >> setup.ttSize))
+        setup.ttSize = TT_SIZE_PER_THREAD * setup.threads;
+    else
+        setup.originalInvocation += " " + std::to_string(setup.ttSize);
+
+    if (!(is >> desiredTimeS))
+        desiredTimeS = DEFAULT_DURATION_S;
+    else
+        setup.originalInvocation += " " + std::to_string(desiredTimeS);
+
+    setup.filledInvocation += std::to_string(setup.threads) + " " + std::to_string(setup.ttSize)
+                            + " " + std::to_string(desiredTimeS);
+
+    auto getCorrectedTime = [&](int ply) {
+        // time per move is fit roughly based on LTC games
+        // seconds = 50/{ply+15}
+        // ms = 50000/{ply+15}
+        // with this fit 10th move gets 2000ms
+        // adjust for desired 10th move time
+        return 50000.0 / (static_cast<double>(ply) + 15.0);
+    };
+
+    float totalTime = 0;
+    for (const auto& game : BenchmarkPositions)
+    {
+        setup.commands.emplace_back("ucinewgame");
+        int ply = 1;
+        for (int i = 0; i < static_cast<int>(game.size()); ++i)
+        {
+            const float correctedTime = getCorrectedTime(ply);
+            totalTime += correctedTime;
+            ply += 1;
+        }
+    }
+
+    float timeScaleFactor = static_cast<float>(desiredTimeS * 1000) / totalTime;
+
+    for (const auto& game : BenchmarkPositions)
+    {
+        setup.commands.emplace_back("ucinewgame");
+        int ply = 1;
+        for (const std::string& fen : game)
+        {
+            setup.commands.emplace_back("position fen " + fen);
+
+            const int correctedTime = static_cast<int>(getCorrectedTime(ply) * timeScaleFactor);
+            setup.commands.emplace_back("go movetime " + std::to_string(correctedTime));
+
+            ply += 1;
+        }
+    }
+
+    return setup;
+}
+
 }  // namespace Stockfish
\ No newline at end of file
diff --git a/src/benchmark.h b/src/benchmark.h
index b1eba40f..eb3a52d8 100644
--- a/src/benchmark.h
+++ b/src/benchmark.h
@@ -27,6 +27,16 @@ namespace Stockfish::Benchmark {
 
 std::vector<std::string> setup_bench(const std::string&, std::istream&);
 
+struct BenchmarkSetup {
+    int                      ttSize;
+    int                      threads;
+    std::vector<std::string> commands;
+    std::string              originalInvocation;
+    std::string              filledInvocation;
+};
+
+BenchmarkSetup setup_benchmark(std::istream&);
+
 }  // namespace Stockfish
 
 #endif  // #ifndef BENCHMARK_H_INCLUDED
diff --git a/src/bitboard.cpp b/src/bitboard.cpp
index 32c626d4..deda6da2 100644
--- a/src/bitboard.cpp
+++ b/src/bitboard.cpp
@@ -34,15 +34,14 @@ Bitboard BetweenBB[SQUARE_NB][SQUARE_NB];
 Bitboard PseudoAttacks[PIECE_TYPE_NB][SQUARE_NB];
 Bitboard PawnAttacks[COLOR_NB][SQUARE_NB];
 
-Magic RookMagics[SQUARE_NB];
-Magic BishopMagics[SQUARE_NB];
+alignas(64) Magic Magics[SQUARE_NB][2];
 
 namespace {
 
 Bitboard RookTable[0x19000];   // To store rook attacks
 Bitboard BishopTable[0x1480];  // To store bishop attacks
 
-void init_magics(PieceType pt, Bitboard table[], Magic magics[]);
+void init_magics(PieceType pt, Bitboard table[], Magic magics[][2]);
 
 // Returns the bitboard of target square for the given step
 // from the given square. If the step is off the board, returns empty bitboard.
@@ -82,8 +81,8 @@ void Bitboards::init() {
         for (Square s2 = SQ_A1; s2 <= SQ_H8; ++s2)
             SquareDistance[s1][s2] = std::max(distance<File>(s1, s2), distance<Rank>(s1, s2));
 
-    init_magics(ROOK, RookTable, RookMagics);
-    init_magics(BISHOP, BishopTable, BishopMagics);
+    init_magics(ROOK, RookTable, Magics);
+    init_magics(BISHOP, BishopTable, Magics);
 
     for (Square s1 = SQ_A1; s1 <= SQ_H8; ++s1)
     {
@@ -124,8 +123,14 @@ Bitboard sliding_attack(PieceType pt, Square sq, Bitboard occupied) {
     for (Direction d : (pt == ROOK ? RookDirections : BishopDirections))
     {
         Square s = sq;
-        while (safe_destination(s, d) && !(occupied & s))
+        while (safe_destination(s, d))
+        {
             attacks |= (s += d);
+            if (occupied & s)
+            {
+                break;
+            }
+        }
     }
 
     return attacks;
@@ -134,41 +139,49 @@ Bitboard sliding_attack(PieceType pt, Square sq, Bitboard occupied) {
 
 // Computes all rook and bishop attacks at startup. Magic
 // bitboards are used to look up attacks of sliding pieces. As a reference see
-// www.chessprogramming.org/Magic_Bitboards. In particular, here we use the so
-// called "fancy" approach.
-void init_magics(PieceType pt, Bitboard table[], Magic magics[]) {
+// https://www.chessprogramming.org/Magic_Bitboards. In particular, here we use
+// the so called "fancy" approach.
+void init_magics(PieceType pt, Bitboard table[], Magic magics[][2]) {
 
+#ifndef USE_PEXT
     // Optimal PRNG seeds to pick the correct magics in the shortest time
     int seeds[][RANK_NB] = {{8977, 44560, 54343, 38998, 5731, 95205, 104912, 17020},
                             {728, 10316, 55013, 32803, 12281, 15100, 16645, 255}};
 
-    Bitboard occupancy[4096], reference[4096], edges, b;
-    int      epoch[4096] = {}, cnt = 0, size = 0;
+    Bitboard occupancy[4096];
+    int      epoch[4096] = {}, cnt = 0;
+#endif
+    Bitboard reference[4096];
+    int      size = 0;
 
     for (Square s = SQ_A1; s <= SQ_H8; ++s)
     {
         // Board edges are not considered in the relevant occupancies
-        edges = ((Rank1BB | Rank8BB) & ~rank_bb(s)) | ((FileABB | FileHBB) & ~file_bb(s));
+        Bitboard edges = ((Rank1BB | Rank8BB) & ~rank_bb(s)) | ((FileABB | FileHBB) & ~file_bb(s));
 
         // Given a square 's', the mask is the bitboard of sliding attacks from
         // 's' computed on an empty board. The index must be big enough to contain
         // all the attacks for each possible subset of the mask and so is 2 power
         // the number of 1s of the mask. Hence we deduce the size of the shift to
         // apply to the 64 or 32 bits word to get the index.
-        Magic& m = magics[s];
+        Magic& m = magics[s][pt - BISHOP];
         m.mask   = sliding_attack(pt, s, 0) & ~edges;
-        m.shift  = (Is64Bit ? 64 : 32) - popcount(m.mask);
-
+#ifndef USE_PEXT
+        m.shift = (Is64Bit ? 64 : 32) - popcount(m.mask);
+#endif
         // Set the offset for the attacks table of the square. We have individual
         // table sizes for each square with "Fancy Magic Bitboards".
-        m.attacks = s == SQ_A1 ? table : magics[s - 1].attacks + size;
+        m.attacks = s == SQ_A1 ? table : magics[s - 1][pt - BISHOP].attacks + size;
+        size      = 0;
 
         // Use Carry-Rippler trick to enumerate all subsets of masks[s] and
         // store the corresponding sliding attack bitboard in reference[].
-        b = size = 0;
+        Bitboard b = 0;
         do
         {
+#ifndef USE_PEXT
             occupancy[size] = b;
+#endif
             reference[size] = sliding_attack(pt, s, b);
 
             if (HasPext)
@@ -178,9 +191,7 @@ void init_magics(PieceType pt, Bitboard table[], Magic magics[]) {
             b = (b - m.mask) & m.mask;
         } while (b);
 
-        if (HasPext)
-            continue;
-
+#ifndef USE_PEXT
         PRNG rng(seeds[Is64Bit][rank_of(s)]);
 
         // Find a magic for square 's' picking up an (almost) random number
@@ -209,6 +220,7 @@ void init_magics(PieceType pt, Bitboard table[], Magic magics[]) {
                     break;
             }
         }
+#endif
     }
 }
 }
diff --git a/src/bitboard.h b/src/bitboard.h
index cdff4c75..c4bf18b5 100644
--- a/src/bitboard.h
+++ b/src/bitboard.h
@@ -67,27 +67,31 @@ extern Bitboard PawnAttacks[COLOR_NB][SQUARE_NB];
 // Magic holds all magic bitboards relevant data for a single square
 struct Magic {
     Bitboard  mask;
-    Bitboard  magic;
     Bitboard* attacks;
-    unsigned  shift;
+#ifndef USE_PEXT
+    Bitboard magic;
+    unsigned shift;
+#endif
 
     // Compute the attack's index using the 'magic bitboards' approach
     unsigned index(Bitboard occupied) const {
 
-        if (HasPext)
-            return unsigned(pext(occupied, mask));
-
+#ifdef USE_PEXT
+        return unsigned(pext(occupied, mask));
+#else
         if (Is64Bit)
             return unsigned(((occupied & mask) * magic) >> shift);
 
         unsigned lo = unsigned(occupied) & unsigned(mask);
         unsigned hi = unsigned(occupied >> 32) & unsigned(mask >> 32);
         return (lo * unsigned(magic) ^ hi * unsigned(magic >> 32)) >> shift;
+#endif
     }
+
+    Bitboard attacks_bb(Bitboard occupied) const { return attacks[index(occupied)]; }
 };
 
-extern Magic RookMagics[SQUARE_NB];
-extern Magic BishopMagics[SQUARE_NB];
+extern Magic Magics[SQUARE_NB][2];
 
 constexpr Bitboard square_bb(Square s) {
     assert(is_ok(s));
@@ -229,9 +233,8 @@ inline Bitboard attacks_bb(Square s, Bitboard occupied) {
     switch (Pt)
     {
     case BISHOP :
-        return BishopMagics[s].attacks[BishopMagics[s].index(occupied)];
     case ROOK :
-        return RookMagics[s].attacks[RookMagics[s].index(occupied)];
+        return Magics[s][Pt - BISHOP].attacks_bb(occupied);
     case QUEEN :
         return attacks_bb<BISHOP>(s, occupied) | attacks_bb<ROOK>(s, occupied);
     default :
diff --git a/src/engine.cpp b/src/engine.cpp
index e8da24aa..85c84099 100644
--- a/src/engine.cpp
+++ b/src/engine.cpp
@@ -18,15 +18,15 @@
 
 #include "engine.h"
 
+#include <cassert>
 #include <deque>
+#include <iosfwd>
 #include <memory>
 #include <ostream>
+#include <sstream>
 #include <string_view>
 #include <utility>
 #include <vector>
-#include <sstream>
-#include <iosfwd>
-#include <cassert>
 
 #include "evaluate.h"
 #include "misc.h"
@@ -44,16 +44,76 @@ namespace Stockfish {
 
 namespace NN = Eval::NNUE;
 
-constexpr auto StartFEN = "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1";
+constexpr auto StartFEN  = "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1";
+constexpr int  MaxHashMB = Is64Bit ? 33554432 : 2048;
 
-Engine::Engine(std::string path) :
-    binaryDirectory(CommandLine::get_binary_directory(path)),
+Engine::Engine(std::optional<std::string> path) :
+    binaryDirectory(path ? CommandLine::get_binary_directory(*path) : ""),
+    numaContext(NumaConfig::from_system()),
     states(new std::deque<StateInfo>(1)),
-    networks(NN::Networks(
-      NN::NetworkBig({EvalFileDefaultNameBig, "None", ""}, NN::EmbeddedNNUEType::BIG),
-      NN::NetworkSmall({EvalFileDefaultNameSmall, "None", ""}, NN::EmbeddedNNUEType::SMALL))) {
+    threads(),
+    networks(
+      numaContext,
+      NN::Networks(
+        NN::NetworkBig({EvalFileDefaultNameBig, "None", ""}, NN::EmbeddedNNUEType::BIG),
+        NN::NetworkSmall({EvalFileDefaultNameSmall, "None", ""}, NN::EmbeddedNNUEType::SMALL))) {
     pos.set(StartFEN, false, &states->back());
     capSq = SQ_NONE;
+
+    options["Debug Log File"] << Option("", [](const Option& o) {
+        start_logger(o);
+        return std::nullopt;
+    });
+
+    options["NumaPolicy"] << Option("auto", [this](const Option& o) {
+        set_numa_config_from_option(o);
+        return numa_config_information_as_string() + "\n"
+             + thread_allocation_information_as_string();
+    });
+
+    options["Threads"] << Option(1, 1, 1024, [this](const Option&) {
+        resize_threads();
+        return thread_allocation_information_as_string();
+    });
+
+    options["Hash"] << Option(16, 1, MaxHashMB, [this](const Option& o) {
+        set_tt_size(o);
+        return std::nullopt;
+    });
+
+    options["Clear Hash"] << Option([this](const Option&) {
+        search_clear();
+        return std::nullopt;
+    });
+    options["Ponder"] << Option(false);
+    options["MultiPV"] << Option(1, 1, MAX_MOVES);
+    options["Skill Level"] << Option(20, 0, 20);
+    options["Move Overhead"] << Option(10, 0, 5000);
+    options["nodestime"] << Option(0, 0, 10000);
+    options["UCI_Chess960"] << Option(false);
+    options["UCI_LimitStrength"] << Option(false);
+    options["UCI_Elo"] << Option(Stockfish::Search::Skill::LowestElo,
+                                 Stockfish::Search::Skill::LowestElo,
+                                 Stockfish::Search::Skill::HighestElo);
+    options["UCI_ShowWDL"] << Option(false);
+    options["SyzygyPath"] << Option("", [](const Option& o) {
+        Tablebases::init(o);
+        return std::nullopt;
+    });
+    options["SyzygyProbeDepth"] << Option(1, 1, 100);
+    options["Syzygy50MoveRule"] << Option(true);
+    options["SyzygyProbeLimit"] << Option(7, 0, 7);
+    options["EvalFile"] << Option(EvalFileDefaultNameBig, [this](const Option& o) {
+        load_big_network(o);
+        return std::nullopt;
+    });
+    options["EvalFileSmall"] << Option(EvalFileDefaultNameSmall, [this](const Option& o) {
+        load_small_network(o);
+        return std::nullopt;
+    });
+
+    load_networks();
+    resize_threads();
 }
 
 std::uint64_t Engine::perft(const std::string& fen, Depth depth, bool isChess960) {
@@ -74,7 +134,7 @@ void Engine::stop() { threads.stop = true; }
 void Engine::search_clear() {
     wait_for_search_finished();
 
-    tt.clear(options["Threads"]);
+    tt.clear(threads);
     threads.clear();
 
     // @TODO wont work with multiple instances
@@ -97,6 +157,10 @@ void Engine::set_on_bestmove(std::function<void(std::string_view, std::string_vi
     updateContext.onBestmove = std::move(f);
 }
 
+void Engine::set_on_verify_networks(std::function<void(std::string_view)>&& f) {
+    onVerifyNetworks = std::move(f);
+}
+
 void Engine::wait_for_search_finished() { threads.main_thread()->wait_for_search_finished(); }
 
 void Engine::set_position(const std::string& fen, const std::vector<std::string>& moves) {
@@ -124,11 +188,42 @@ void Engine::set_position(const std::string& fen, const std::vector<std::string>
 
 // modifiers
 
-void Engine::resize_threads() { threads.set({options, threads, tt, networks}, updateContext); }
+void Engine::set_numa_config_from_option(const std::string& o) {
+    if (o == "auto" || o == "system")
+    {
+        numaContext.set_numa_config(NumaConfig::from_system());
+    }
+    else if (o == "hardware")
+    {
+        // Don't respect affinity set in the system.
+        numaContext.set_numa_config(NumaConfig::from_system(false));
+    }
+    else if (o == "none")
+    {
+        numaContext.set_numa_config(NumaConfig{});
+    }
+    else
+    {
+        numaContext.set_numa_config(NumaConfig::from_string(o));
+    }
+
+    // Force reallocation of threads in case affinities need to change.
+    resize_threads();
+    threads.ensure_network_replicated();
+}
+
+void Engine::resize_threads() {
+    threads.wait_for_search_finished();
+    threads.set(numaContext.get_numa_config(), {options, threads, tt, networks}, updateContext);
+
+    // Reallocate the hash with the new threadpool size
+    set_tt_size(options["Hash"]);
+    threads.ensure_network_replicated();
+}
 
 void Engine::set_tt_size(size_t mb) {
     wait_for_search_finished();
-    tt.resize(mb, options["Threads"]);
+    tt.resize(mb, threads);
 }
 
 void Engine::set_ponderhit(bool b) { threads.main_manager()->ponder = b; }
@@ -136,28 +231,38 @@ void Engine::set_ponderhit(bool b) { threads.main_manager()->ponder = b; }
 // network related
 
 void Engine::verify_networks() const {
-    networks.big.verify(options["EvalFile"]);
-    networks.small.verify(options["EvalFileSmall"]);
+    networks->big.verify(options["EvalFile"], onVerifyNetworks);
+    networks->small.verify(options["EvalFileSmall"], onVerifyNetworks);
 }
 
 void Engine::load_networks() {
-    load_big_network(options["EvalFile"]);
-    load_small_network(options["EvalFileSmall"]);
+    networks.modify_and_replicate([this](NN::Networks& networks_) {
+        networks_.big.load(binaryDirectory, options["EvalFile"]);
+        networks_.small.load(binaryDirectory, options["EvalFileSmall"]);
+    });
+    threads.clear();
+    threads.ensure_network_replicated();
 }
 
 void Engine::load_big_network(const std::string& file) {
-    networks.big.load(binaryDirectory, file);
+    networks.modify_and_replicate(
+      [this, &file](NN::Networks& networks_) { networks_.big.load(binaryDirectory, file); });
     threads.clear();
+    threads.ensure_network_replicated();
 }
 
 void Engine::load_small_network(const std::string& file) {
-    networks.small.load(binaryDirectory, file);
+    networks.modify_and_replicate(
+      [this, &file](NN::Networks& networks_) { networks_.small.load(binaryDirectory, file); });
     threads.clear();
+    threads.ensure_network_replicated();
 }
 
 void Engine::save_network(const std::pair<std::optional<std::string>, std::string> files[2]) {
-    networks.big.save(files[0].first);
-    networks.small.save(files[1].first);
+    networks.modify_and_replicate([&files](NN::Networks& networks_) {
+        networks_.big.save(files[0].first);
+        networks_.small.save(files[1].first);
+    });
 }
 
 // utility functions
@@ -169,10 +274,11 @@ void Engine::trace_eval() const {
 
     verify_networks();
 
-    sync_cout << "\n" << Eval::trace(p, networks) << sync_endl;
+    sync_cout << "\n" << Eval::trace(p, *networks) << sync_endl;
 }
 
-OptionsMap& Engine::get_options() { return options; }
+const OptionsMap& Engine::get_options() const { return options; }
+OptionsMap&       Engine::get_options() { return options; }
 
 std::string Engine::fen() const { return pos.fen(); }
 
@@ -184,4 +290,63 @@ std::string Engine::visualize() const {
     return ss.str();
 }
 
+int Engine::get_hashfull(int maxAge) const { return tt.hashfull(maxAge); }
+
+std::vector<std::pair<size_t, size_t>> Engine::get_bound_thread_count_by_numa_node() const {
+    auto                                   counts = threads.get_bound_thread_count_by_numa_node();
+    const NumaConfig&                      cfg    = numaContext.get_numa_config();
+    std::vector<std::pair<size_t, size_t>> ratios;
+    NumaIndex                              n = 0;
+    for (; n < counts.size(); ++n)
+        ratios.emplace_back(counts[n], cfg.num_cpus_in_numa_node(n));
+    if (!counts.empty())
+        for (; n < cfg.num_numa_nodes(); ++n)
+            ratios.emplace_back(0, cfg.num_cpus_in_numa_node(n));
+    return ratios;
+}
+
+std::string Engine::get_numa_config_as_string() const {
+    return numaContext.get_numa_config().to_string();
+}
+
+std::string Engine::numa_config_information_as_string() const {
+    auto cfgStr = get_numa_config_as_string();
+    return "Available processors: " + cfgStr;
+}
+
+std::string Engine::thread_binding_information_as_string() const {
+    auto              boundThreadsByNode = get_bound_thread_count_by_numa_node();
+    std::stringstream ss;
+    if (boundThreadsByNode.empty())
+        return ss.str();
+
+    bool isFirst = true;
+
+    for (auto&& [current, total] : boundThreadsByNode)
+    {
+        if (!isFirst)
+            ss << ":";
+        ss << current << "/" << total;
+        isFirst = false;
+    }
+
+    return ss.str();
+}
+
+std::string Engine::thread_allocation_information_as_string() const {
+    std::stringstream ss;
+
+    size_t threadsSize = threads.size();
+    ss << "Using " << threadsSize << (threadsSize > 1 ? " threads" : " thread");
+
+    auto boundThreadsByNodeStr = thread_binding_information_as_string();
+    if (boundThreadsByNodeStr.empty())
+        return ss.str();
+
+    ss << " with NUMA node thread binding: ";
+    ss << boundThreadsByNodeStr;
+
+    return ss.str();
+}
+
 }
diff --git a/src/engine.h b/src/engine.h
index 64a814cb..25782693 100644
--- a/src/engine.h
+++ b/src/engine.h
@@ -29,6 +29,7 @@
 #include <vector>
 
 #include "nnue/network.h"
+#include "numa.h"
 #include "position.h"
 #include "search.h"
 #include "syzygy/tbprobe.h"  // for Stockfish::Depth
@@ -46,7 +47,14 @@ class Engine {
     using InfoFull  = Search::InfoFull;
     using InfoIter  = Search::InfoIteration;
 
-    Engine(std::string path = "");
+    Engine(std::optional<std::string> path = std::nullopt);
+
+    // Cannot be movable due to components holding backreferences to fields
+    Engine(const Engine&)            = delete;
+    Engine(Engine&&)                 = delete;
+    Engine& operator=(const Engine&) = delete;
+    Engine& operator=(Engine&&)      = delete;
+
     ~Engine() { wait_for_search_finished(); }
 
     std::uint64_t perft(const std::string& fen, Depth depth, bool isChess960);
@@ -63,6 +71,7 @@ class Engine {
 
     // modifiers
 
+    void set_numa_config_from_option(const std::string& o);
     void resize_threads();
     void set_tt_size(size_t mb);
     void set_ponderhit(bool);
@@ -72,6 +81,7 @@ class Engine {
     void set_on_update_full(std::function<void(const InfoFull&)>&&);
     void set_on_iter(std::function<void(const InfoIter&)>&&);
     void set_on_bestmove(std::function<void(std::string_view, std::string_view)>&&);
+    void set_on_verify_networks(std::function<void(std::string_view)>&&);
 
     // network related
 
@@ -83,25 +93,38 @@ class Engine {
 
     // utility functions
 
-    void        trace_eval() const;
-    OptionsMap& get_options();
-    std::string fen() const;
-    void        flip();
-    std::string visualize() const;
+    void trace_eval() const;
+
+    const OptionsMap& get_options() const;
+    OptionsMap&       get_options();
+
+    int get_hashfull(int maxAge = 0) const;
+
+    std::string                            fen() const;
+    void                                   flip();
+    std::string                            visualize() const;
+    std::vector<std::pair<size_t, size_t>> get_bound_thread_count_by_numa_node() const;
+    std::string                            get_numa_config_as_string() const;
+    std::string                            numa_config_information_as_string() const;
+    std::string                            thread_allocation_information_as_string() const;
+    std::string                            thread_binding_information_as_string() const;
 
    private:
     const std::string binaryDirectory;
 
+    NumaReplicationContext numaContext;
+
     Position     pos;
     StateListPtr states;
     Square       capSq;
 
-    OptionsMap           options;
-    ThreadPool           threads;
-    TranspositionTable   tt;
-    Eval::NNUE::Networks networks;
+    OptionsMap                               options;
+    ThreadPool                               threads;
+    TranspositionTable                       tt;
+    LazyNumaReplicated<Eval::NNUE::Networks> networks;
 
-    Search::SearchManager::UpdateContext updateContext;
+    Search::SearchManager::UpdateContext  updateContext;
+    std::function<void(std::string_view)> onVerifyNetworks;
 };
 
 }  // namespace Stockfish
diff --git a/src/evaluate.cpp b/src/evaluate.cpp
index dd0480fa..5e580da5 100644
--- a/src/evaluate.cpp
+++ b/src/evaluate.cpp
@@ -24,8 +24,9 @@
 #include <cstdlib>
 #include <iomanip>
 #include <iostream>
-#include <sstream>
 #include <memory>
+#include <sstream>
+#include <tuple>
 
 #include "nnue/network.h"
 #include "nnue/nnue_misc.h"
@@ -44,6 +45,10 @@ int Eval::simple_eval(const Position& pos, Color c) {
          + (pos.non_pawn_material(c) - pos.non_pawn_material(~c));
 }
 
+bool Eval::use_smallnet(const Position& pos) {
+    int simpleEval = simple_eval(pos, pos.side_to_move());
+    return std::abs(simpleEval) > 962;
+}
 
 // Evaluate is the evaluator for the outer world. It returns a static evaluation
 // of the position from the point of view of the side to move.
@@ -54,34 +59,30 @@ Value Eval::evaluate(const Eval::NNUE::Networks&    networks,
 
     assert(!pos.checkers());
 
-    int  simpleEval = simple_eval(pos, pos.side_to_move());
-    bool smallNet   = std::abs(simpleEval) > SmallNetThreshold;
-    int  nnueComplexity;
-    int  v;
+    bool smallNet           = use_smallnet(pos);
+    auto [psqt, positional] = smallNet ? networks.small.evaluate(pos, &caches.small)
+                                       : networks.big.evaluate(pos, &caches.big);
 
-    Value nnue = smallNet ? networks.small.evaluate(pos, &caches.small, true, &nnueComplexity)
-                          : networks.big.evaluate(pos, &caches.big, true, &nnueComplexity);
+    Value nnue = (125 * psqt + 131 * positional) / 128;
 
-    const auto adjustEval = [&](int nnueDiv, int pawnCountConstant, int pawnCountMul,
-                                int npmConstant, int evalDiv, int shufflingConstant) {
-        // Blend optimism and eval with nnue complexity and material imbalance
-        optimism += optimism * (nnueComplexity + std::abs(simpleEval - nnue)) / 584;
-        nnue -= nnue * (nnueComplexity * 5 / 3) / nnueDiv;
+    // Re-evaluate the position when higher eval accuracy is worth the time spent
+    if (smallNet && (std::abs(nnue) < 236))
+    {
+        std::tie(psqt, positional) = networks.big.evaluate(pos, &caches.big);
+        nnue                       = (125 * psqt + 131 * positional) / 128;
+        smallNet                   = false;
+    }
 
-        int npm = pos.non_pawn_material() / 64;
-        v       = (nnue * (npm + pawnCountConstant + pawnCountMul * pos.count<PAWN>())
-             + optimism * (npmConstant + npm))
-          / evalDiv;
+    // Blend optimism and eval with nnue complexity
+    int nnueComplexity = std::abs(psqt - positional);
+    optimism += optimism * nnueComplexity / 468;
+    nnue -= nnue * nnueComplexity / (smallNet ? 20233 : 17879);
 
-        // Damp down the evaluation linearly when shuffling
-        int shuffling = pos.rule50_count();
-        v             = v * (shufflingConstant - shuffling) / 207;
-    };
+    int material = (smallNet ? 553 : 532) * pos.count<PAWN>() + pos.non_pawn_material();
+    int v        = (nnue * (77777 + material) + optimism * (7777 + material)) / 77777;
 
-    if (!smallNet)
-        adjustEval(32395, 942, 11, 139, 1058, 178);
-    else
-        adjustEval(32793, 944, 9, 140, 1067, 206);
+    // Damp down the evaluation linearly when shuffling
+    v -= v * pos.rule50_count() / 212;
 
     // Guarantee evaluation does not hit the tablebase range
     v = std::clamp(v, VALUE_TB_LOSS_IN_MAX_PLY + 1, VALUE_TB_WIN_IN_MAX_PLY - 1);
@@ -108,8 +109,9 @@ std::string Eval::trace(Position& pos, const Eval::NNUE::Networks& networks) {
 
     ss << std::showpoint << std::showpos << std::fixed << std::setprecision(2) << std::setw(15);
 
-    Value v = networks.big.evaluate(pos, &caches->big, false);
-    v       = pos.side_to_move() == WHITE ? v : -v;
+    auto [psqt, positional] = networks.big.evaluate(pos, &caches->big);
+    Value v                 = psqt + positional;
+    v                       = pos.side_to_move() == WHITE ? v : -v;
     ss << "NNUE evaluation        " << 0.01 * UCIEngine::to_cp(v, pos) << " (white side)\n";
 
     v = evaluate(networks, pos, *caches, VALUE_ZERO);
diff --git a/src/evaluate.h b/src/evaluate.h
index 2d244ff6..4604321d 100644
--- a/src/evaluate.h
+++ b/src/evaluate.h
@@ -29,14 +29,12 @@ class Position;
 
 namespace Eval {
 
-constexpr inline int SmallNetThreshold = 1274;
-
 // The default net name MUST follow the format nn-[SHA256 first 12 digits].nnue
 // for the build process (profile-build and fishtest) to work. Do not change the
 // name of the macro or the location where this macro is defined, as it is used
 // in the Makefile/Fishtest.
-#define EvalFileDefaultNameBig "nn-ae6a388e4a1a.nnue"
-#define EvalFileDefaultNameSmall "nn-baff1ede1f90.nnue"
+#define EvalFileDefaultNameBig "nn-1c0000000000.nnue"
+#define EvalFileDefaultNameSmall "nn-37f18f62d772.nnue"
 
 namespace NNUE {
 struct Networks;
@@ -46,6 +44,7 @@ struct AccumulatorCaches;
 std::string trace(Position& pos, const Eval::NNUE::Networks& networks);
 
 int   simple_eval(const Position& pos, Color c);
+bool  use_smallnet(const Position& pos);
 Value evaluate(const NNUE::Networks&          networks,
                const Position&                pos,
                Eval::NNUE::AccumulatorCaches& caches,
diff --git a/src/history.h b/src/history.h
new file mode 100644
index 00000000..8d14a7a7
--- /dev/null
+++ b/src/history.h
@@ -0,0 +1,185 @@
+/*
+  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
+  Copyright (C) 2004-2024 The Stockfish developers (see AUTHORS file)
+
+  Stockfish is free software: you can redistribute it and/or modify
+  it under the terms of the GNU General Public License as published by
+  the Free Software Foundation, either version 3 of the License, or
+  (at your option) any later version.
+
+  Stockfish is distributed in the hope that it will be useful,
+  but WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+  GNU General Public License for more details.
+
+  You should have received a copy of the GNU General Public License
+  along with this program.  If not, see <http://www.gnu.org/licenses/>.
+*/
+
+#ifndef HISTORY_H_INCLUDED
+#define HISTORY_H_INCLUDED
+
+#include <algorithm>
+#include <array>
+#include <cassert>
+#include <cmath>
+#include <cstdint>
+#include <cstdlib>
+#include <limits>
+#include <type_traits>  // IWYU pragma: keep
+
+#include "position.h"
+
+namespace Stockfish {
+
+constexpr int PAWN_HISTORY_SIZE        = 512;    // has to be a power of 2
+constexpr int CORRECTION_HISTORY_SIZE  = 32768;  // has to be a power of 2
+constexpr int CORRECTION_HISTORY_LIMIT = 1024;
+constexpr int LOW_PLY_HISTORY_SIZE     = 4;
+
+static_assert((PAWN_HISTORY_SIZE & (PAWN_HISTORY_SIZE - 1)) == 0,
+              "PAWN_HISTORY_SIZE has to be a power of 2");
+
+static_assert((CORRECTION_HISTORY_SIZE & (CORRECTION_HISTORY_SIZE - 1)) == 0,
+              "CORRECTION_HISTORY_SIZE has to be a power of 2");
+
+enum PawnHistoryType {
+    Normal,
+    Correction
+};
+
+template<PawnHistoryType T = Normal>
+inline int pawn_structure_index(const Position& pos) {
+    return pos.pawn_key() & ((T == Normal ? PAWN_HISTORY_SIZE : CORRECTION_HISTORY_SIZE) - 1);
+}
+
+inline int major_piece_index(const Position& pos) {
+    return pos.major_piece_key() & (CORRECTION_HISTORY_SIZE - 1);
+}
+
+inline int minor_piece_index(const Position& pos) {
+    return pos.minor_piece_key() & (CORRECTION_HISTORY_SIZE - 1);
+}
+
+template<Color c>
+inline int non_pawn_index(const Position& pos) {
+    return pos.non_pawn_key(c) & (CORRECTION_HISTORY_SIZE - 1);
+}
+
+// StatsEntry stores the stat table value. It is usually a number but could
+// be a move or even a nested history. We use a class instead of a naked value
+// to directly call history update operator<<() on the entry so to use stats
+// tables at caller sites as simple multi-dim arrays.
+template<typename T, int D>
+class StatsEntry {
+
+    T entry;
+
+   public:
+    void operator=(const T& v) { entry = v; }
+    T*   operator&() { return &entry; }
+    T*   operator->() { return &entry; }
+    operator const T&() const { return entry; }
+
+    void operator<<(int bonus) {
+        static_assert(D <= std::numeric_limits<T>::max(), "D overflows T");
+
+        // Make sure that bonus is in range [-D, D]
+        int clampedBonus = std::clamp(bonus, -D, D);
+        entry += clampedBonus - entry * std::abs(clampedBonus) / D;
+
+        assert(std::abs(entry) <= D);
+    }
+};
+
+// Stats is a generic N-dimensional array used to store various statistics.
+// The first template parameter T is the base type of the array, and the second
+// template parameter D limits the range of updates in [-D, D] when we update
+// values with the << operator, while the last parameters (Size and Sizes)
+// encode the dimensions of the array.
+template<typename T, int D, int Size, int... Sizes>
+struct Stats: public std::array<Stats<T, D, Sizes...>, Size> {
+    using stats = Stats<T, D, Size, Sizes...>;
+
+    void fill(const T& v) {
+
+        // For standard-layout 'this' points to the first struct member
+        assert(std::is_standard_layout_v<stats>);
+
+        using entry = StatsEntry<T, D>;
+        entry* p    = reinterpret_cast<entry*>(this);
+        std::fill(p, p + sizeof(*this) / sizeof(entry), v);
+    }
+};
+
+template<typename T, int D, int Size>
+struct Stats<T, D, Size>: public std::array<StatsEntry<T, D>, Size> {};
+
+// In stats table, D=0 means that the template parameter is not used
+enum StatsParams {
+    NOT_USED = 0
+};
+enum StatsType {
+    NoCaptures,
+    Captures
+};
+
+// ButterflyHistory records how often quiet moves have been successful or unsuccessful
+// during the current search, and is used for reduction and move ordering decisions.
+// It uses 2 tables (one for each color) indexed by the move's from and to squares,
+// see https://www.chessprogramming.org/Butterfly_Boards (~11 elo)
+using ButterflyHistory = Stats<int16_t, 7183, COLOR_NB, int(SQUARE_NB) * int(SQUARE_NB)>;
+
+// LowPlyHistory is adressed by play and move's from and to squares, used
+// to improve move ordering near the root
+using LowPlyHistory = Stats<int16_t, 7183, LOW_PLY_HISTORY_SIZE, int(SQUARE_NB) * int(SQUARE_NB)>;
+
+// CapturePieceToHistory is addressed by a move's [piece][to][captured piece type]
+using CapturePieceToHistory = Stats<int16_t, 10692, PIECE_NB, SQUARE_NB, PIECE_TYPE_NB>;
+
+// PieceToHistory is like ButterflyHistory but is addressed by a move's [piece][to]
+using PieceToHistory = Stats<int16_t, 29952, PIECE_NB, SQUARE_NB>;
+
+// ContinuationHistory is the combined history of a given pair of moves, usually
+// the current one given a previous one. The nested history table is based on
+// PieceToHistory instead of ButterflyBoards.
+// (~63 elo)
+using ContinuationHistory = Stats<PieceToHistory, NOT_USED, PIECE_NB, SQUARE_NB>;
+
+// PawnHistory is addressed by the pawn structure and a move's [piece][to]
+using PawnHistory = Stats<int16_t, 8192, PAWN_HISTORY_SIZE, PIECE_NB, SQUARE_NB>;
+
+// Correction histories record differences between the static evaluation of
+// positions and their search score. It is used to improve the static evaluation
+// used by some search heuristics.
+// see https://www.chessprogramming.org/Static_Evaluation_Correction_History
+enum CorrHistType {
+    Pawn,          // By color and pawn structure
+    Major,         // By color and positions of major pieces (Queen, Rook) and King
+    Minor,         // By color and positions of minor pieces (Knight, Bishop) and King
+    NonPawn,       // By color and non-pawn material positions
+    PieceTo,       // By [piece][to] move
+    Continuation,  // Combined history of move pairs
+};
+
+template<CorrHistType _>
+struct CorrHistTypedef {
+    using type = Stats<int16_t, CORRECTION_HISTORY_LIMIT, COLOR_NB, CORRECTION_HISTORY_SIZE>;
+};
+
+template<>
+struct CorrHistTypedef<PieceTo> {
+    using type = Stats<int16_t, CORRECTION_HISTORY_LIMIT, PIECE_NB, SQUARE_NB>;
+};
+
+template<>
+struct CorrHistTypedef<Continuation> {
+    using type = Stats<CorrHistTypedef<PieceTo>::type, NOT_USED, PIECE_NB, SQUARE_NB>;
+};
+
+template<CorrHistType T>
+using CorrectionHistory = typename CorrHistTypedef<T>::type;
+
+}  // namespace Stockfish
+
+#endif  // #ifndef HISTORY_H_INCLUDED
diff --git a/src/memory.cpp b/src/memory.cpp
new file mode 100644
index 00000000..47c901b4
--- /dev/null
+++ b/src/memory.cpp
@@ -0,0 +1,268 @@
+/*
+  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
+  Copyright (C) 2004-2024 The Stockfish developers (see AUTHORS file)
+
+  Stockfish is free software: you can redistribute it and/or modify
+  it under the terms of the GNU General Public License as published by
+  the Free Software Foundation, either version 3 of the License, or
+  (at your option) any later version.
+
+  Stockfish is distributed in the hope that it will be useful,
+  but WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+  GNU General Public License for more details.
+
+  You should have received a copy of the GNU General Public License
+  along with this program.  If not, see <http://www.gnu.org/licenses/>.
+*/
+
+#include "memory.h"
+
+#include <cstdlib>
+
+#if __has_include("features.h")
+    #include <features.h>
+#endif
+
+#if defined(__linux__) && !defined(__ANDROID__)
+    #include <sys/mman.h>
+#endif
+
+#if defined(__APPLE__) || defined(__ANDROID__) || defined(__OpenBSD__) \
+  || (defined(__GLIBCXX__) && !defined(_GLIBCXX_HAVE_ALIGNED_ALLOC) && !defined(_WIN32)) \
+  || defined(__e2k__)
+    #define POSIXALIGNEDALLOC
+    #include <stdlib.h>
+#endif
+
+#ifdef _WIN32
+    #if _WIN32_WINNT < 0x0601
+        #undef _WIN32_WINNT
+        #define _WIN32_WINNT 0x0601  // Force to include needed API prototypes
+    #endif
+
+    #ifndef NOMINMAX
+        #define NOMINMAX
+    #endif
+
+    #include <ios>       // std::hex, std::dec
+    #include <iostream>  // std::cerr
+    #include <ostream>   // std::endl
+    #include <windows.h>
+
+// The needed Windows API for processor groups could be missed from old Windows
+// versions, so instead of calling them directly (forcing the linker to resolve
+// the calls at compile time), try to load them at runtime. To do this we need
+// first to define the corresponding function pointers.
+
+extern "C" {
+using OpenProcessToken_t      = bool (*)(HANDLE, DWORD, PHANDLE);
+using LookupPrivilegeValueA_t = bool (*)(LPCSTR, LPCSTR, PLUID);
+using AdjustTokenPrivileges_t =
+  bool (*)(HANDLE, BOOL, PTOKEN_PRIVILEGES, DWORD, PTOKEN_PRIVILEGES, PDWORD);
+}
+#endif
+
+
+namespace Stockfish {
+
+// Wrappers for systems where the c++17 implementation does not guarantee the
+// availability of aligned_alloc(). Memory allocated with std_aligned_alloc()
+// must be freed with std_aligned_free().
+
+void* std_aligned_alloc(size_t alignment, size_t size) {
+#if defined(_ISOC11_SOURCE)
+    return aligned_alloc(alignment, size);
+#elif defined(POSIXALIGNEDALLOC)
+    void* mem = nullptr;
+    posix_memalign(&mem, alignment, size);
+    return mem;
+#elif defined(_WIN32) && !defined(_M_ARM) && !defined(_M_ARM64)
+    return _mm_malloc(size, alignment);
+#elif defined(_WIN32)
+    return _aligned_malloc(size, alignment);
+#else
+    return std::aligned_alloc(alignment, size);
+#endif
+}
+
+void std_aligned_free(void* ptr) {
+
+#if defined(POSIXALIGNEDALLOC)
+    free(ptr);
+#elif defined(_WIN32) && !defined(_M_ARM) && !defined(_M_ARM64)
+    _mm_free(ptr);
+#elif defined(_WIN32)
+    _aligned_free(ptr);
+#else
+    free(ptr);
+#endif
+}
+
+// aligned_large_pages_alloc() will return suitably aligned memory,
+// if possible using large pages.
+
+#if defined(_WIN32)
+
+static void* aligned_large_pages_alloc_windows([[maybe_unused]] size_t allocSize) {
+
+    #if !defined(_WIN64)
+    return nullptr;
+    #else
+
+    HANDLE hProcessToken{};
+    LUID   luid{};
+    void*  mem = nullptr;
+
+    const size_t largePageSize = GetLargePageMinimum();
+    if (!largePageSize)
+        return nullptr;
+
+    // Dynamically link OpenProcessToken, LookupPrivilegeValue and AdjustTokenPrivileges
+
+    HMODULE hAdvapi32 = GetModuleHandle(TEXT("advapi32.dll"));
+
+    if (!hAdvapi32)
+        hAdvapi32 = LoadLibrary(TEXT("advapi32.dll"));
+
+    auto OpenProcessToken_f =
+      OpenProcessToken_t((void (*)()) GetProcAddress(hAdvapi32, "OpenProcessToken"));
+    if (!OpenProcessToken_f)
+        return nullptr;
+    auto LookupPrivilegeValueA_f =
+      LookupPrivilegeValueA_t((void (*)()) GetProcAddress(hAdvapi32, "LookupPrivilegeValueA"));
+    if (!LookupPrivilegeValueA_f)
+        return nullptr;
+    auto AdjustTokenPrivileges_f =
+      AdjustTokenPrivileges_t((void (*)()) GetProcAddress(hAdvapi32, "AdjustTokenPrivileges"));
+    if (!AdjustTokenPrivileges_f)
+        return nullptr;
+
+    // We need SeLockMemoryPrivilege, so try to enable it for the process
+
+    if (!OpenProcessToken_f(  // OpenProcessToken()
+          GetCurrentProcess(), TOKEN_ADJUST_PRIVILEGES | TOKEN_QUERY, &hProcessToken))
+        return nullptr;
+
+    if (LookupPrivilegeValueA_f(nullptr, "SeLockMemoryPrivilege", &luid))
+    {
+        TOKEN_PRIVILEGES tp{};
+        TOKEN_PRIVILEGES prevTp{};
+        DWORD            prevTpLen = 0;
+
+        tp.PrivilegeCount           = 1;
+        tp.Privileges[0].Luid       = luid;
+        tp.Privileges[0].Attributes = SE_PRIVILEGE_ENABLED;
+
+        // Try to enable SeLockMemoryPrivilege. Note that even if AdjustTokenPrivileges()
+        // succeeds, we still need to query GetLastError() to ensure that the privileges
+        // were actually obtained.
+
+        if (AdjustTokenPrivileges_f(hProcessToken, FALSE, &tp, sizeof(TOKEN_PRIVILEGES), &prevTp,
+                                    &prevTpLen)
+            && GetLastError() == ERROR_SUCCESS)
+        {
+            // Round up size to full pages and allocate
+            allocSize = (allocSize + largePageSize - 1) & ~size_t(largePageSize - 1);
+            mem       = VirtualAlloc(nullptr, allocSize, MEM_RESERVE | MEM_COMMIT | MEM_LARGE_PAGES,
+                                     PAGE_READWRITE);
+
+            // Privilege no longer needed, restore previous state
+            AdjustTokenPrivileges_f(hProcessToken, FALSE, &prevTp, 0, nullptr, nullptr);
+        }
+    }
+
+    CloseHandle(hProcessToken);
+
+    return mem;
+
+    #endif
+}
+
+void* aligned_large_pages_alloc(size_t allocSize) {
+
+    // Try to allocate large pages
+    void* mem = aligned_large_pages_alloc_windows(allocSize);
+
+    // Fall back to regular, page-aligned, allocation if necessary
+    if (!mem)
+        mem = VirtualAlloc(nullptr, allocSize, MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE);
+
+    return mem;
+}
+
+#else
+
+void* aligned_large_pages_alloc(size_t allocSize) {
+
+    #if defined(__linux__)
+    constexpr size_t alignment = 2 * 1024 * 1024;  // 2MB page size assumed
+    #else
+    constexpr size_t alignment = 4096;  // small page size assumed
+    #endif
+
+    // Round up to multiples of alignment
+    size_t size = ((allocSize + alignment - 1) / alignment) * alignment;
+    void*  mem  = std_aligned_alloc(alignment, size);
+    #if defined(MADV_HUGEPAGE)
+    madvise(mem, size, MADV_HUGEPAGE);
+    #endif
+    return mem;
+}
+
+#endif
+
+bool has_large_pages() {
+
+#if defined(_WIN32)
+
+    constexpr size_t page_size = 2 * 1024 * 1024;  // 2MB page size assumed
+    void*            mem       = aligned_large_pages_alloc_windows(page_size);
+    if (mem == nullptr)
+    {
+        return false;
+    }
+    else
+    {
+        aligned_large_pages_free(mem);
+        return true;
+    }
+
+#elif defined(__linux__)
+
+    #if defined(MADV_HUGEPAGE)
+    return true;
+    #else
+    return false;
+    #endif
+
+#else
+
+    return false;
+
+#endif
+}
+
+
+// aligned_large_pages_free() will free the previously memory allocated
+// by aligned_large_pages_alloc(). The effect is a nop if mem == nullptr.
+
+#if defined(_WIN32)
+
+void aligned_large_pages_free(void* mem) {
+
+    if (mem && !VirtualFree(mem, 0, MEM_RELEASE))
+    {
+        DWORD err = GetLastError();
+        std::cerr << "Failed to free large page memory. Error code: 0x" << std::hex << err
+                  << std::dec << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+#else
+
+void aligned_large_pages_free(void* mem) { std_aligned_free(mem); }
+
+#endif
+}  // namespace Stockfish
diff --git a/src/memory.h b/src/memory.h
new file mode 100644
index 00000000..eaf0261a
--- /dev/null
+++ b/src/memory.h
@@ -0,0 +1,218 @@
+/*
+  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
+  Copyright (C) 2004-2024 The Stockfish developers (see AUTHORS file)
+
+  Stockfish is free software: you can redistribute it and/or modify
+  it under the terms of the GNU General Public License as published by
+  the Free Software Foundation, either version 3 of the License, or
+  (at your option) any later version.
+
+  Stockfish is distributed in the hope that it will be useful,
+  but WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+  GNU General Public License for more details.
+
+  You should have received a copy of the GNU General Public License
+  along with this program.  If not, see <http://www.gnu.org/licenses/>.
+*/
+
+#ifndef MEMORY_H_INCLUDED
+#define MEMORY_H_INCLUDED
+
+#include <algorithm>
+#include <cstddef>
+#include <cstdint>
+#include <memory>
+#include <new>
+#include <type_traits>
+#include <utility>
+
+#include "types.h"
+
+namespace Stockfish {
+
+void* std_aligned_alloc(size_t alignment, size_t size);
+void  std_aligned_free(void* ptr);
+
+// Memory aligned by page size, min alignment: 4096 bytes
+void* aligned_large_pages_alloc(size_t size);
+void  aligned_large_pages_free(void* mem);
+
+bool has_large_pages();
+
+// Frees memory which was placed there with placement new.
+// Works for both single objects and arrays of unknown bound.
+template<typename T, typename FREE_FUNC>
+void memory_deleter(T* ptr, FREE_FUNC free_func) {
+    if (!ptr)
+        return;
+
+    // Explicitly needed to call the destructor
+    if constexpr (!std::is_trivially_destructible_v<T>)
+        ptr->~T();
+
+    free_func(ptr);
+    return;
+}
+
+// Frees memory which was placed there with placement new.
+// Works for both single objects and arrays of unknown bound.
+template<typename T, typename FREE_FUNC>
+void memory_deleter_array(T* ptr, FREE_FUNC free_func) {
+    if (!ptr)
+        return;
+
+
+    // Move back on the pointer to where the size is allocated
+    const size_t array_offset = std::max(sizeof(size_t), alignof(T));
+    char*        raw_memory   = reinterpret_cast<char*>(ptr) - array_offset;
+
+    if constexpr (!std::is_trivially_destructible_v<T>)
+    {
+        const size_t size = *reinterpret_cast<size_t*>(raw_memory);
+
+        // Explicitly call the destructor for each element in reverse order
+        for (size_t i = size; i-- > 0;)
+            ptr[i].~T();
+    }
+
+    free_func(raw_memory);
+}
+
+// Allocates memory for a single object and places it there with placement new
+template<typename T, typename ALLOC_FUNC, typename... Args>
+inline std::enable_if_t<!std::is_array_v<T>, T*> memory_allocator(ALLOC_FUNC alloc_func,
+                                                                  Args&&... args) {
+    void* raw_memory = alloc_func(sizeof(T));
+    ASSERT_ALIGNED(raw_memory, alignof(T));
+    return new (raw_memory) T(std::forward<Args>(args)...);
+}
+
+// Allocates memory for an array of unknown bound and places it there with placement new
+template<typename T, typename ALLOC_FUNC>
+inline std::enable_if_t<std::is_array_v<T>, std::remove_extent_t<T>*>
+memory_allocator(ALLOC_FUNC alloc_func, size_t num) {
+    using ElementType = std::remove_extent_t<T>;
+
+    const size_t array_offset = std::max(sizeof(size_t), alignof(ElementType));
+
+    // Save the array size in the memory location
+    char* raw_memory =
+      reinterpret_cast<char*>(alloc_func(array_offset + num * sizeof(ElementType)));
+    ASSERT_ALIGNED(raw_memory, alignof(T));
+
+    new (raw_memory) size_t(num);
+
+    for (size_t i = 0; i < num; ++i)
+        new (raw_memory + array_offset + i * sizeof(ElementType)) ElementType();
+
+    // Need to return the pointer at the start of the array so that
+    // the indexing in unique_ptr<T[]> works.
+    return reinterpret_cast<ElementType*>(raw_memory + array_offset);
+}
+
+//
+//
+// aligned large page unique ptr
+//
+//
+
+template<typename T>
+struct LargePageDeleter {
+    void operator()(T* ptr) const { return memory_deleter<T>(ptr, aligned_large_pages_free); }
+};
+
+template<typename T>
+struct LargePageArrayDeleter {
+    void operator()(T* ptr) const { return memory_deleter_array<T>(ptr, aligned_large_pages_free); }
+};
+
+template<typename T>
+using LargePagePtr =
+  std::conditional_t<std::is_array_v<T>,
+                     std::unique_ptr<T, LargePageArrayDeleter<std::remove_extent_t<T>>>,
+                     std::unique_ptr<T, LargePageDeleter<T>>>;
+
+// make_unique_large_page for single objects
+template<typename T, typename... Args>
+std::enable_if_t<!std::is_array_v<T>, LargePagePtr<T>> make_unique_large_page(Args&&... args) {
+    static_assert(alignof(T) <= 4096,
+                  "aligned_large_pages_alloc() may fail for such a big alignment requirement of T");
+
+    T* obj = memory_allocator<T>(aligned_large_pages_alloc, std::forward<Args>(args)...);
+
+    return LargePagePtr<T>(obj);
+}
+
+// make_unique_large_page for arrays of unknown bound
+template<typename T>
+std::enable_if_t<std::is_array_v<T>, LargePagePtr<T>> make_unique_large_page(size_t num) {
+    using ElementType = std::remove_extent_t<T>;
+
+    static_assert(alignof(ElementType) <= 4096,
+                  "aligned_large_pages_alloc() may fail for such a big alignment requirement of T");
+
+    ElementType* memory = memory_allocator<T>(aligned_large_pages_alloc, num);
+
+    return LargePagePtr<T>(memory);
+}
+
+//
+//
+// aligned unique ptr
+//
+//
+
+template<typename T>
+struct AlignedDeleter {
+    void operator()(T* ptr) const { return memory_deleter<T>(ptr, std_aligned_free); }
+};
+
+template<typename T>
+struct AlignedArrayDeleter {
+    void operator()(T* ptr) const { return memory_deleter_array<T>(ptr, std_aligned_free); }
+};
+
+template<typename T>
+using AlignedPtr =
+  std::conditional_t<std::is_array_v<T>,
+                     std::unique_ptr<T, AlignedArrayDeleter<std::remove_extent_t<T>>>,
+                     std::unique_ptr<T, AlignedDeleter<T>>>;
+
+// make_unique_aligned for single objects
+template<typename T, typename... Args>
+std::enable_if_t<!std::is_array_v<T>, AlignedPtr<T>> make_unique_aligned(Args&&... args) {
+    const auto func = [](size_t size) { return std_aligned_alloc(alignof(T), size); };
+    T*         obj  = memory_allocator<T>(func, std::forward<Args>(args)...);
+
+    return AlignedPtr<T>(obj);
+}
+
+// make_unique_aligned for arrays of unknown bound
+template<typename T>
+std::enable_if_t<std::is_array_v<T>, AlignedPtr<T>> make_unique_aligned(size_t num) {
+    using ElementType = std::remove_extent_t<T>;
+
+    const auto   func   = [](size_t size) { return std_aligned_alloc(alignof(ElementType), size); };
+    ElementType* memory = memory_allocator<T>(func, num);
+
+    return AlignedPtr<T>(memory);
+}
+
+
+// Get the first aligned element of an array.
+// ptr must point to an array of size at least `sizeof(T) * N + alignment` bytes,
+// where N is the number of elements in the array.
+template<uintptr_t Alignment, typename T>
+T* align_ptr_up(T* ptr) {
+    static_assert(alignof(T) < Alignment);
+
+    const uintptr_t ptrint = reinterpret_cast<uintptr_t>(reinterpret_cast<char*>(ptr));
+    return reinterpret_cast<T*>(
+      reinterpret_cast<char*>((ptrint + (Alignment - 1)) / Alignment * Alignment));
+}
+
+
+}  // namespace Stockfish
+
+#endif  // #ifndef MEMORY_H_INCLUDED
diff --git a/src/misc.cpp b/src/misc.cpp
index 1abb81b1..10c86b7a 100644
--- a/src/misc.cpp
+++ b/src/misc.cpp
@@ -18,58 +18,21 @@
 
 #include "misc.h"
 
-#ifdef _WIN32
-    #if _WIN32_WINNT < 0x0601
-        #undef _WIN32_WINNT
-        #define _WIN32_WINNT 0x0601  // Force to include needed API prototypes
-    #endif
-
-    #ifndef NOMINMAX
-        #define NOMINMAX
-    #endif
-
-    #include <windows.h>
-// The needed Windows API for processor groups could be missed from old Windows
-// versions, so instead of calling them directly (forcing the linker to resolve
-// the calls at compile time), try to load them at runtime. To do this we need
-// first to define the corresponding function pointers.
-extern "C" {
-using fun1_t = bool (*)(LOGICAL_PROCESSOR_RELATIONSHIP,
-                        PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX,
-                        PDWORD);
-using fun2_t = bool (*)(USHORT, PGROUP_AFFINITY);
-using fun3_t = bool (*)(HANDLE, CONST GROUP_AFFINITY*, PGROUP_AFFINITY);
-using fun4_t = bool (*)(USHORT, PGROUP_AFFINITY, USHORT, PUSHORT);
-using fun5_t = WORD (*)();
-using fun6_t = bool (*)(HANDLE, DWORD, PHANDLE);
-using fun7_t = bool (*)(LPCSTR, LPCSTR, PLUID);
-using fun8_t = bool (*)(HANDLE, BOOL, PTOKEN_PRIVILEGES, DWORD, PTOKEN_PRIVILEGES, PDWORD);
-}
-#endif
-
 #include <atomic>
+#include <cctype>
 #include <cmath>
 #include <cstdlib>
 #include <fstream>
 #include <iomanip>
 #include <iostream>
+#include <iterator>
+#include <limits>
 #include <mutex>
 #include <sstream>
 #include <string_view>
 
 #include "types.h"
 
-#if defined(__linux__) && !defined(__ANDROID__)
-    #include <sys/mman.h>
-#endif
-
-#if defined(__APPLE__) || defined(__ANDROID__) || defined(__OpenBSD__) \
-  || (defined(__GLIBCXX__) && !defined(_GLIBCXX_HAVE_ALIGNED_ALLOC) && !defined(_WIN32)) \
-  || defined(__e2k__)
-    #define POSIXALIGNEDALLOC
-    #include <stdlib.h>
-#endif
-
 namespace Stockfish {
 
 namespace {
@@ -149,15 +112,17 @@ class Logger {
 
 
 // Returns the full name of the current Stockfish version.
-// For local dev compiles we try to append the commit sha and commit date
-// from git if that fails only the local compilation date is set and "nogit" is specified:
-// Stockfish dev-YYYYMMDD-SHA
-// or
-// Stockfish dev-YYYYMMDD-nogit
+//
+// For local dev compiles we try to append the commit SHA and
+// commit date from git. If that fails only the local compilation
+// date is set and "nogit" is specified:
+//      Stockfish dev-YYYYMMDD-SHA
+//      or
+//      Stockfish dev-YYYYMMDD-nogit
 //
 // For releases (non-dev builds) we only include the version number:
-// Stockfish version
-std::string engine_info(bool to_uci) {
+//      Stockfish version
+std::string engine_version_info() {
     std::stringstream ss;
     ss << "Stockfish " << version << std::setfill('0');
 
@@ -168,8 +133,9 @@ std::string engine_info(bool to_uci) {
         ss << stringify(GIT_DATE);
 #else
         constexpr std::string_view months("Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec");
-        std::string                month, day, year;
-        std::stringstream          date(__DATE__);  // From compiler, format is "Sep 21 2008"
+
+        std::string       month, day, year;
+        std::stringstream date(__DATE__);  // From compiler, format is "Sep 21 2008"
 
         date >> month >> day >> year;
         ss << year << std::setw(2) << std::setfill('0') << (1 + months.find(month) / 4)
@@ -185,11 +151,14 @@ std::string engine_info(bool to_uci) {
 #endif
     }
 
-    ss << (to_uci ? "\nid author " : " by ") << "the Stockfish developers (see AUTHORS file)";
-
     return ss.str();
 }
 
+std::string engine_info(bool to_uci) {
+    return engine_version_info() + (to_uci ? "\nid author " : " by ")
+         + "the Stockfish developers (see AUTHORS file)";
+}
+
 
 // Returns a string trying to describe the compiler we use
 std::string compiler_info() {
@@ -318,13 +287,21 @@ template<size_t N>
 struct DebugInfo {
     std::atomic<int64_t> data[N] = {0};
 
-    constexpr inline std::atomic<int64_t>& operator[](int index) { return data[index]; }
+    constexpr std::atomic<int64_t>& operator[](int index) { return data[index]; }
 };
 
-DebugInfo<2> hit[MaxDebugSlots];
-DebugInfo<2> mean[MaxDebugSlots];
-DebugInfo<3> stdev[MaxDebugSlots];
-DebugInfo<6> correl[MaxDebugSlots];
+struct DebugExtremes: public DebugInfo<3> {
+    DebugExtremes() {
+        data[1] = std::numeric_limits<int64_t>::min();
+        data[2] = std::numeric_limits<int64_t>::max();
+    }
+};
+
+DebugInfo<2>  hit[MaxDebugSlots];
+DebugInfo<2>  mean[MaxDebugSlots];
+DebugInfo<3>  stdev[MaxDebugSlots];
+DebugInfo<6>  correl[MaxDebugSlots];
+DebugExtremes extremes[MaxDebugSlots];
 
 }  // namespace
 
@@ -348,6 +325,18 @@ void dbg_stdev_of(int64_t value, int slot) {
     stdev[slot][2] += value * value;
 }
 
+void dbg_extremes_of(int64_t value, int slot) {
+    ++extremes[slot][0];
+
+    int64_t current_max = extremes[slot][1].load();
+    while (current_max < value && !extremes[slot][1].compare_exchange_weak(current_max, value))
+    {}
+
+    int64_t current_min = extremes[slot][2].load();
+    while (current_min > value && !extremes[slot][2].compare_exchange_weak(current_min, value))
+    {}
+}
+
 void dbg_correl_of(int64_t value1, int64_t value2, int slot) {
 
     ++correl[slot][0];
@@ -382,6 +371,13 @@ void dbg_print() {
             std::cerr << "Stdev #" << i << ": Total " << n << " Stdev " << r << std::endl;
         }
 
+    for (int i = 0; i < MaxDebugSlots; ++i)
+        if ((n = extremes[i][0]))
+        {
+            std::cerr << "Extremity #" << i << ": Total " << n << " Min " << extremes[i][2]
+                      << " Max " << extremes[i][1] << std::endl;
+        }
+
     for (int i = 0; i < MaxDebugSlots; ++i)
         if ((n = correl[i][0]))
         {
@@ -408,6 +404,8 @@ std::ostream& operator<<(std::ostream& os, SyncCout sc) {
     return os;
 }
 
+void sync_cout_start() { std::cout << IO_LOCK; }
+void sync_cout_end() { std::cout << IO_UNLOCK; }
 
 // Trampoline helper to avoid moving Logger to misc.h
 void start_logger(const std::string& fname) { Logger::start(fname); }
@@ -415,14 +413,14 @@ void start_logger(const std::string& fname) { Logger::start(fname); }
 
 #ifdef NO_PREFETCH
 
-void prefetch(void*) {}
+void prefetch(const void*) {}
 
 #else
 
-void prefetch(void* addr) {
+void prefetch(const void* addr) {
 
     #if defined(_MSC_VER)
-    _mm_prefetch((char*) addr, _MM_HINT_T0);
+    _mm_prefetch((char const*) addr, _MM_HINT_T0);
     #else
     __builtin_prefetch(addr);
     #endif
@@ -430,291 +428,6 @@ void prefetch(void* addr) {
 
 #endif
 
-
-// Wrapper for systems where the c++17 implementation
-// does not guarantee the availability of aligned_alloc(). Memory allocated with
-// std_aligned_alloc() must be freed with std_aligned_free().
-void* std_aligned_alloc(size_t alignment, size_t size) {
-
-#if defined(POSIXALIGNEDALLOC)
-    void* mem;
-    return posix_memalign(&mem, alignment, size) ? nullptr : mem;
-#elif defined(_WIN32) && !defined(_M_ARM) && !defined(_M_ARM64)
-    return _mm_malloc(size, alignment);
-#elif defined(_WIN32)
-    return _aligned_malloc(size, alignment);
-#else
-    return std::aligned_alloc(alignment, size);
-#endif
-}
-
-void std_aligned_free(void* ptr) {
-
-#if defined(POSIXALIGNEDALLOC)
-    free(ptr);
-#elif defined(_WIN32) && !defined(_M_ARM) && !defined(_M_ARM64)
-    _mm_free(ptr);
-#elif defined(_WIN32)
-    _aligned_free(ptr);
-#else
-    free(ptr);
-#endif
-}
-
-// aligned_large_pages_alloc() will return suitably aligned memory, if possible using large pages.
-
-#if defined(_WIN32)
-
-static void* aligned_large_pages_alloc_windows([[maybe_unused]] size_t allocSize) {
-
-    #if !defined(_WIN64)
-    return nullptr;
-    #else
-
-    HANDLE hProcessToken{};
-    LUID   luid{};
-    void*  mem = nullptr;
-
-    const size_t largePageSize = GetLargePageMinimum();
-    if (!largePageSize)
-        return nullptr;
-
-    // Dynamically link OpenProcessToken, LookupPrivilegeValue and AdjustTokenPrivileges
-
-    HMODULE hAdvapi32 = GetModuleHandle(TEXT("advapi32.dll"));
-
-    if (!hAdvapi32)
-        hAdvapi32 = LoadLibrary(TEXT("advapi32.dll"));
-
-    auto fun6 = fun6_t((void (*)()) GetProcAddress(hAdvapi32, "OpenProcessToken"));
-    if (!fun6)
-        return nullptr;
-    auto fun7 = fun7_t((void (*)()) GetProcAddress(hAdvapi32, "LookupPrivilegeValueA"));
-    if (!fun7)
-        return nullptr;
-    auto fun8 = fun8_t((void (*)()) GetProcAddress(hAdvapi32, "AdjustTokenPrivileges"));
-    if (!fun8)
-        return nullptr;
-
-    // We need SeLockMemoryPrivilege, so try to enable it for the process
-    if (!fun6(  // OpenProcessToken()
-          GetCurrentProcess(), TOKEN_ADJUST_PRIVILEGES | TOKEN_QUERY, &hProcessToken))
-        return nullptr;
-
-    if (fun7(  // LookupPrivilegeValue(nullptr, SE_LOCK_MEMORY_NAME, &luid)
-          nullptr, "SeLockMemoryPrivilege", &luid))
-    {
-        TOKEN_PRIVILEGES tp{};
-        TOKEN_PRIVILEGES prevTp{};
-        DWORD            prevTpLen = 0;
-
-        tp.PrivilegeCount           = 1;
-        tp.Privileges[0].Luid       = luid;
-        tp.Privileges[0].Attributes = SE_PRIVILEGE_ENABLED;
-
-        // Try to enable SeLockMemoryPrivilege. Note that even if AdjustTokenPrivileges() succeeds,
-        // we still need to query GetLastError() to ensure that the privileges were actually obtained.
-        if (fun8(  // AdjustTokenPrivileges()
-              hProcessToken, FALSE, &tp, sizeof(TOKEN_PRIVILEGES), &prevTp, &prevTpLen)
-            && GetLastError() == ERROR_SUCCESS)
-        {
-            // Round up size to full pages and allocate
-            allocSize = (allocSize + largePageSize - 1) & ~size_t(largePageSize - 1);
-            mem       = VirtualAlloc(nullptr, allocSize, MEM_RESERVE | MEM_COMMIT | MEM_LARGE_PAGES,
-                                     PAGE_READWRITE);
-
-            // Privilege no longer needed, restore previous state
-            fun8(  // AdjustTokenPrivileges ()
-              hProcessToken, FALSE, &prevTp, 0, nullptr, nullptr);
-        }
-    }
-
-    CloseHandle(hProcessToken);
-
-    return mem;
-
-    #endif
-}
-
-void* aligned_large_pages_alloc(size_t allocSize) {
-
-    // Try to allocate large pages
-    void* mem = aligned_large_pages_alloc_windows(allocSize);
-
-    // Fall back to regular, page-aligned, allocation if necessary
-    if (!mem)
-        mem = VirtualAlloc(nullptr, allocSize, MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE);
-
-    return mem;
-}
-
-#else
-
-void* aligned_large_pages_alloc(size_t allocSize) {
-
-    #if defined(__linux__)
-    constexpr size_t alignment = 2 * 1024 * 1024;  // assumed 2MB page size
-    #else
-    constexpr size_t alignment = 4096;  // assumed small page size
-    #endif
-
-    // Round up to multiples of alignment
-    size_t size = ((allocSize + alignment - 1) / alignment) * alignment;
-    void*  mem  = std_aligned_alloc(alignment, size);
-    #if defined(MADV_HUGEPAGE)
-    madvise(mem, size, MADV_HUGEPAGE);
-    #endif
-    return mem;
-}
-
-#endif
-
-
-// aligned_large_pages_free() will free the previously allocated ttmem
-
-#if defined(_WIN32)
-
-void aligned_large_pages_free(void* mem) {
-
-    if (mem && !VirtualFree(mem, 0, MEM_RELEASE))
-    {
-        DWORD err = GetLastError();
-        std::cerr << "Failed to free large page memory. Error code: 0x" << std::hex << err
-                  << std::dec << std::endl;
-        exit(EXIT_FAILURE);
-    }
-}
-
-#else
-
-void aligned_large_pages_free(void* mem) { std_aligned_free(mem); }
-
-#endif
-
-
-namespace WinProcGroup {
-
-#ifndef _WIN32
-
-void bind_this_thread(size_t) {}
-
-#else
-
-namespace {
-// Retrieves logical processor information using Windows-specific
-// API and returns the best node id for the thread with index idx. Original
-// code from Texel by Peter Österlund.
-int best_node(size_t idx) {
-
-    int   threads      = 0;
-    int   nodes        = 0;
-    int   cores        = 0;
-    DWORD returnLength = 0;
-    DWORD byteOffset   = 0;
-
-    // Early exit if the needed API is not available at runtime
-    HMODULE k32  = GetModuleHandle(TEXT("Kernel32.dll"));
-    auto    fun1 = (fun1_t) (void (*)()) GetProcAddress(k32, "GetLogicalProcessorInformationEx");
-    if (!fun1)
-        return -1;
-
-    // First call to GetLogicalProcessorInformationEx() to get returnLength.
-    // We expect the call to fail due to null buffer.
-    if (fun1(RelationAll, nullptr, &returnLength))
-        return -1;
-
-    // Once we know returnLength, allocate the buffer
-    SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX *buffer, *ptr;
-    ptr = buffer = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*) malloc(returnLength);
-
-    // Second call to GetLogicalProcessorInformationEx(), now we expect to succeed
-    if (!fun1(RelationAll, buffer, &returnLength))
-    {
-        free(buffer);
-        return -1;
-    }
-
-    while (byteOffset < returnLength)
-    {
-        if (ptr->Relationship == RelationNumaNode)
-            nodes++;
-
-        else if (ptr->Relationship == RelationProcessorCore)
-        {
-            cores++;
-            threads += (ptr->Processor.Flags == LTP_PC_SMT) ? 2 : 1;
-        }
-
-        assert(ptr->Size);
-        byteOffset += ptr->Size;
-        ptr = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*) (((char*) ptr) + ptr->Size);
-    }
-
-    free(buffer);
-
-    std::vector<int> groups;
-
-    // Run as many threads as possible on the same node until the core limit is
-    // reached, then move on to filling the next node.
-    for (int n = 0; n < nodes; n++)
-        for (int i = 0; i < cores / nodes; i++)
-            groups.push_back(n);
-
-    // In case a core has more than one logical processor (we assume 2) and we
-    // still have threads to allocate, spread them evenly across available nodes.
-    for (int t = 0; t < threads - cores; t++)
-        groups.push_back(t % nodes);
-
-    // If we still have more threads than the total number of logical processors
-    // then return -1 and let the OS to decide what to do.
-    return idx < groups.size() ? groups[idx] : -1;
-}
-}
-
-
-// Sets the group affinity of the current thread
-void bind_this_thread(size_t idx) {
-
-    // Use only local variables to be thread-safe
-    int node = best_node(idx);
-
-    if (node == -1)
-        return;
-
-    // Early exit if the needed API are not available at runtime
-    HMODULE k32  = GetModuleHandle(TEXT("Kernel32.dll"));
-    auto    fun2 = fun2_t((void (*)()) GetProcAddress(k32, "GetNumaNodeProcessorMaskEx"));
-    auto    fun3 = fun3_t((void (*)()) GetProcAddress(k32, "SetThreadGroupAffinity"));
-    auto    fun4 = fun4_t((void (*)()) GetProcAddress(k32, "GetNumaNodeProcessorMask2"));
-    auto    fun5 = fun5_t((void (*)()) GetProcAddress(k32, "GetMaximumProcessorGroupCount"));
-
-    if (!fun2 || !fun3)
-        return;
-
-    if (!fun4 || !fun5)
-    {
-        GROUP_AFFINITY affinity;
-        if (fun2(node, &affinity))                         // GetNumaNodeProcessorMaskEx
-            fun3(GetCurrentThread(), &affinity, nullptr);  // SetThreadGroupAffinity
-    }
-    else
-    {
-        // If a numa node has more than one processor group, we assume they are
-        // sized equal and we spread threads evenly across the groups.
-        USHORT elements, returnedElements;
-        elements                 = fun5();  // GetMaximumProcessorGroupCount
-        GROUP_AFFINITY* affinity = (GROUP_AFFINITY*) malloc(elements * sizeof(GROUP_AFFINITY));
-        if (fun4(node, affinity, elements, &returnedElements))  // GetNumaNodeProcessorMask2
-            fun3(GetCurrentThread(), &affinity[idx % returnedElements],
-                 nullptr);  // SetThreadGroupAffinity
-        free(affinity);
-    }
-}
-
-#endif
-
-}  // namespace WinProcGroup
-
 #ifdef _WIN32
     #include <direct.h>
     #define GETCWD _getcwd
@@ -723,6 +436,27 @@ void bind_this_thread(size_t idx) {
     #define GETCWD getcwd
 #endif
 
+size_t str_to_size_t(const std::string& s) {
+    unsigned long long value = std::stoull(s);
+    if (value > std::numeric_limits<size_t>::max())
+        std::exit(EXIT_FAILURE);
+    return static_cast<size_t>(value);
+}
+
+std::optional<std::string> read_file_to_string(const std::string& path) {
+    std::ifstream f(path, std::ios_base::binary);
+    if (!f)
+        return std::nullopt;
+    return std::string(std::istreambuf_iterator<char>(f), std::istreambuf_iterator<char>());
+}
+
+void remove_whitespace(std::string& s) {
+    s.erase(std::remove_if(s.begin(), s.end(), [](char c) { return std::isspace(c); }), s.end());
+}
+
+bool is_whitespace(std::string_view s) {
+    return std::all_of(s.begin(), s.end(), [](char c) { return std::isspace(c); });
+}
 
 std::string CommandLine::get_binary_directory(std::string argv0) {
     std::string pathSeparator;
diff --git a/src/misc.h b/src/misc.h
index d75b236f..21093769 100644
--- a/src/misc.h
+++ b/src/misc.h
@@ -24,9 +24,11 @@
 #include <chrono>
 #include <cstddef>
 #include <cstdint>
+#include <cstdio>
 #include <iosfwd>
-#include <memory>
+#include <optional>
 #include <string>
+#include <string_view>
 #include <vector>
 
 #define stringify2(x) #x
@@ -34,49 +36,40 @@
 
 namespace Stockfish {
 
+std::string engine_version_info();
 std::string engine_info(bool to_uci = false);
 std::string compiler_info();
 
 // Preloads the given address in L1/L2 cache. This is a non-blocking
 // function that doesn't stall the CPU waiting for data to be loaded from memory,
 // which can be quite slow.
-void prefetch(void* addr);
+void prefetch(const void* addr);
 
-void  start_logger(const std::string& fname);
-void* std_aligned_alloc(size_t alignment, size_t size);
-void  std_aligned_free(void* ptr);
-// memory aligned by page size, min alignment: 4096 bytes
-void* aligned_large_pages_alloc(size_t size);
-// nop if mem == nullptr
-void aligned_large_pages_free(void* mem);
+void start_logger(const std::string& fname);
 
-// Deleter for automating release of memory area
-template<typename T>
-struct AlignedDeleter {
-    void operator()(T* ptr) const {
-        ptr->~T();
-        std_aligned_free(ptr);
+size_t str_to_size_t(const std::string& s);
+
+#if defined(__linux__)
+
+struct PipeDeleter {
+    void operator()(FILE* file) const {
+        if (file != nullptr)
+        {
+            pclose(file);
+        }
     }
 };
 
-template<typename T>
-struct LargePageDeleter {
-    void operator()(T* ptr) const {
-        ptr->~T();
-        aligned_large_pages_free(ptr);
-    }
-};
-
-template<typename T>
-using AlignedPtr = std::unique_ptr<T, AlignedDeleter<T>>;
-
-template<typename T>
-using LargePagePtr = std::unique_ptr<T, LargePageDeleter<T>>;
+#endif
 
+// Reads the file as bytes.
+// Returns std::nullopt if the file does not exist.
+std::optional<std::string> read_file_to_string(const std::string& path);
 
 void dbg_hit_on(bool cond, int slot = 0);
 void dbg_mean_of(int64_t value, int slot = 0);
 void dbg_stdev_of(int64_t value, int slot = 0);
+void dbg_extremes_of(int64_t value, int slot = 0);
 void dbg_correl_of(int64_t value1, int64_t value2, int slot = 0);
 void dbg_print();
 
@@ -88,6 +81,30 @@ inline TimePoint now() {
       .count();
 }
 
+inline std::vector<std::string_view> split(std::string_view s, std::string_view delimiter) {
+    std::vector<std::string_view> res;
+
+    if (s.empty())
+        return res;
+
+    size_t begin = 0;
+    for (;;)
+    {
+        const size_t end = s.find(delimiter, begin);
+        if (end == std::string::npos)
+            break;
+
+        res.emplace_back(s.substr(begin, end - begin));
+        begin = end + delimiter.size();
+    }
+
+    res.emplace_back(s.substr(begin));
+
+    return res;
+}
+
+void remove_whitespace(std::string& s);
+bool is_whitespace(std::string_view s);
 
 enum SyncCout {
     IO_LOCK,
@@ -98,19 +115,8 @@ std::ostream& operator<<(std::ostream&, SyncCout);
 #define sync_cout std::cout << IO_LOCK
 #define sync_endl std::endl << IO_UNLOCK
 
-
-// Get the first aligned element of an array.
-// ptr must point to an array of size at least `sizeof(T) * N + alignment` bytes,
-// where N is the number of elements in the array.
-template<uintptr_t Alignment, typename T>
-T* align_ptr_up(T* ptr) {
-    static_assert(alignof(T) < Alignment);
-
-    const uintptr_t ptrint = reinterpret_cast<uintptr_t>(reinterpret_cast<char*>(ptr));
-    return reinterpret_cast<T*>(
-      reinterpret_cast<char*>((ptrint + (Alignment - 1)) / Alignment * Alignment));
-}
-
+void sync_cout_start();
+void sync_cout_end();
 
 // True if and only if the binary is compiled on a little-endian machine
 static inline const union {
@@ -194,15 +200,6 @@ inline uint64_t mul_hi64(uint64_t a, uint64_t b) {
 #endif
 }
 
-// Under Windows it is not possible for a process to run on more than one
-// logical processor group. This usually means being limited to using max 64
-// cores. To overcome this, some special platform-specific API should be
-// called to set group affinity for each thread. Original code from Texel by
-// Peter Österlund.
-namespace WinProcGroup {
-void bind_this_thread(size_t idx);
-}
-
 
 struct CommandLine {
    public:
diff --git a/src/movegen.cpp b/src/movegen.cpp
index e6923067..69b8fe6a 100644
--- a/src/movegen.cpp
+++ b/src/movegen.cpp
@@ -75,17 +75,6 @@ ExtMove* generate_pawn_moves(const Position& pos, ExtMove* moveList, Bitboard ta
             b2 &= target;
         }
 
-        if constexpr (Type == QUIET_CHECKS)
-        {
-            // To make a quiet check, you either make a direct check by pushing a pawn
-            // or push a blocker pawn that is not on the same file as the enemy king.
-            // Discovered check promotion has been already generated amongst the captures.
-            Square   ksq              = pos.square<KING>(Them);
-            Bitboard dcCandidatePawns = pos.blockers_for_king(Them) & ~file_bb(ksq);
-            b1 &= pawn_attacks_bb(Them, ksq) | shift<Up>(dcCandidatePawns);
-            b2 &= pawn_attacks_bb(Them, ksq) | shift<Up + Up>(dcCandidatePawns);
-        }
-
         while (b1)
         {
             Square to   = pop_lsb(b1);
@@ -158,7 +147,7 @@ ExtMove* generate_pawn_moves(const Position& pos, ExtMove* moveList, Bitboard ta
 }
 
 
-template<Color Us, PieceType Pt, bool Checks>
+template<Color Us, PieceType Pt>
 ExtMove* generate_moves(const Position& pos, ExtMove* moveList, Bitboard target) {
 
     static_assert(Pt != KING && Pt != PAWN, "Unsupported piece type in generate_moves()");
@@ -170,10 +159,6 @@ ExtMove* generate_moves(const Position& pos, ExtMove* moveList, Bitboard target)
         Square   from = pop_lsb(bb);
         Bitboard b    = attacks_bb<Pt>(from, pos.pieces()) & target;
 
-        // To check, you either move freely a blocker or make a direct check.
-        if (Checks && (Pt == QUEEN || !(pos.blockers_for_king(~Us) & from)))
-            b &= pos.check_squares(Pt);
-
         while (b)
             *moveList++ = Move(from, pop_lsb(b));
     }
@@ -187,9 +172,8 @@ ExtMove* generate_all(const Position& pos, ExtMove* moveList) {
 
     static_assert(Type != LEGAL, "Unsupported type in generate_all()");
 
-    constexpr bool Checks = Type == QUIET_CHECKS;  // Reduce template instantiations
-    const Square   ksq    = pos.square<KING>(Us);
-    Bitboard       target;
+    const Square ksq = pos.square<KING>(Us);
+    Bitboard     target;
 
     // Skip generating non-king moves when in double check
     if (Type != EVASIONS || !more_than_one(pos.checkers()))
@@ -197,29 +181,24 @@ ExtMove* generate_all(const Position& pos, ExtMove* moveList) {
         target = Type == EVASIONS     ? between_bb(ksq, lsb(pos.checkers()))
                : Type == NON_EVASIONS ? ~pos.pieces(Us)
                : Type == CAPTURES     ? pos.pieces(~Us)
-                                      : ~pos.pieces();  // QUIETS || QUIET_CHECKS
+                                      : ~pos.pieces();  // QUIETS
 
         moveList = generate_pawn_moves<Us, Type>(pos, moveList, target);
-        moveList = generate_moves<Us, KNIGHT, Checks>(pos, moveList, target);
-        moveList = generate_moves<Us, BISHOP, Checks>(pos, moveList, target);
-        moveList = generate_moves<Us, ROOK, Checks>(pos, moveList, target);
-        moveList = generate_moves<Us, QUEEN, Checks>(pos, moveList, target);
+        moveList = generate_moves<Us, KNIGHT>(pos, moveList, target);
+        moveList = generate_moves<Us, BISHOP>(pos, moveList, target);
+        moveList = generate_moves<Us, ROOK>(pos, moveList, target);
+        moveList = generate_moves<Us, QUEEN>(pos, moveList, target);
     }
 
-    if (!Checks || pos.blockers_for_king(~Us) & ksq)
-    {
-        Bitboard b = attacks_bb<KING>(ksq) & (Type == EVASIONS ? ~pos.pieces(Us) : target);
-        if (Checks)
-            b &= ~attacks_bb<QUEEN>(pos.square<KING>(~Us));
+    Bitboard b = attacks_bb<KING>(ksq) & (Type == EVASIONS ? ~pos.pieces(Us) : target);
 
-        while (b)
-            *moveList++ = Move(ksq, pop_lsb(b));
+    while (b)
+        *moveList++ = Move(ksq, pop_lsb(b));
 
-        if ((Type == QUIETS || Type == NON_EVASIONS) && pos.can_castle(Us & ANY_CASTLING))
-            for (CastlingRights cr : {Us & KING_SIDE, Us & QUEEN_SIDE})
-                if (!pos.castling_impeded(cr) && pos.can_castle(cr))
-                    *moveList++ = Move::make<CASTLING>(ksq, pos.castling_rook_square(cr));
-    }
+    if ((Type == QUIETS || Type == NON_EVASIONS) && pos.can_castle(Us & ANY_CASTLING))
+        for (CastlingRights cr : {Us & KING_SIDE, Us & QUEEN_SIDE})
+            if (!pos.castling_impeded(cr) && pos.can_castle(cr))
+                *moveList++ = Move::make<CASTLING>(ksq, pos.castling_rook_square(cr));
 
     return moveList;
 }
@@ -231,8 +210,6 @@ ExtMove* generate_all(const Position& pos, ExtMove* moveList) {
 // <QUIETS>       Generates all pseudo-legal non-captures and underpromotions
 // <EVASIONS>     Generates all pseudo-legal check evasions
 // <NON_EVASIONS> Generates all pseudo-legal captures and non-captures
-// <QUIET_CHECKS> Generates all pseudo-legal non-captures giving check,
-//                except castling and promotions
 //
 // Returns a pointer to the end of the move list.
 template<GenType Type>
@@ -251,7 +228,6 @@ ExtMove* generate(const Position& pos, ExtMove* moveList) {
 template ExtMove* generate<CAPTURES>(const Position&, ExtMove*);
 template ExtMove* generate<QUIETS>(const Position&, ExtMove*);
 template ExtMove* generate<EVASIONS>(const Position&, ExtMove*);
-template ExtMove* generate<QUIET_CHECKS>(const Position&, ExtMove*);
 template ExtMove* generate<NON_EVASIONS>(const Position&, ExtMove*);
 
 
diff --git a/src/movegen.h b/src/movegen.h
index 5f650d2e..f067f880 100644
--- a/src/movegen.h
+++ b/src/movegen.h
@@ -31,7 +31,6 @@ class Position;
 enum GenType {
     CAPTURES,
     QUIETS,
-    QUIET_CHECKS,
     EVASIONS,
     NON_EVASIONS,
     LEGAL
diff --git a/src/movepick.cpp b/src/movepick.cpp
index 4a93662d..96f03171 100644
--- a/src/movepick.cpp
+++ b/src/movepick.cpp
@@ -18,10 +18,9 @@
 
 #include "movepick.h"
 
-#include <algorithm>
+#include <array>
 #include <cassert>
-#include <iterator>
-#include <utility>
+#include <limits>
 
 #include "bitboard.h"
 #include "position.h"
@@ -35,7 +34,6 @@ enum Stages {
     MAIN_TT,
     CAPTURE_INIT,
     GOOD_CAPTURE,
-    REFUTATION,
     QUIET_INIT,
     GOOD_QUIET,
     BAD_CAPTURE,
@@ -54,13 +52,11 @@ enum Stages {
     // generate qsearch moves
     QSEARCH_TT,
     QCAPTURE_INIT,
-    QCAPTURE,
-    QCHECK_INIT,
-    QCHECK
+    QCAPTURE
 };
 
-// Sort moves in descending order up to and including
-// a given limit. The order of moves smaller than the limit is left unspecified.
+// Sort moves in descending order up to and including a given limit.
+// The order of moves smaller than the limit is left unspecified.
 void partial_insertion_sort(ExtMove* begin, ExtMove* end, int limit) {
 
     for (ExtMove *sortedEnd = begin, *p = begin + 1; p < end; ++p)
@@ -78,56 +74,38 @@ void partial_insertion_sort(ExtMove* begin, ExtMove* end, int limit) {
 
 
 // Constructors of the MovePicker class. As arguments, we pass information
-// to help it return the (presumably) good moves first, to decide which
-// moves to return (in the quiescence search, for instance, we only want to
-// search captures, promotions, and some checks) and how important a good
-// move ordering is at the current node.
+// to decide which class of moves to emit, to help sorting the (presumably)
+// good moves first, and how important move ordering is at the current node.
 
-// MovePicker constructor for the main search
+// MovePicker constructor for the main search and for the quiescence search
 MovePicker::MovePicker(const Position&              p,
                        Move                         ttm,
                        Depth                        d,
                        const ButterflyHistory*      mh,
+                       const LowPlyHistory*         lph,
                        const CapturePieceToHistory* cph,
                        const PieceToHistory**       ch,
                        const PawnHistory*           ph,
-                       Move                         cm,
-                       const Move*                  killers) :
+                       int                          pl) :
     pos(p),
     mainHistory(mh),
+    lowPlyHistory(lph),
     captureHistory(cph),
     continuationHistory(ch),
     pawnHistory(ph),
     ttMove(ttm),
-    refutations{{killers[0], 0}, {killers[1], 0}, {cm, 0}},
-    depth(d) {
-    assert(d > 0);
+    depth(d),
+    ply(pl) {
 
-    stage = (pos.checkers() ? EVASION_TT : MAIN_TT) + !(ttm && pos.pseudo_legal(ttm));
+    if (pos.checkers())
+        stage = EVASION_TT + !(ttm && pos.pseudo_legal(ttm));
+
+    else
+        stage = (depth > 0 ? MAIN_TT : QSEARCH_TT) + !(ttm && pos.pseudo_legal(ttm));
 }
 
-// Constructor for quiescence search
-MovePicker::MovePicker(const Position&              p,
-                       Move                         ttm,
-                       Depth                        d,
-                       const ButterflyHistory*      mh,
-                       const CapturePieceToHistory* cph,
-                       const PieceToHistory**       ch,
-                       const PawnHistory*           ph) :
-    pos(p),
-    mainHistory(mh),
-    captureHistory(cph),
-    continuationHistory(ch),
-    pawnHistory(ph),
-    ttMove(ttm),
-    depth(d) {
-    assert(d <= 0);
-
-    stage = (pos.checkers() ? EVASION_TT : QSEARCH_TT) + !(ttm && pos.pseudo_legal(ttm));
-}
-
-// Constructor for ProbCut: we generate captures with SEE greater
-// than or equal to the given threshold.
+// MovePicker constructor for ProbCut: we generate captures with Static Exchange
+// Evaluation (SEE) greater than or equal to the given threshold.
 MovePicker::MovePicker(const Position& p, Move ttm, int th, const CapturePieceToHistory* cph) :
     pos(p),
     captureHistory(cph),
@@ -139,9 +117,9 @@ MovePicker::MovePicker(const Position& p, Move ttm, int th, const CapturePieceTo
           + !(ttm && pos.capture_stage(ttm) && pos.pseudo_legal(ttm) && pos.see_ge(ttm, threshold));
 }
 
-// Assigns a numerical value to each move in a list, used
-// for sorting. Captures are ordered by Most Valuable Victim (MVV), preferring
-// captures with a good history. Quiets moves are ordered using the history tables.
+// Assigns a numerical value to each move in a list, used for sorting.
+// Captures are ordered by Most Valuable Victim (MVV), preferring captures
+// with a good history. Quiets moves are ordered using the history tables.
 template<GenType Type>
 void MovePicker::score() {
 
@@ -180,9 +158,9 @@ void MovePicker::score() {
             // histories
             m.value = 2 * (*mainHistory)[pos.side_to_move()][m.from_to()];
             m.value += 2 * (*pawnHistory)[pawn_structure_index(pos)][pc][to];
-            m.value += 2 * (*continuationHistory[0])[pc][to];
+            m.value += (*continuationHistory[0])[pc][to];
             m.value += (*continuationHistory[1])[pc][to];
-            m.value += (*continuationHistory[2])[pc][to] / 4;
+            m.value += (*continuationHistory[2])[pc][to];
             m.value += (*continuationHistory[3])[pc][to];
             m.value += (*continuationHistory[5])[pc][to];
 
@@ -197,13 +175,12 @@ void MovePicker::score() {
                                                : 0;
 
             // malus for putting piece en prise
-            m.value -= !(threatenedPieces & from)
-                       ? (pt == QUEEN ? bool(to & threatenedByRook) * 48150
-                                          + bool(to & threatenedByMinor) * 10650
-                          : pt == ROOK ? bool(to & threatenedByMinor) * 24335
-                          : pt != PAWN ? bool(to & threatenedByPawn) * 14950
-                                       : 0)
-                       : 0;
+            m.value -= (pt == QUEEN ? bool(to & threatenedByRook) * 49000
+                        : pt == ROOK && bool(to & threatenedByMinor) ? 24335
+                                                                     : 0);
+
+            if (ply < LOW_PLY_HISTORY_SIZE)
+                m.value += 8 * (*lowPlyHistory)[ply][m.from_to()] / (1 + 2 * ply);
         }
 
         else  // Type == EVASIONS
@@ -219,27 +196,21 @@ void MovePicker::score() {
 }
 
 // Returns the next move satisfying a predicate function.
-// It never returns the TT move.
-template<MovePicker::PickType T, typename Pred>
+// This never returns the TT move, as it was emitted before.
+template<typename Pred>
 Move MovePicker::select(Pred filter) {
 
-    while (cur < endMoves)
-    {
-        if constexpr (T == Best)
-            std::swap(*cur, *std::max_element(cur, endMoves));
-
+    for (; cur < endMoves; ++cur)
         if (*cur != ttMove && filter())
             return *cur++;
 
-        cur++;
-    }
     return Move::none();
 }
 
-// Most important method of the MovePicker class. It
-// returns a new pseudo-legal move every time it is called until there are no more
-// moves left, picking the move with the highest score from a list of generated moves.
-Move MovePicker::next_move(bool skipQuiets) {
+// This is the most important method of the MovePicker class. We emit one
+// new pseudo-legal move on every call until there are no more moves left,
+// picking the move with the highest score from a list of generated moves.
+Move MovePicker::next_move() {
 
     auto quiet_threshold = [](Depth d) { return -3560 * d; };
 
@@ -266,29 +237,13 @@ top:
         goto top;
 
     case GOOD_CAPTURE :
-        if (select<Next>([&]() {
+        if (select([&]() {
                 // Move losing capture to endBadCaptures to be tried later
                 return pos.see_ge(*cur, -cur->value / 18) ? true
                                                           : (*endBadCaptures++ = *cur, false);
             }))
             return *(cur - 1);
 
-        // Prepare the pointers to loop over the refutations array
-        cur      = std::begin(refutations);
-        endMoves = std::end(refutations);
-
-        // If the countermove is the same as a killer, skip it
-        if (refutations[0] == refutations[2] || refutations[1] == refutations[2])
-            --endMoves;
-
-        ++stage;
-        [[fallthrough]];
-
-    case REFUTATION :
-        if (select<Next>([&]() {
-                return *cur != Move::none() && !pos.capture_stage(*cur) && pos.pseudo_legal(*cur);
-            }))
-            return *(cur - 1);
         ++stage;
         [[fallthrough]];
 
@@ -306,9 +261,7 @@ top:
         [[fallthrough]];
 
     case GOOD_QUIET :
-        if (!skipQuiets && select<Next>([&]() {
-                return *cur != refutations[0] && *cur != refutations[1] && *cur != refutations[2];
-            }))
+        if (!skipQuiets && select([]() { return true; }))
         {
             if ((cur - 1)->value > -7998 || (cur - 1)->value <= quiet_threshold(depth))
                 return *(cur - 1);
@@ -325,7 +278,7 @@ top:
         [[fallthrough]];
 
     case BAD_CAPTURE :
-        if (select<Next>([]() { return true; }))
+        if (select([]() { return true; }))
             return *(cur - 1);
 
         // Prepare the pointers to loop over the bad quiets
@@ -337,9 +290,7 @@ top:
 
     case BAD_QUIET :
         if (!skipQuiets)
-            return select<Next>([&]() {
-                return *cur != refutations[0] && *cur != refutations[1] && *cur != refutations[2];
-            });
+            return select([]() { return true; });
 
         return Move::none();
 
@@ -348,39 +299,22 @@ top:
         endMoves = generate<EVASIONS>(pos, cur);
 
         score<EVASIONS>();
+        partial_insertion_sort(cur, endMoves, std::numeric_limits<int>::min());
         ++stage;
         [[fallthrough]];
 
     case EVASION :
-        return select<Best>([]() { return true; });
+    case QCAPTURE :
+        return select([]() { return true; });
 
     case PROBCUT :
-        return select<Next>([&]() { return pos.see_ge(*cur, threshold); });
-
-    case QCAPTURE :
-        if (select<Next>([]() { return true; }))
-            return *(cur - 1);
-
-        // If we did not find any move and we do not try checks, we have finished
-        if (depth != DEPTH_QS_CHECKS)
-            return Move::none();
-
-        ++stage;
-        [[fallthrough]];
-
-    case QCHECK_INIT :
-        cur      = moves;
-        endMoves = generate<QUIET_CHECKS>(pos, cur);
-
-        ++stage;
-        [[fallthrough]];
-
-    case QCHECK :
-        return select<Next>([]() { return true; });
+        return select([&]() { return pos.see_ge(*cur, threshold); });
     }
 
     assert(false);
     return Move::none();  // Silence warning
 }
 
+void MovePicker::skip_quiet_moves() { skipQuiets = true; }
+
 }  // namespace Stockfish
diff --git a/src/movepick.h b/src/movepick.h
index b81f76e1..ab4e832f 100644
--- a/src/movepick.h
+++ b/src/movepick.h
@@ -19,141 +19,22 @@
 #ifndef MOVEPICK_H_INCLUDED
 #define MOVEPICK_H_INCLUDED
 
-#include <algorithm>
-#include <array>
-#include <cassert>
-#include <cmath>
-#include <cstdint>
-#include <cstdlib>
-#include <limits>
-#include <type_traits>  // IWYU pragma: keep
-
+#include "history.h"
 #include "movegen.h"
-#include "position.h"
 #include "types.h"
 
 namespace Stockfish {
 
-constexpr int PAWN_HISTORY_SIZE        = 512;    // has to be a power of 2
-constexpr int CORRECTION_HISTORY_SIZE  = 16384;  // has to be a power of 2
-constexpr int CORRECTION_HISTORY_LIMIT = 1024;
+class Position;
 
-static_assert((PAWN_HISTORY_SIZE & (PAWN_HISTORY_SIZE - 1)) == 0,
-              "PAWN_HISTORY_SIZE has to be a power of 2");
-
-static_assert((CORRECTION_HISTORY_SIZE & (CORRECTION_HISTORY_SIZE - 1)) == 0,
-              "CORRECTION_HISTORY_SIZE has to be a power of 2");
-
-enum PawnHistoryType {
-    Normal,
-    Correction
-};
-
-template<PawnHistoryType T = Normal>
-inline int pawn_structure_index(const Position& pos) {
-    return pos.pawn_key() & ((T == Normal ? PAWN_HISTORY_SIZE : CORRECTION_HISTORY_SIZE) - 1);
-}
-
-// StatsEntry stores the stat table value. It is usually a number but could
-// be a move or even a nested history. We use a class instead of a naked value
-// to directly call history update operator<<() on the entry so to use stats
-// tables at caller sites as simple multi-dim arrays.
-template<typename T, int D>
-class StatsEntry {
-
-    T entry;
-
-   public:
-    void operator=(const T& v) { entry = v; }
-    T*   operator&() { return &entry; }
-    T*   operator->() { return &entry; }
-    operator const T&() const { return entry; }
-
-    void operator<<(int bonus) {
-        static_assert(D <= std::numeric_limits<T>::max(), "D overflows T");
-
-        // Make sure that bonus is in range [-D, D]
-        int clampedBonus = std::clamp(bonus, -D, D);
-        entry += clampedBonus - entry * std::abs(clampedBonus) / D;
-
-        assert(std::abs(entry) <= D);
-    }
-};
-
-// Stats is a generic N-dimensional array used to store various statistics.
-// The first template parameter T is the base type of the array, and the second
-// template parameter D limits the range of updates in [-D, D] when we update
-// values with the << operator, while the last parameters (Size and Sizes)
-// encode the dimensions of the array.
-template<typename T, int D, int Size, int... Sizes>
-struct Stats: public std::array<Stats<T, D, Sizes...>, Size> {
-    using stats = Stats<T, D, Size, Sizes...>;
-
-    void fill(const T& v) {
-
-        // For standard-layout 'this' points to the first struct member
-        assert(std::is_standard_layout_v<stats>);
-
-        using entry = StatsEntry<T, D>;
-        entry* p    = reinterpret_cast<entry*>(this);
-        std::fill(p, p + sizeof(*this) / sizeof(entry), v);
-    }
-};
-
-template<typename T, int D, int Size>
-struct Stats<T, D, Size>: public std::array<StatsEntry<T, D>, Size> {};
-
-// In stats table, D=0 means that the template parameter is not used
-enum StatsParams {
-    NOT_USED = 0
-};
-enum StatsType {
-    NoCaptures,
-    Captures
-};
-
-// ButterflyHistory records how often quiet moves have been successful or unsuccessful
-// during the current search, and is used for reduction and move ordering decisions.
-// It uses 2 tables (one for each color) indexed by the move's from and to squares,
-// see www.chessprogramming.org/Butterfly_Boards (~11 elo)
-using ButterflyHistory = Stats<int16_t, 7183, COLOR_NB, int(SQUARE_NB) * int(SQUARE_NB)>;
-
-// CounterMoveHistory stores counter moves indexed by [piece][to] of the previous
-// move, see www.chessprogramming.org/Countermove_Heuristic
-using CounterMoveHistory = Stats<Move, NOT_USED, PIECE_NB, SQUARE_NB>;
-
-// CapturePieceToHistory is addressed by a move's [piece][to][captured piece type]
-using CapturePieceToHistory = Stats<int16_t, 10692, PIECE_NB, SQUARE_NB, PIECE_TYPE_NB>;
-
-// PieceToHistory is like ButterflyHistory but is addressed by a move's [piece][to]
-using PieceToHistory = Stats<int16_t, 29952, PIECE_NB, SQUARE_NB>;
-
-// ContinuationHistory is the combined history of a given pair of moves, usually
-// the current one given a previous one. The nested history table is based on
-// PieceToHistory instead of ButterflyBoards.
-// (~63 elo)
-using ContinuationHistory = Stats<PieceToHistory, NOT_USED, PIECE_NB, SQUARE_NB>;
-
-// PawnHistory is addressed by the pawn structure and a move's [piece][to]
-using PawnHistory = Stats<int16_t, 8192, PAWN_HISTORY_SIZE, PIECE_NB, SQUARE_NB>;
-
-// CorrectionHistory is addressed by color and pawn structure
-using CorrectionHistory =
-  Stats<int16_t, CORRECTION_HISTORY_LIMIT, COLOR_NB, CORRECTION_HISTORY_SIZE>;
-
-// MovePicker class is used to pick one pseudo-legal move at a time from the
-// current position. The most important method is next_move(), which returns a
-// new pseudo-legal move each time it is called, until there are no moves left,
-// when Move::none() is returned. In order to improve the efficiency of the
-// alpha-beta algorithm, MovePicker attempts to return the moves which are most
-// likely to get a cut-off first.
+// The MovePicker class is used to pick one pseudo-legal move at a time from the
+// current position. The most important method is next_move(), which emits one
+// new pseudo-legal move on every call, until there are no moves left, when
+// Move::none() is returned. In order to improve the efficiency of the alpha-beta
+// algorithm, MovePicker attempts to return the moves which are most likely to get
+// a cut-off first.
 class MovePicker {
 
-    enum PickType {
-        Next,
-        Best
-    };
-
    public:
     MovePicker(const MovePicker&)            = delete;
     MovePicker& operator=(const MovePicker&) = delete;
@@ -161,23 +42,17 @@ class MovePicker {
                Move,
                Depth,
                const ButterflyHistory*,
+               const LowPlyHistory*,
                const CapturePieceToHistory*,
                const PieceToHistory**,
                const PawnHistory*,
-               Move,
-               const Move*);
-    MovePicker(const Position&,
-               Move,
-               Depth,
-               const ButterflyHistory*,
-               const CapturePieceToHistory*,
-               const PieceToHistory**,
-               const PawnHistory*);
+               int);
     MovePicker(const Position&, Move, int, const CapturePieceToHistory*);
-    Move next_move(bool skipQuiets = false);
+    Move next_move();
+    void skip_quiet_moves();
 
    private:
-    template<PickType T, typename Pred>
+    template<typename Pred>
     Move select(Pred);
     template<GenType>
     void     score();
@@ -186,15 +61,18 @@ class MovePicker {
 
     const Position&              pos;
     const ButterflyHistory*      mainHistory;
+    const LowPlyHistory*         lowPlyHistory;
     const CapturePieceToHistory* captureHistory;
     const PieceToHistory**       continuationHistory;
     const PawnHistory*           pawnHistory;
     Move                         ttMove;
-    ExtMove refutations[3], *cur, *endMoves, *endBadCaptures, *beginBadQuiets, *endBadQuiets;
-    int     stage;
-    int     threshold;
-    Depth   depth;
-    ExtMove moves[MAX_MOVES];
+    ExtMove *                    cur, *endMoves, *endBadCaptures, *beginBadQuiets, *endBadQuiets;
+    int                          stage;
+    int                          threshold;
+    Depth                        depth;
+    int                          ply;
+    bool                         skipQuiets = false;
+    ExtMove                      moves[MAX_MOVES];
 };
 
 }  // namespace Stockfish
diff --git a/src/nnue/layers/affine_transform.h b/src/nnue/layers/affine_transform.h
index ad9167c0..59a6149f 100644
--- a/src/nnue/layers/affine_transform.h
+++ b/src/nnue/layers/affine_transform.h
@@ -39,25 +39,26 @@
 
 namespace Stockfish::Eval::NNUE::Layers {
 
+#if defined(USE_SSSE3) || defined(USE_NEON_DOTPROD)
+    #define ENABLE_SEQ_OPT
+#endif
+
 // Fallback implementation for older/other architectures.
 // Requires the input to be padded to at least 16 values.
-#if !defined(USE_SSSE3)
+#ifndef ENABLE_SEQ_OPT
+
 template<IndexType InputDimensions, IndexType PaddedInputDimensions, IndexType OutputDimensions>
 static void affine_transform_non_ssse3(std::int32_t*       output,
                                        const std::int8_t*  weights,
                                        const std::int32_t* biases,
                                        const std::uint8_t* input) {
-    #if defined(USE_SSE2) || defined(USE_NEON_DOTPROD) || defined(USE_NEON)
+    #if defined(USE_SSE2) || defined(USE_NEON)
         #if defined(USE_SSE2)
     // At least a multiple of 16, with SSE2.
     constexpr IndexType NumChunks   = ceil_to_multiple<IndexType>(InputDimensions, 16) / 16;
     const __m128i       Zeros       = _mm_setzero_si128();
     const auto          inputVector = reinterpret_cast<const __m128i*>(input);
 
-        #elif defined(USE_NEON_DOTPROD)
-    constexpr IndexType NumChunks   = ceil_to_multiple<IndexType>(InputDimensions, 16) / 16;
-    const auto          inputVector = reinterpret_cast<const int8x16_t*>(input);
-
         #elif defined(USE_NEON)
     constexpr IndexType NumChunks   = ceil_to_multiple<IndexType>(InputDimensions, 16) / 16;
     const auto          inputVector = reinterpret_cast<const int8x8_t*>(input);
@@ -91,16 +92,8 @@ static void affine_transform_non_ssse3(std::int32_t*       output,
         sum                   = _mm_add_epi32(sum, sum_second_32);
         output[i]             = _mm_cvtsi128_si32(sum);
 
-        #elif defined(USE_NEON_DOTPROD)
-        int32x4_t  sum = {biases[i]};
-        const auto row = reinterpret_cast<const int8x16_t*>(&weights[offset]);
-        for (IndexType j = 0; j < NumChunks; ++j)
-        {
-            sum = vdotq_s32(sum, inputVector[j], row[j]);
-        }
-        output[i] = vaddvq_s32(sum);
-
         #elif defined(USE_NEON)
+
         int32x4_t  sum = {biases[i]};
         const auto row = reinterpret_cast<const int8x8_t*>(&weights[offset]);
         for (IndexType j = 0; j < NumChunks; ++j)
@@ -127,7 +120,8 @@ static void affine_transform_non_ssse3(std::int32_t*       output,
         }
     #endif
 }
-#endif
+
+#endif  // !ENABLE_SEQ_OPT
 
 template<IndexType InDims, IndexType OutDims>
 class AffineTransform {
@@ -162,7 +156,7 @@ class AffineTransform {
     }
 
     static constexpr IndexType get_weight_index(IndexType i) {
-#if defined(USE_SSSE3)
+#ifdef ENABLE_SEQ_OPT
         return get_weight_index_scrambled(i);
 #else
         return i;
@@ -190,29 +184,28 @@ class AffineTransform {
     // Forward propagation
     void propagate(const InputType* input, OutputType* output) const {
 
-#if defined(USE_SSSE3)
+#ifdef ENABLE_SEQ_OPT
 
         if constexpr (OutputDimensions > 1)
         {
-
     #if defined(USE_AVX512)
             using vec_t = __m512i;
-        #define vec_setzero _mm512_setzero_si512
         #define vec_set_32 _mm512_set1_epi32
         #define vec_add_dpbusd_32 Simd::m512_add_dpbusd_epi32
-        #define vec_hadd Simd::m512_hadd
     #elif defined(USE_AVX2)
             using vec_t = __m256i;
-        #define vec_setzero _mm256_setzero_si256
         #define vec_set_32 _mm256_set1_epi32
         #define vec_add_dpbusd_32 Simd::m256_add_dpbusd_epi32
-        #define vec_hadd Simd::m256_hadd
     #elif defined(USE_SSSE3)
             using vec_t = __m128i;
-        #define vec_setzero _mm_setzero_si128
         #define vec_set_32 _mm_set1_epi32
         #define vec_add_dpbusd_32 Simd::m128_add_dpbusd_epi32
-        #define vec_hadd Simd::m128_hadd
+    #elif defined(USE_NEON_DOTPROD)
+            using vec_t = int32x4_t;
+        #define vec_set_32 vdupq_n_s32
+        #define vec_add_dpbusd_32(acc, a, b) \
+            Simd::dotprod_m128_add_dpbusd_epi32(acc, vreinterpretq_s8_s32(a), \
+                                                vreinterpretq_s8_s32(b))
     #endif
 
             static constexpr IndexType OutputSimdWidth = sizeof(vec_t) / sizeof(OutputType);
@@ -242,28 +235,33 @@ class AffineTransform {
             for (IndexType k = 0; k < NumRegs; ++k)
                 outptr[k] = acc[k];
 
-    #undef vec_setzero
     #undef vec_set_32
     #undef vec_add_dpbusd_32
-    #undef vec_hadd
         }
         else if constexpr (OutputDimensions == 1)
         {
-
     // We cannot use AVX512 for the last layer because there are only 32 inputs
     // and the buffer is not padded to 64 elements.
     #if defined(USE_AVX2)
             using vec_t = __m256i;
-        #define vec_setzero _mm256_setzero_si256
+        #define vec_setzero() _mm256_setzero_si256()
         #define vec_set_32 _mm256_set1_epi32
         #define vec_add_dpbusd_32 Simd::m256_add_dpbusd_epi32
         #define vec_hadd Simd::m256_hadd
     #elif defined(USE_SSSE3)
             using vec_t = __m128i;
-        #define vec_setzero _mm_setzero_si128
+        #define vec_setzero() _mm_setzero_si128()
         #define vec_set_32 _mm_set1_epi32
         #define vec_add_dpbusd_32 Simd::m128_add_dpbusd_epi32
         #define vec_hadd Simd::m128_hadd
+    #elif defined(USE_NEON_DOTPROD)
+            using vec_t = int32x4_t;
+        #define vec_setzero() vdupq_n_s32(0)
+        #define vec_set_32 vdupq_n_s32
+        #define vec_add_dpbusd_32(acc, a, b) \
+            Simd::dotprod_m128_add_dpbusd_epi32(acc, vreinterpretq_s8_s32(a), \
+                                                vreinterpretq_s8_s32(b))
+        #define vec_hadd Simd::neon_m128_hadd
     #endif
 
             const auto inputVector = reinterpret_cast<const vec_t*>(input);
diff --git a/src/nnue/layers/clipped_relu.h b/src/nnue/layers/clipped_relu.h
index 813234c5..2ee378ad 100644
--- a/src/nnue/layers/clipped_relu.h
+++ b/src/nnue/layers/clipped_relu.h
@@ -65,41 +65,37 @@ class ClippedReLU {
         if constexpr (InputDimensions % SimdWidth == 0)
         {
             constexpr IndexType NumChunks = InputDimensions / SimdWidth;
-            const __m256i       Zero      = _mm256_setzero_si256();
             const __m256i       Offsets   = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
             const auto          in        = reinterpret_cast<const __m256i*>(input);
             const auto          out       = reinterpret_cast<__m256i*>(output);
             for (IndexType i = 0; i < NumChunks; ++i)
             {
                 const __m256i words0 =
-                  _mm256_srai_epi16(_mm256_packs_epi32(_mm256_load_si256(&in[i * 4 + 0]),
-                                                       _mm256_load_si256(&in[i * 4 + 1])),
+                  _mm256_srli_epi16(_mm256_packus_epi32(_mm256_load_si256(&in[i * 4 + 0]),
+                                                        _mm256_load_si256(&in[i * 4 + 1])),
                                     WeightScaleBits);
                 const __m256i words1 =
-                  _mm256_srai_epi16(_mm256_packs_epi32(_mm256_load_si256(&in[i * 4 + 2]),
-                                                       _mm256_load_si256(&in[i * 4 + 3])),
+                  _mm256_srli_epi16(_mm256_packus_epi32(_mm256_load_si256(&in[i * 4 + 2]),
+                                                        _mm256_load_si256(&in[i * 4 + 3])),
                                     WeightScaleBits);
-                _mm256_store_si256(
-                  &out[i], _mm256_permutevar8x32_epi32(
-                             _mm256_max_epi8(_mm256_packs_epi16(words0, words1), Zero), Offsets));
+                _mm256_store_si256(&out[i], _mm256_permutevar8x32_epi32(
+                                              _mm256_packs_epi16(words0, words1), Offsets));
             }
         }
         else
         {
             constexpr IndexType NumChunks = InputDimensions / (SimdWidth / 2);
-            const __m128i       Zero      = _mm_setzero_si128();
             const auto          in        = reinterpret_cast<const __m128i*>(input);
             const auto          out       = reinterpret_cast<__m128i*>(output);
             for (IndexType i = 0; i < NumChunks; ++i)
             {
-                const __m128i words0 = _mm_srai_epi16(
-                  _mm_packs_epi32(_mm_load_si128(&in[i * 4 + 0]), _mm_load_si128(&in[i * 4 + 1])),
+                const __m128i words0 = _mm_srli_epi16(
+                  _mm_packus_epi32(_mm_load_si128(&in[i * 4 + 0]), _mm_load_si128(&in[i * 4 + 1])),
                   WeightScaleBits);
-                const __m128i words1 = _mm_srai_epi16(
-                  _mm_packs_epi32(_mm_load_si128(&in[i * 4 + 2]), _mm_load_si128(&in[i * 4 + 3])),
+                const __m128i words1 = _mm_srli_epi16(
+                  _mm_packus_epi32(_mm_load_si128(&in[i * 4 + 2]), _mm_load_si128(&in[i * 4 + 3])),
                   WeightScaleBits);
-                const __m128i packedbytes = _mm_packs_epi16(words0, words1);
-                _mm_store_si128(&out[i], _mm_max_epi8(packedbytes, Zero));
+                _mm_store_si128(&out[i], _mm_packs_epi16(words0, words1));
             }
         }
         constexpr IndexType Start = InputDimensions % SimdWidth == 0
@@ -109,9 +105,7 @@ class ClippedReLU {
 #elif defined(USE_SSE2)
         constexpr IndexType NumChunks = InputDimensions / SimdWidth;
 
-    #ifdef USE_SSE41
-        const __m128i Zero = _mm_setzero_si128();
-    #else
+    #ifndef USE_SSE41
         const __m128i k0x80s = _mm_set1_epi8(-128);
     #endif
 
@@ -119,6 +113,15 @@ class ClippedReLU {
         const auto out = reinterpret_cast<__m128i*>(output);
         for (IndexType i = 0; i < NumChunks; ++i)
         {
+    #if defined(USE_SSE41)
+            const __m128i words0 = _mm_srli_epi16(
+              _mm_packus_epi32(_mm_load_si128(&in[i * 4 + 0]), _mm_load_si128(&in[i * 4 + 1])),
+              WeightScaleBits);
+            const __m128i words1 = _mm_srli_epi16(
+              _mm_packus_epi32(_mm_load_si128(&in[i * 4 + 2]), _mm_load_si128(&in[i * 4 + 3])),
+              WeightScaleBits);
+            _mm_store_si128(&out[i], _mm_packs_epi16(words0, words1));
+    #else
             const __m128i words0 = _mm_srai_epi16(
               _mm_packs_epi32(_mm_load_si128(&in[i * 4 + 0]), _mm_load_si128(&in[i * 4 + 1])),
               WeightScaleBits);
@@ -126,15 +129,8 @@ class ClippedReLU {
               _mm_packs_epi32(_mm_load_si128(&in[i * 4 + 2]), _mm_load_si128(&in[i * 4 + 3])),
               WeightScaleBits);
             const __m128i packedbytes = _mm_packs_epi16(words0, words1);
-            _mm_store_si128(&out[i],
-
-    #ifdef USE_SSE41
-                            _mm_max_epi8(packedbytes, Zero)
-    #else
-                            _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)
+            _mm_store_si128(&out[i], _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s));
     #endif
-
-            );
         }
         constexpr IndexType Start = NumChunks * SimdWidth;
 
diff --git a/src/nnue/layers/simd.h b/src/nnue/layers/simd.h
index cec41474..55cb7df1 100644
--- a/src/nnue/layers/simd.h
+++ b/src/nnue/layers/simd.h
@@ -43,39 +43,6 @@ namespace Stockfish::Simd {
     return _mm512_reduce_add_epi32(sum) + bias;
 }
 
-/*
-      Parameters:
-        sum0 = [zmm0.i128[0], zmm0.i128[1], zmm0.i128[2], zmm0.i128[3]]
-        sum1 = [zmm1.i128[0], zmm1.i128[1], zmm1.i128[2], zmm1.i128[3]]
-        sum2 = [zmm2.i128[0], zmm2.i128[1], zmm2.i128[2], zmm2.i128[3]]
-        sum3 = [zmm3.i128[0], zmm3.i128[1], zmm3.i128[2], zmm3.i128[3]]
-
-      Returns:
-        ret = [
-          reduce_add_epi32(zmm0.i128[0]), reduce_add_epi32(zmm1.i128[0]), reduce_add_epi32(zmm2.i128[0]), reduce_add_epi32(zmm3.i128[0]),
-          reduce_add_epi32(zmm0.i128[1]), reduce_add_epi32(zmm1.i128[1]), reduce_add_epi32(zmm2.i128[1]), reduce_add_epi32(zmm3.i128[1]),
-          reduce_add_epi32(zmm0.i128[2]), reduce_add_epi32(zmm1.i128[2]), reduce_add_epi32(zmm2.i128[2]), reduce_add_epi32(zmm3.i128[2]),
-          reduce_add_epi32(zmm0.i128[3]), reduce_add_epi32(zmm1.i128[3]), reduce_add_epi32(zmm2.i128[3]), reduce_add_epi32(zmm3.i128[3])
-        ]
-    */
-[[maybe_unused]] static __m512i
-m512_hadd128x16_interleave(__m512i sum0, __m512i sum1, __m512i sum2, __m512i sum3) {
-
-    __m512i sum01a = _mm512_unpacklo_epi32(sum0, sum1);
-    __m512i sum01b = _mm512_unpackhi_epi32(sum0, sum1);
-
-    __m512i sum23a = _mm512_unpacklo_epi32(sum2, sum3);
-    __m512i sum23b = _mm512_unpackhi_epi32(sum2, sum3);
-
-    __m512i sum01 = _mm512_add_epi32(sum01a, sum01b);
-    __m512i sum23 = _mm512_add_epi32(sum23a, sum23b);
-
-    __m512i sum0123a = _mm512_unpacklo_epi64(sum01, sum23);
-    __m512i sum0123b = _mm512_unpackhi_epi64(sum01, sum23);
-
-    return _mm512_add_epi32(sum0123a, sum0123b);
-}
-
 [[maybe_unused]] static void m512_add_dpbusd_epi32(__m512i& acc, __m512i a, __m512i b) {
 
     #if defined(USE_VNNI)
diff --git a/src/nnue/network.cpp b/src/nnue/network.cpp
index de2c7eca..a8e901a0 100644
--- a/src/nnue/network.cpp
+++ b/src/nnue/network.cpp
@@ -18,17 +18,17 @@
 
 #include "network.h"
 
-#include <cmath>
 #include <cstdlib>
-#include <cstring>
 #include <fstream>
 #include <iostream>
+#include <memory>
 #include <optional>
 #include <type_traits>
 #include <vector>
 
 #include "../evaluate.h"
 #include "../incbin/incbin.h"
+#include "../memory.h"
 #include "../misc.h"
 #include "../position.h"
 #include "../types.h"
@@ -85,23 +85,6 @@ namespace Stockfish::Eval::NNUE {
 
 namespace Detail {
 
-// Initialize the evaluation function parameters
-template<typename T>
-void initialize(AlignedPtr<T>& pointer) {
-
-    pointer.reset(reinterpret_cast<T*>(std_aligned_alloc(alignof(T), sizeof(T))));
-    std::memset(pointer.get(), 0, sizeof(T));
-}
-
-template<typename T>
-void initialize(LargePagePtr<T>& pointer) {
-
-    static_assert(alignof(T) <= 4096,
-                  "aligned_large_pages_alloc() may fail for such a big alignment requirement of T");
-    pointer.reset(reinterpret_cast<T*>(aligned_large_pages_alloc(sizeof(T))));
-    std::memset(pointer.get(), 0, sizeof(T));
-}
-
 // Read evaluation function parameters
 template<typename T>
 bool read_parameters(std::istream& stream, T& reference) {
@@ -123,6 +106,42 @@ bool write_parameters(std::ostream& stream, const T& reference) {
 
 }  // namespace Detail
 
+template<typename Arch, typename Transformer>
+Network<Arch, Transformer>::Network(const Network<Arch, Transformer>& other) :
+    evalFile(other.evalFile),
+    embeddedType(other.embeddedType) {
+
+    if (other.featureTransformer)
+        featureTransformer = make_unique_large_page<Transformer>(*other.featureTransformer);
+
+    network = make_unique_aligned<Arch[]>(LayerStacks);
+
+    if (!other.network)
+        return;
+
+    for (std::size_t i = 0; i < LayerStacks; ++i)
+        network[i] = other.network[i];
+}
+
+template<typename Arch, typename Transformer>
+Network<Arch, Transformer>&
+Network<Arch, Transformer>::operator=(const Network<Arch, Transformer>& other) {
+    evalFile     = other.evalFile;
+    embeddedType = other.embeddedType;
+
+    if (other.featureTransformer)
+        featureTransformer = make_unique_large_page<Transformer>(*other.featureTransformer);
+
+    network = make_unique_aligned<Arch[]>(LayerStacks);
+
+    if (!other.network)
+        return *this;
+
+    for (std::size_t i = 0; i < LayerStacks; ++i)
+        network[i] = other.network[i];
+
+    return *this;
+}
 
 template<typename Arch, typename Transformer>
 void Network<Arch, Transformer>::load(const std::string& rootDirectory, std::string evalfilePath) {
@@ -186,15 +205,13 @@ bool Network<Arch, Transformer>::save(const std::optional<std::string>& filename
 
 
 template<typename Arch, typename Transformer>
-Value Network<Arch, Transformer>::evaluate(const Position&                         pos,
-                                           AccumulatorCaches::Cache<FTDimensions>* cache,
-                                           bool                                    adjusted,
-                                           int* complexity) const {
+NetworkOutput
+Network<Arch, Transformer>::evaluate(const Position&                         pos,
+                                     AccumulatorCaches::Cache<FTDimensions>* cache) const {
     // We manually align the arrays on the stack because with gcc < 9.3
     // overaligning stack variables with alignas() doesn't work correctly.
 
     constexpr uint64_t alignment = CacheLineSize;
-    constexpr int      delta     = 24;
 
 #if defined(ALIGNAS_ON_STACK_VARIABLES_BROKEN)
     TransformedFeatureType
@@ -211,50 +228,50 @@ Value Network<Arch, Transformer>::evaluate(const Position&
 
     const int  bucket     = (pos.count<ALL_PIECES>() - 1) / 4;
     const auto psqt       = featureTransformer->transform(pos, cache, transformedFeatures, bucket);
-    const auto positional = network[bucket]->propagate(transformedFeatures);
-
-    if (complexity)
-        *complexity = std::abs(psqt - positional) / OutputScale;
-
-    // Give more value to positional evaluation when adjusted flag is set
-    if (adjusted)
-        return static_cast<Value>(((1024 - delta) * psqt + (1024 + delta) * positional)
-                                  / (1024 * OutputScale));
-    else
-        return static_cast<Value>((psqt + positional) / OutputScale);
+    const auto positional = network[bucket].propagate(transformedFeatures);
+    return {static_cast<Value>(psqt / OutputScale), static_cast<Value>(positional / OutputScale)};
 }
 
 
 template<typename Arch, typename Transformer>
-void Network<Arch, Transformer>::verify(std::string evalfilePath) const {
+void Network<Arch, Transformer>::verify(std::string                                  evalfilePath,
+                                        const std::function<void(std::string_view)>& f) const {
     if (evalfilePath.empty())
         evalfilePath = evalFile.defaultName;
 
     if (evalFile.current != evalfilePath)
     {
-        std::string msg1 =
-          "Network evaluation parameters compatible with the engine must be available.";
-        std::string msg2 = "The network file " + evalfilePath + " was not loaded successfully.";
-        std::string msg3 = "The UCI option EvalFile might need to specify the full path, "
-                           "including the directory name, to the network file.";
-        std::string msg4 = "The default net can be downloaded from: "
-                           "https://tests.stockfishchess.org/api/nn/"
-                         + evalFile.defaultName;
-        std::string msg5 = "The engine will be terminated now.";
+        if (f)
+        {
+            std::string msg1 =
+              "Network evaluation parameters compatible with the engine must be available.";
+            std::string msg2 = "The network file " + evalfilePath + " was not loaded successfully.";
+            std::string msg3 = "The UCI option EvalFile might need to specify the full path, "
+                               "including the directory name, to the network file.";
+            std::string msg4 = "The default net can be downloaded from: "
+                               "https://tests.stockfishchess.org/api/nn/"
+                             + evalFile.defaultName;
+            std::string msg5 = "The engine will be terminated now.";
+
+            std::string msg = "ERROR: " + msg1 + '\n' + "ERROR: " + msg2 + '\n' + "ERROR: " + msg3
+                            + '\n' + "ERROR: " + msg4 + '\n' + "ERROR: " + msg5 + '\n';
+
+            f(msg);
+        }
 
-        sync_cout << "info string ERROR: " << msg1 << sync_endl;
-        sync_cout << "info string ERROR: " << msg2 << sync_endl;
-        sync_cout << "info string ERROR: " << msg3 << sync_endl;
-        sync_cout << "info string ERROR: " << msg4 << sync_endl;
-        sync_cout << "info string ERROR: " << msg5 << sync_endl;
         exit(EXIT_FAILURE);
     }
 
-    size_t size = sizeof(*featureTransformer) + sizeof(*network) * LayerStacks;
-    sync_cout << "info string NNUE evaluation using " << evalfilePath << " ("
-              << size / (1024 * 1024) << "MiB, (" << featureTransformer->InputDimensions << ", "
-              << network[0]->TransformedFeatureDimensions << ", " << network[0]->FC_0_OUTPUTS
-              << ", " << network[0]->FC_1_OUTPUTS << ", 1))" << sync_endl;
+    if (f)
+    {
+        size_t size = sizeof(*featureTransformer) + sizeof(Arch) * LayerStacks;
+        f("info string NNUE evaluation using " + evalfilePath + " ("
+          + std::to_string(size / (1024 * 1024)) + "MiB, ("
+          + std::to_string(featureTransformer->InputDimensions) + ", "
+          + std::to_string(network[0].TransformedFeatureDimensions) + ", "
+          + std::to_string(network[0].FC_0_OUTPUTS) + ", " + std::to_string(network[0].FC_1_OUTPUTS)
+          + ", 1))");
+    }
 }
 
 
@@ -291,7 +308,7 @@ Network<Arch, Transformer>::trace_evaluate(const Position&
     {
         const auto materialist =
           featureTransformer->transform(pos, cache, transformedFeatures, bucket);
-        const auto positional = network[bucket]->propagate(transformedFeatures);
+        const auto positional = network[bucket].propagate(transformedFeatures);
 
         t.psqt[bucket]       = static_cast<Value>(materialist / OutputScale);
         t.positional[bucket] = static_cast<Value>(positional / OutputScale);
@@ -344,9 +361,8 @@ void Network<Arch, Transformer>::load_internal() {
 
 template<typename Arch, typename Transformer>
 void Network<Arch, Transformer>::initialize() {
-    Detail::initialize(featureTransformer);
-    for (std::size_t i = 0; i < LayerStacks; ++i)
-        Detail::initialize(network[i]);
+    featureTransformer = make_unique_large_page<Transformer>();
+    network            = make_unique_aligned<Arch[]>(LayerStacks);
 }
 
 
@@ -413,7 +429,7 @@ bool Network<Arch, Transformer>::read_parameters(std::istream& stream,
         return false;
     for (std::size_t i = 0; i < LayerStacks; ++i)
     {
-        if (!Detail::read_parameters(stream, *(network[i])))
+        if (!Detail::read_parameters(stream, network[i]))
             return false;
     }
     return stream && stream.peek() == std::ios::traits_type::eof();
@@ -429,7 +445,7 @@ bool Network<Arch, Transformer>::write_parameters(std::ostream&      stream,
         return false;
     for (std::size_t i = 0; i < LayerStacks; ++i)
     {
-        if (!Detail::write_parameters(stream, *(network[i])))
+        if (!Detail::write_parameters(stream, network[i]))
             return false;
     }
     return bool(stream);
diff --git a/src/nnue/network.h b/src/nnue/network.h
index 23f56663..95253595 100644
--- a/src/nnue/network.h
+++ b/src/nnue/network.h
@@ -20,18 +20,21 @@
 #define NETWORK_H_INCLUDED
 
 #include <cstdint>
+#include <functional>
 #include <iostream>
 #include <optional>
 #include <string>
+#include <string_view>
+#include <tuple>
 #include <utility>
 
-#include "../misc.h"
+#include "../memory.h"
 #include "../position.h"
 #include "../types.h"
+#include "nnue_accumulator.h"
 #include "nnue_architecture.h"
 #include "nnue_feature_transformer.h"
 #include "nnue_misc.h"
-#include "nnue_accumulator.h"
 
 namespace Stockfish::Eval::NNUE {
 
@@ -40,6 +43,7 @@ enum class EmbeddedNNUEType {
     SMALL,
 };
 
+using NetworkOutput = std::tuple<Value, Value>;
 
 template<typename Arch, typename Transformer>
 class Network {
@@ -50,19 +54,23 @@ class Network {
         evalFile(file),
         embeddedType(type) {}
 
+    Network(const Network& other);
+    Network(Network&& other) = default;
+
+    Network& operator=(const Network& other);
+    Network& operator=(Network&& other) = default;
+
     void load(const std::string& rootDirectory, std::string evalfilePath);
     bool save(const std::optional<std::string>& filename) const;
 
-    Value evaluate(const Position&                         pos,
-                   AccumulatorCaches::Cache<FTDimensions>* cache,
-                   bool                                    adjusted   = false,
-                   int*                                    complexity = nullptr) const;
+    NetworkOutput evaluate(const Position&                         pos,
+                           AccumulatorCaches::Cache<FTDimensions>* cache) const;
 
 
     void hint_common_access(const Position&                         pos,
                             AccumulatorCaches::Cache<FTDimensions>* cache) const;
 
-    void          verify(std::string evalfilePath) const;
+    void verify(std::string evalfilePath, const std::function<void(std::string_view)>&) const;
     NnueEvalTrace trace_evaluate(const Position&                         pos,
                                  AccumulatorCaches::Cache<FTDimensions>* cache) const;
 
@@ -85,7 +93,7 @@ class Network {
     LargePagePtr<Transformer> featureTransformer;
 
     // Evaluation function
-    AlignedPtr<Arch> network[LayerStacks];
+    AlignedPtr<Arch[]> network;
 
     EvalFile         evalFile;
     EmbeddedNNUEType embeddedType;
diff --git a/src/nnue/nnue_accumulator.h b/src/nnue/nnue_accumulator.h
index b8dcf1e4..b92901e4 100644
--- a/src/nnue/nnue_accumulator.h
+++ b/src/nnue/nnue_accumulator.h
@@ -80,11 +80,6 @@ struct AccumulatorCaches {
                     entry.clear(network.featureTransformer->biases);
         }
 
-        void clear(const BiasType* biases) {
-            for (auto& entry : entries)
-                entry.clear(biases);
-        }
-
         std::array<Entry, COLOR_NB>& operator[](Square sq) { return entries[sq]; }
 
         std::array<std::array<Entry, COLOR_NB>, SQUARE_NB> entries;
diff --git a/src/nnue/nnue_feature_transformer.h b/src/nnue/nnue_feature_transformer.h
index 2b11adef..fa180678 100644
--- a/src/nnue/nnue_feature_transformer.h
+++ b/src/nnue/nnue_feature_transformer.h
@@ -55,14 +55,14 @@ using psqt_vec_t = __m256i;
     #define vec_store(a, b) _mm512_store_si512(a, b)
     #define vec_add_16(a, b) _mm512_add_epi16(a, b)
     #define vec_sub_16(a, b) _mm512_sub_epi16(a, b)
-    #define vec_mul_16(a, b) _mm512_mullo_epi16(a, b)
+    #define vec_mulhi_16(a, b) _mm512_mulhi_epi16(a, b)
     #define vec_zero() _mm512_setzero_epi32()
     #define vec_set_16(a) _mm512_set1_epi16(a)
     #define vec_max_16(a, b) _mm512_max_epi16(a, b)
     #define vec_min_16(a, b) _mm512_min_epi16(a, b)
+    #define vec_slli_16(a, b) _mm512_slli_epi16(a, b)
     // Inverse permuted at load time
-    #define vec_msb_pack_16(a, b) \
-        _mm512_packs_epi16(_mm512_srli_epi16(a, 7), _mm512_srli_epi16(b, 7))
+    #define vec_packus_16(a, b) _mm512_packus_epi16(a, b)
     #define vec_load_psqt(a) _mm256_load_si256(a)
     #define vec_store_psqt(a, b) _mm256_store_si256(a, b)
     #define vec_add_psqt_32(a, b) _mm256_add_epi32(a, b)
@@ -78,14 +78,14 @@ using psqt_vec_t = __m256i;
     #define vec_store(a, b) _mm256_store_si256(a, b)
     #define vec_add_16(a, b) _mm256_add_epi16(a, b)
     #define vec_sub_16(a, b) _mm256_sub_epi16(a, b)
-    #define vec_mul_16(a, b) _mm256_mullo_epi16(a, b)
+    #define vec_mulhi_16(a, b) _mm256_mulhi_epi16(a, b)
     #define vec_zero() _mm256_setzero_si256()
     #define vec_set_16(a) _mm256_set1_epi16(a)
     #define vec_max_16(a, b) _mm256_max_epi16(a, b)
     #define vec_min_16(a, b) _mm256_min_epi16(a, b)
+    #define vec_slli_16(a, b) _mm256_slli_epi16(a, b)
     // Inverse permuted at load time
-    #define vec_msb_pack_16(a, b) \
-        _mm256_packs_epi16(_mm256_srli_epi16(a, 7), _mm256_srli_epi16(b, 7))
+    #define vec_packus_16(a, b) _mm256_packus_epi16(a, b)
     #define vec_load_psqt(a) _mm256_load_si256(a)
     #define vec_store_psqt(a, b) _mm256_store_si256(a, b)
     #define vec_add_psqt_32(a, b) _mm256_add_epi32(a, b)
@@ -101,12 +101,13 @@ using psqt_vec_t = __m128i;
     #define vec_store(a, b) *(a) = (b)
     #define vec_add_16(a, b) _mm_add_epi16(a, b)
     #define vec_sub_16(a, b) _mm_sub_epi16(a, b)
-    #define vec_mul_16(a, b) _mm_mullo_epi16(a, b)
+    #define vec_mulhi_16(a, b) _mm_mulhi_epi16(a, b)
     #define vec_zero() _mm_setzero_si128()
     #define vec_set_16(a) _mm_set1_epi16(a)
     #define vec_max_16(a, b) _mm_max_epi16(a, b)
     #define vec_min_16(a, b) _mm_min_epi16(a, b)
-    #define vec_msb_pack_16(a, b) _mm_packs_epi16(_mm_srli_epi16(a, 7), _mm_srli_epi16(b, 7))
+    #define vec_slli_16(a, b) _mm_slli_epi16(a, b)
+    #define vec_packus_16(a, b) _mm_packus_epi16(a, b)
     #define vec_load_psqt(a) (*(a))
     #define vec_store_psqt(a, b) *(a) = (b)
     #define vec_add_psqt_32(a, b) _mm_add_epi32(a, b)
@@ -122,18 +123,14 @@ using psqt_vec_t = int32x4_t;
     #define vec_store(a, b) *(a) = (b)
     #define vec_add_16(a, b) vaddq_s16(a, b)
     #define vec_sub_16(a, b) vsubq_s16(a, b)
-    #define vec_mul_16(a, b) vmulq_s16(a, b)
+    #define vec_mulhi_16(a, b) vqdmulhq_s16(a, b)
     #define vec_zero() \
         vec_t { 0 }
     #define vec_set_16(a) vdupq_n_s16(a)
     #define vec_max_16(a, b) vmaxq_s16(a, b)
     #define vec_min_16(a, b) vminq_s16(a, b)
-inline vec_t vec_msb_pack_16(vec_t a, vec_t b) {
-    const int8x8_t  shifta    = vshrn_n_s16(a, 7);
-    const int8x8_t  shiftb    = vshrn_n_s16(b, 7);
-    const int8x16_t compacted = vcombine_s8(shifta, shiftb);
-    return *reinterpret_cast<const vec_t*>(&compacted);
-}
+    #define vec_slli_16(a, b) vshlq_s16(a, vec_set_16(b))
+    #define vec_packus_16(a, b) reinterpret_cast<vec_t>(vcombine_u8(vqmovun_s16(a), vqmovun_s16(b)))
     #define vec_load_psqt(a) (*(a))
     #define vec_store_psqt(a, b) *(a) = (b)
     #define vec_add_psqt_32(a, b) vaddq_s32(a, b)
@@ -281,6 +278,19 @@ class FeatureTransformer {
 #endif
     }
 
+    inline void scale_weights(bool read) const {
+        for (IndexType j = 0; j < InputDimensions; ++j)
+        {
+            WeightType* w = const_cast<WeightType*>(&weights[j * HalfDimensions]);
+            for (IndexType i = 0; i < HalfDimensions; ++i)
+                w[i] = read ? w[i] * 2 : w[i] / 2;
+        }
+
+        BiasType* b = const_cast<BiasType*>(biases);
+        for (IndexType i = 0; i < HalfDimensions; ++i)
+            b[i] = read ? b[i] * 2 : b[i] / 2;
+    }
+
     // Read network parameters
     bool read_parameters(std::istream& stream) {
 
@@ -289,6 +299,7 @@ class FeatureTransformer {
         read_leb_128<PSQTWeightType>(stream, psqtWeights, PSQTBuckets * InputDimensions);
 
         permute_weights(inverse_order_packs);
+        scale_weights(true);
         return !stream.fail();
     }
 
@@ -296,12 +307,14 @@ class FeatureTransformer {
     bool write_parameters(std::ostream& stream) const {
 
         permute_weights(order_packs);
+        scale_weights(false);
 
         write_leb_128<BiasType>(stream, biases, HalfDimensions);
         write_leb_128<WeightType>(stream, weights, HalfDimensions * InputDimensions);
         write_leb_128<PSQTWeightType>(stream, psqtWeights, PSQTBuckets * InputDimensions);
 
         permute_weights(inverse_order_packs);
+        scale_weights(true);
         return !stream.fail();
     }
 
@@ -332,24 +345,86 @@ class FeatureTransformer {
             constexpr IndexType NumOutputChunks = HalfDimensions / 2 / OutputChunkSize;
 
             const vec_t Zero = vec_zero();
-            const vec_t One  = vec_set_16(127);
+            const vec_t One  = vec_set_16(127 * 2);
 
             const vec_t* in0 = reinterpret_cast<const vec_t*>(&(accumulation[perspectives[p]][0]));
             const vec_t* in1 =
               reinterpret_cast<const vec_t*>(&(accumulation[perspectives[p]][HalfDimensions / 2]));
             vec_t* out = reinterpret_cast<vec_t*>(output + offset);
 
+            // Per the NNUE architecture, here we want to multiply pairs of
+            // clipped elements and divide the product by 128. To do this,
+            // we can naively perform min/max operation to clip each of the
+            // four int16 vectors, mullo pairs together, then pack them into
+            // one int8 vector. However, there exists a faster way.
+
+            // The idea here is to use the implicit clipping from packus to
+            // save us two vec_max_16 instructions. This clipping works due
+            // to the fact that any int16 integer below zero will be zeroed
+            // on packus.
+
+            // Consider the case where the second element is negative.
+            // If we do standard clipping, that element will be zero, which
+            // means our pairwise product is zero. If we perform packus and
+            // remove the lower-side clip for the second element, then our
+            // product before packus will be negative, and is zeroed on pack.
+            // The two operation produce equivalent results, but the second
+            // one (using packus) saves one max operation per pair.
+
+            // But here we run into a problem: mullo does not preserve the
+            // sign of the multiplication. We can get around this by doing
+            // mulhi, which keeps the sign. But that requires an additional
+            // tweak.
+
+            // mulhi cuts off the last 16 bits of the resulting product,
+            // which is the same as performing a rightward shift of 16 bits.
+            // We can use this to our advantage. Recall that we want to
+            // divide the final product by 128, which is equivalent to a
+            // 7-bit right shift. Intuitively, if we shift the clipped
+            // value left by 9, and perform mulhi, which shifts the product
+            // right by 16 bits, then we will net a right shift of 7 bits.
+            // However, this won't work as intended. Since we clip the
+            // values to have a maximum value of 127, shifting it by 9 bits
+            // might occupy the signed bit, resulting in some positive
+            // values being interpreted as negative after the shift.
+
+            // There is a way, however, to get around this limitation. When
+            // loading the network, scale accumulator weights and biases by
+            // 2. To get the same pairwise multiplication result as before,
+            // we need to divide the product by 128 * 2 * 2 = 512, which
+            // amounts to a right shift of 9 bits. So now we only have to
+            // shift left by 7 bits, perform mulhi (shifts right by 16 bits)
+            // and net a 9 bit right shift. Since we scaled everything by
+            // two, the values are clipped at 127 * 2 = 254, which occupies
+            // 8 bits. Shifting it by 7 bits left will no longer occupy the
+            // signed bit, so we are safe.
+
+            // Note that on NEON processors, we shift left by 6 instead
+            // because the instruction "vqdmulhq_s16" also doubles the
+            // return value after the multiplication, adding an extra shift
+            // to the left by 1, so we compensate by shifting less before
+            // the multiplication.
+
+            constexpr int shift =
+    #if defined(USE_SSE2)
+              7;
+    #else
+              6;
+    #endif
+
             for (IndexType j = 0; j < NumOutputChunks; ++j)
             {
-                const vec_t sum0a = vec_max_16(vec_min_16(in0[j * 2 + 0], One), Zero);
-                const vec_t sum0b = vec_max_16(vec_min_16(in0[j * 2 + 1], One), Zero);
-                const vec_t sum1a = vec_max_16(vec_min_16(in1[j * 2 + 0], One), Zero);
-                const vec_t sum1b = vec_max_16(vec_min_16(in1[j * 2 + 1], One), Zero);
+                const vec_t sum0a =
+                  vec_slli_16(vec_max_16(vec_min_16(in0[j * 2 + 0], One), Zero), shift);
+                const vec_t sum0b =
+                  vec_slli_16(vec_max_16(vec_min_16(in0[j * 2 + 1], One), Zero), shift);
+                const vec_t sum1a = vec_min_16(in1[j * 2 + 0], One);
+                const vec_t sum1b = vec_min_16(in1[j * 2 + 1], One);
 
-                const vec_t pa = vec_mul_16(sum0a, sum1a);
-                const vec_t pb = vec_mul_16(sum0b, sum1b);
+                const vec_t pa = vec_mulhi_16(sum0a, sum1a);
+                const vec_t pb = vec_mulhi_16(sum0b, sum1b);
 
-                out[j] = vec_msb_pack_16(pa, pb);
+                out[j] = vec_packus_16(pa, pb);
             }
 
 #else
@@ -359,9 +434,9 @@ class FeatureTransformer {
                 BiasType sum0 = accumulation[static_cast<int>(perspectives[p])][j + 0];
                 BiasType sum1 =
                   accumulation[static_cast<int>(perspectives[p])][j + HalfDimensions / 2];
-                sum0               = std::clamp<BiasType>(sum0, 0, 127);
-                sum1               = std::clamp<BiasType>(sum1, 0, 127);
-                output[offset + j] = static_cast<OutputType>(unsigned(sum0 * sum1) / 128);
+                sum0               = std::clamp<BiasType>(sum0, 0, 127 * 2);
+                sum1               = std::clamp<BiasType>(sum1, 0, 127 * 2);
+                output[offset + j] = static_cast<OutputType>(unsigned(sum0 * sum1) / 512);
             }
 
 #endif
@@ -378,11 +453,10 @@ class FeatureTransformer {
 
    private:
     template<Color Perspective>
-    [[nodiscard]] std::pair<StateInfo*, StateInfo*>
-    try_find_computed_accumulator(const Position& pos) const {
+    StateInfo* try_find_computed_accumulator(const Position& pos) const {
         // Look for a usable accumulator of an earlier position. We keep track
         // of the estimated gain in terms of features to be added/subtracted.
-        StateInfo *st = pos.state(), *next = nullptr;
+        StateInfo* st   = pos.state();
         int        gain = FeatureSet::refresh_cost(pos);
         while (st->previous && !(st->*accPtr).computed[Perspective])
         {
@@ -391,236 +465,205 @@ class FeatureTransformer {
             if (FeatureSet::requires_refresh(st, Perspective)
                 || (gain -= FeatureSet::update_cost(st) + 1) < 0)
                 break;
-            next = st;
-            st   = st->previous;
+            st = st->previous;
         }
-        return {st, next};
+        return st;
     }
 
-    // NOTE: The parameter states_to_update is an array of position states.
-    //       All states must be sequential, that is states_to_update[i] must either be reachable
-    //       by repeatedly applying ->previous from states_to_update[i+1].
-    //       computed_st must be reachable by repeatedly applying ->previous on
-    //       states_to_update[0].
-    template<Color Perspective, size_t N>
-    void update_accumulator_incremental(const Position& pos,
-                                        StateInfo*      computed_st,
-                                        StateInfo*      states_to_update[N]) const {
-        static_assert(N > 0);
-        assert([&]() {
-            for (size_t i = 0; i < N; ++i)
-            {
-                if (states_to_update[i] == nullptr)
-                    return false;
-            }
-            return true;
-        }());
+    // It computes the accumulator of the next position, or updates the
+    // current position's accumulator if CurrentOnly is true.
+    template<Color Perspective, bool CurrentOnly>
+    void update_accumulator_incremental(const Position& pos, StateInfo* computed) const {
+        assert((computed->*accPtr).computed[Perspective]);
+        assert(computed->next != nullptr);
 
 #ifdef VECTOR
         // Gcc-10.2 unnecessarily spills AVX2 registers if this array
-        // is defined in the VECTOR code below, once in each branch
+        // is defined in the VECTOR code below, once in each branch.
         vec_t      acc[NumRegs];
         psqt_vec_t psqt[NumPsqtRegs];
 #endif
 
-        // Update incrementally going back through states_to_update.
-        // Gather all features to be updated.
         const Square ksq = pos.square<KING>(Perspective);
 
         // The size must be enough to contain the largest possible update.
         // That might depend on the feature set and generally relies on the
         // feature set's update cost calculation to be correct and never allow
         // updates with more added/removed features than MaxActiveDimensions.
-        FeatureSet::IndexList removed[N], added[N];
+        FeatureSet::IndexList removed, added;
 
-        for (int i = N - 1; i >= 0; --i)
-        {
-            (states_to_update[i]->*accPtr).computed[Perspective] = true;
+        if constexpr (CurrentOnly)
+            for (StateInfo* st = pos.state(); st != computed; st = st->previous)
+                FeatureSet::append_changed_indices<Perspective>(ksq, st->dirtyPiece, removed,
+                                                                added);
+        else
+            FeatureSet::append_changed_indices<Perspective>(ksq, computed->next->dirtyPiece,
+                                                            removed, added);
 
-            const StateInfo* end_state = i == 0 ? computed_st : states_to_update[i - 1];
+        StateInfo* next = CurrentOnly ? pos.state() : computed->next;
+        assert(!(next->*accPtr).computed[Perspective]);
 
-            for (StateInfo* st2 = states_to_update[i]; st2 != end_state; st2 = st2->previous)
-                FeatureSet::append_changed_indices<Perspective>(ksq, st2->dirtyPiece, removed[i],
-                                                                added[i]);
-        }
-
-        StateInfo* st = computed_st;
-
-        // Now update the accumulators listed in states_to_update[], where the last element is a sentinel.
 #ifdef VECTOR
-
-        if (N == 1 && (removed[0].size() == 1 || removed[0].size() == 2) && added[0].size() == 1)
+        if ((removed.size() == 1 || removed.size() == 2) && added.size() == 1)
         {
-            assert(states_to_update[0]);
-
             auto accIn =
-              reinterpret_cast<const vec_t*>(&(st->*accPtr).accumulation[Perspective][0]);
-            auto accOut = reinterpret_cast<vec_t*>(
-              &(states_to_update[0]->*accPtr).accumulation[Perspective][0]);
+              reinterpret_cast<const vec_t*>(&(computed->*accPtr).accumulation[Perspective][0]);
+            auto accOut = reinterpret_cast<vec_t*>(&(next->*accPtr).accumulation[Perspective][0]);
 
-            const IndexType offsetR0 = HalfDimensions * removed[0][0];
+            const IndexType offsetR0 = HalfDimensions * removed[0];
             auto            columnR0 = reinterpret_cast<const vec_t*>(&weights[offsetR0]);
-            const IndexType offsetA  = HalfDimensions * added[0][0];
+            const IndexType offsetA  = HalfDimensions * added[0];
             auto            columnA  = reinterpret_cast<const vec_t*>(&weights[offsetA]);
 
-            if (removed[0].size() == 1)
+            if (removed.size() == 1)
             {
-                for (IndexType k = 0; k < HalfDimensions * sizeof(std::int16_t) / sizeof(vec_t);
-                     ++k)
-                    accOut[k] = vec_add_16(vec_sub_16(accIn[k], columnR0[k]), columnA[k]);
+                for (IndexType i = 0; i < HalfDimensions * sizeof(WeightType) / sizeof(vec_t); ++i)
+                    accOut[i] = vec_add_16(vec_sub_16(accIn[i], columnR0[i]), columnA[i]);
             }
             else
             {
-                const IndexType offsetR1 = HalfDimensions * removed[0][1];
+                const IndexType offsetR1 = HalfDimensions * removed[1];
                 auto            columnR1 = reinterpret_cast<const vec_t*>(&weights[offsetR1]);
 
-                for (IndexType k = 0; k < HalfDimensions * sizeof(std::int16_t) / sizeof(vec_t);
-                     ++k)
-                    accOut[k] = vec_sub_16(vec_add_16(accIn[k], columnA[k]),
-                                           vec_add_16(columnR0[k], columnR1[k]));
+                for (IndexType i = 0; i < HalfDimensions * sizeof(WeightType) / sizeof(vec_t); ++i)
+                    accOut[i] = vec_sub_16(vec_add_16(accIn[i], columnA[i]),
+                                           vec_add_16(columnR0[i], columnR1[i]));
             }
 
-            auto accPsqtIn =
-              reinterpret_cast<const psqt_vec_t*>(&(st->*accPtr).psqtAccumulation[Perspective][0]);
-            auto accPsqtOut = reinterpret_cast<psqt_vec_t*>(
-              &(states_to_update[0]->*accPtr).psqtAccumulation[Perspective][0]);
+            auto accPsqtIn = reinterpret_cast<const psqt_vec_t*>(
+              &(computed->*accPtr).psqtAccumulation[Perspective][0]);
+            auto accPsqtOut =
+              reinterpret_cast<psqt_vec_t*>(&(next->*accPtr).psqtAccumulation[Perspective][0]);
 
-            const IndexType offsetPsqtR0 = PSQTBuckets * removed[0][0];
+            const IndexType offsetPsqtR0 = PSQTBuckets * removed[0];
             auto columnPsqtR0 = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offsetPsqtR0]);
-            const IndexType offsetPsqtA = PSQTBuckets * added[0][0];
+            const IndexType offsetPsqtA = PSQTBuckets * added[0];
             auto columnPsqtA = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offsetPsqtA]);
 
-            if (removed[0].size() == 1)
+            if (removed.size() == 1)
             {
-                for (std::size_t k = 0; k < PSQTBuckets * sizeof(std::int32_t) / sizeof(psqt_vec_t);
-                     ++k)
-                    accPsqtOut[k] = vec_add_psqt_32(vec_sub_psqt_32(accPsqtIn[k], columnPsqtR0[k]),
-                                                    columnPsqtA[k]);
+                for (std::size_t i = 0;
+                     i < PSQTBuckets * sizeof(PSQTWeightType) / sizeof(psqt_vec_t); ++i)
+                    accPsqtOut[i] = vec_add_psqt_32(vec_sub_psqt_32(accPsqtIn[i], columnPsqtR0[i]),
+                                                    columnPsqtA[i]);
             }
             else
             {
-                const IndexType offsetPsqtR1 = PSQTBuckets * removed[0][1];
+                const IndexType offsetPsqtR1 = PSQTBuckets * removed[1];
                 auto columnPsqtR1 = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offsetPsqtR1]);
 
-                for (std::size_t k = 0; k < PSQTBuckets * sizeof(std::int32_t) / sizeof(psqt_vec_t);
-                     ++k)
-                    accPsqtOut[k] =
-                      vec_sub_psqt_32(vec_add_psqt_32(accPsqtIn[k], columnPsqtA[k]),
-                                      vec_add_psqt_32(columnPsqtR0[k], columnPsqtR1[k]));
+                for (std::size_t i = 0;
+                     i < PSQTBuckets * sizeof(PSQTWeightType) / sizeof(psqt_vec_t); ++i)
+                    accPsqtOut[i] =
+                      vec_sub_psqt_32(vec_add_psqt_32(accPsqtIn[i], columnPsqtA[i]),
+                                      vec_add_psqt_32(columnPsqtR0[i], columnPsqtR1[i]));
             }
         }
         else
         {
-            for (IndexType j = 0; j < HalfDimensions / TileHeight; ++j)
+            for (IndexType i = 0; i < HalfDimensions / TileHeight; ++i)
             {
                 // Load accumulator
                 auto accTileIn = reinterpret_cast<const vec_t*>(
-                  &(st->*accPtr).accumulation[Perspective][j * TileHeight]);
-                for (IndexType k = 0; k < NumRegs; ++k)
-                    acc[k] = vec_load(&accTileIn[k]);
+                  &(computed->*accPtr).accumulation[Perspective][i * TileHeight]);
+                for (IndexType j = 0; j < NumRegs; ++j)
+                    acc[j] = vec_load(&accTileIn[j]);
 
-                for (IndexType i = 0; i < N; ++i)
+                // Difference calculation for the deactivated features
+                for (const auto index : removed)
                 {
-                    // Difference calculation for the deactivated features
-                    for (const auto index : removed[i])
-                    {
-                        const IndexType offset = HalfDimensions * index + j * TileHeight;
-                        auto            column = reinterpret_cast<const vec_t*>(&weights[offset]);
-                        for (IndexType k = 0; k < NumRegs; ++k)
-                            acc[k] = vec_sub_16(acc[k], column[k]);
-                    }
-
-                    // Difference calculation for the activated features
-                    for (const auto index : added[i])
-                    {
-                        const IndexType offset = HalfDimensions * index + j * TileHeight;
-                        auto            column = reinterpret_cast<const vec_t*>(&weights[offset]);
-                        for (IndexType k = 0; k < NumRegs; ++k)
-                            acc[k] = vec_add_16(acc[k], column[k]);
-                    }
-
-                    // Store accumulator
-                    auto accTileOut = reinterpret_cast<vec_t*>(
-                      &(states_to_update[i]->*accPtr).accumulation[Perspective][j * TileHeight]);
-                    for (IndexType k = 0; k < NumRegs; ++k)
-                        vec_store(&accTileOut[k], acc[k]);
+                    const IndexType offset = HalfDimensions * index + i * TileHeight;
+                    auto            column = reinterpret_cast<const vec_t*>(&weights[offset]);
+                    for (IndexType j = 0; j < NumRegs; ++j)
+                        acc[j] = vec_sub_16(acc[j], column[j]);
                 }
+
+                // Difference calculation for the activated features
+                for (const auto index : added)
+                {
+                    const IndexType offset = HalfDimensions * index + i * TileHeight;
+                    auto            column = reinterpret_cast<const vec_t*>(&weights[offset]);
+                    for (IndexType j = 0; j < NumRegs; ++j)
+                        acc[j] = vec_add_16(acc[j], column[j]);
+                }
+
+                // Store accumulator
+                auto accTileOut = reinterpret_cast<vec_t*>(
+                  &(next->*accPtr).accumulation[Perspective][i * TileHeight]);
+                for (IndexType j = 0; j < NumRegs; ++j)
+                    vec_store(&accTileOut[j], acc[j]);
             }
 
-            for (IndexType j = 0; j < PSQTBuckets / PsqtTileHeight; ++j)
+            for (IndexType i = 0; i < PSQTBuckets / PsqtTileHeight; ++i)
             {
                 // Load accumulator
                 auto accTilePsqtIn = reinterpret_cast<const psqt_vec_t*>(
-                  &(st->*accPtr).psqtAccumulation[Perspective][j * PsqtTileHeight]);
-                for (std::size_t k = 0; k < NumPsqtRegs; ++k)
-                    psqt[k] = vec_load_psqt(&accTilePsqtIn[k]);
+                  &(computed->*accPtr).psqtAccumulation[Perspective][i * PsqtTileHeight]);
+                for (std::size_t j = 0; j < NumPsqtRegs; ++j)
+                    psqt[j] = vec_load_psqt(&accTilePsqtIn[j]);
 
-                for (IndexType i = 0; i < N; ++i)
+                // Difference calculation for the deactivated features
+                for (const auto index : removed)
                 {
-                    // Difference calculation for the deactivated features
-                    for (const auto index : removed[i])
-                    {
-                        const IndexType offset = PSQTBuckets * index + j * PsqtTileHeight;
-                        auto columnPsqt = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offset]);
-                        for (std::size_t k = 0; k < NumPsqtRegs; ++k)
-                            psqt[k] = vec_sub_psqt_32(psqt[k], columnPsqt[k]);
-                    }
-
-                    // Difference calculation for the activated features
-                    for (const auto index : added[i])
-                    {
-                        const IndexType offset = PSQTBuckets * index + j * PsqtTileHeight;
-                        auto columnPsqt = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offset]);
-                        for (std::size_t k = 0; k < NumPsqtRegs; ++k)
-                            psqt[k] = vec_add_psqt_32(psqt[k], columnPsqt[k]);
-                    }
-
-                    // Store accumulator
-                    auto accTilePsqtOut = reinterpret_cast<psqt_vec_t*>(
-                      &(states_to_update[i]->*accPtr)
-                         .psqtAccumulation[Perspective][j * PsqtTileHeight]);
-                    for (std::size_t k = 0; k < NumPsqtRegs; ++k)
-                        vec_store_psqt(&accTilePsqtOut[k], psqt[k]);
+                    const IndexType offset = PSQTBuckets * index + i * PsqtTileHeight;
+                    auto columnPsqt = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offset]);
+                    for (std::size_t j = 0; j < NumPsqtRegs; ++j)
+                        psqt[j] = vec_sub_psqt_32(psqt[j], columnPsqt[j]);
                 }
+
+                // Difference calculation for the activated features
+                for (const auto index : added)
+                {
+                    const IndexType offset = PSQTBuckets * index + i * PsqtTileHeight;
+                    auto columnPsqt = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offset]);
+                    for (std::size_t j = 0; j < NumPsqtRegs; ++j)
+                        psqt[j] = vec_add_psqt_32(psqt[j], columnPsqt[j]);
+                }
+
+                // Store accumulator
+                auto accTilePsqtOut = reinterpret_cast<psqt_vec_t*>(
+                  &(next->*accPtr).psqtAccumulation[Perspective][i * PsqtTileHeight]);
+                for (std::size_t j = 0; j < NumPsqtRegs; ++j)
+                    vec_store_psqt(&accTilePsqtOut[j], psqt[j]);
             }
         }
 #else
-        for (IndexType i = 0; i < N; ++i)
+        std::memcpy((next->*accPtr).accumulation[Perspective],
+                    (computed->*accPtr).accumulation[Perspective],
+                    HalfDimensions * sizeof(BiasType));
+        std::memcpy((next->*accPtr).psqtAccumulation[Perspective],
+                    (computed->*accPtr).psqtAccumulation[Perspective],
+                    PSQTBuckets * sizeof(PSQTWeightType));
+
+        // Difference calculation for the deactivated features
+        for (const auto index : removed)
         {
-            std::memcpy((states_to_update[i]->*accPtr).accumulation[Perspective],
-                        (st->*accPtr).accumulation[Perspective], HalfDimensions * sizeof(BiasType));
+            const IndexType offset = HalfDimensions * index;
+            for (IndexType i = 0; i < HalfDimensions; ++i)
+                (next->*accPtr).accumulation[Perspective][i] -= weights[offset + i];
 
-            for (std::size_t k = 0; k < PSQTBuckets; ++k)
-                (states_to_update[i]->*accPtr).psqtAccumulation[Perspective][k] =
-                  (st->*accPtr).psqtAccumulation[Perspective][k];
+            for (std::size_t i = 0; i < PSQTBuckets; ++i)
+                (next->*accPtr).psqtAccumulation[Perspective][i] -=
+                  psqtWeights[index * PSQTBuckets + i];
+        }
 
-            st = states_to_update[i];
+        // Difference calculation for the activated features
+        for (const auto index : added)
+        {
+            const IndexType offset = HalfDimensions * index;
+            for (IndexType i = 0; i < HalfDimensions; ++i)
+                (next->*accPtr).accumulation[Perspective][i] += weights[offset + i];
 
-            // Difference calculation for the deactivated features
-            for (const auto index : removed[i])
-            {
-                const IndexType offset = HalfDimensions * index;
-                for (IndexType j = 0; j < HalfDimensions; ++j)
-                    (st->*accPtr).accumulation[Perspective][j] -= weights[offset + j];
-
-                for (std::size_t k = 0; k < PSQTBuckets; ++k)
-                    (st->*accPtr).psqtAccumulation[Perspective][k] -=
-                      psqtWeights[index * PSQTBuckets + k];
-            }
-
-            // Difference calculation for the activated features
-            for (const auto index : added[i])
-            {
-                const IndexType offset = HalfDimensions * index;
-                for (IndexType j = 0; j < HalfDimensions; ++j)
-                    (st->*accPtr).accumulation[Perspective][j] += weights[offset + j];
-
-                for (std::size_t k = 0; k < PSQTBuckets; ++k)
-                    (st->*accPtr).psqtAccumulation[Perspective][k] +=
-                      psqtWeights[index * PSQTBuckets + k];
-            }
+            for (std::size_t i = 0; i < PSQTBuckets; ++i)
+                (next->*accPtr).psqtAccumulation[Perspective][i] +=
+                  psqtWeights[index * PSQTBuckets + i];
         }
 #endif
+
+        (next->*accPtr).computed[Perspective] = true;
+
+        if (!CurrentOnly && next != pos.state())
+            update_accumulator_incremental<Perspective, false>(pos, next);
     }
 
     template<Color Perspective>
@@ -664,7 +707,10 @@ class FeatureTransformer {
 
         for (IndexType j = 0; j < HalfDimensions / TileHeight; ++j)
         {
+            auto accTile =
+              reinterpret_cast<vec_t*>(&accumulator.accumulation[Perspective][j * TileHeight]);
             auto entryTile = reinterpret_cast<vec_t*>(&entry.accumulation[j * TileHeight]);
+
             for (IndexType k = 0; k < NumRegs; ++k)
                 acc[k] = entryTile[k];
 
@@ -679,7 +725,7 @@ class FeatureTransformer {
                 auto            columnA = reinterpret_cast<const vec_t*>(&weights[offsetA]);
 
                 for (unsigned k = 0; k < NumRegs; ++k)
-                    acc[k] = vec_add_16(vec_sub_16(acc[k], columnR[k]), columnA[k]);
+                    acc[k] = vec_add_16(acc[k], vec_sub_16(columnA[k], columnR[k]));
             }
             for (; i < int(removed.size()); ++i)
             {
@@ -702,12 +748,17 @@ class FeatureTransformer {
 
             for (IndexType k = 0; k < NumRegs; k++)
                 vec_store(&entryTile[k], acc[k]);
+            for (IndexType k = 0; k < NumRegs; k++)
+                vec_store(&accTile[k], acc[k]);
         }
 
         for (IndexType j = 0; j < PSQTBuckets / PsqtTileHeight; ++j)
         {
+            auto accTilePsqt = reinterpret_cast<psqt_vec_t*>(
+              &accumulator.psqtAccumulation[Perspective][j * PsqtTileHeight]);
             auto entryTilePsqt =
               reinterpret_cast<psqt_vec_t*>(&entry.psqtAccumulation[j * PsqtTileHeight]);
+
             for (std::size_t k = 0; k < NumPsqtRegs; ++k)
                 psqt[k] = entryTilePsqt[k];
 
@@ -732,6 +783,8 @@ class FeatureTransformer {
 
             for (std::size_t k = 0; k < NumPsqtRegs; ++k)
                 vec_store_psqt(&entryTilePsqt[k], psqt[k]);
+            for (std::size_t k = 0; k < NumPsqtRegs; ++k)
+                vec_store_psqt(&accTilePsqt[k], psqt[k]);
         }
 
 #else
@@ -755,16 +808,15 @@ class FeatureTransformer {
                 entry.psqtAccumulation[k] += psqtWeights[index * PSQTBuckets + k];
         }
 
-#endif
-
         // The accumulator of the refresh entry has been updated.
-        // Now copy its content to the actual accumulator we were refreshing
+        // Now copy its content to the actual accumulator we were refreshing.
 
         std::memcpy(accumulator.accumulation[Perspective], entry.accumulation,
                     sizeof(BiasType) * HalfDimensions);
 
         std::memcpy(accumulator.psqtAccumulation[Perspective], entry.psqtAccumulation,
                     sizeof(int32_t) * PSQTBuckets);
+#endif
 
         for (Color c : {WHITE, BLACK})
             entry.byColorBB[c] = pos.pieces(c);
@@ -786,14 +838,10 @@ class FeatureTransformer {
         if ((pos.state()->*accPtr).computed[Perspective])
             return;
 
-        auto [oldest_st, _] = try_find_computed_accumulator<Perspective>(pos);
+        StateInfo* oldest = try_find_computed_accumulator<Perspective>(pos);
 
-        if ((oldest_st->*accPtr).computed[Perspective])
-        {
-            // Only update current position accumulator to minimize work.
-            StateInfo* states_to_update[1] = {pos.state()};
-            update_accumulator_incremental<Perspective, 1>(pos, oldest_st, states_to_update);
-        }
+        if ((oldest->*accPtr).computed[Perspective] && oldest != pos.state())
+            update_accumulator_incremental<Perspective, true>(pos, oldest);
         else
             update_accumulator_refresh_cache<Perspective>(pos, cache);
     }
@@ -802,31 +850,12 @@ class FeatureTransformer {
     void update_accumulator(const Position&                           pos,
                             AccumulatorCaches::Cache<HalfDimensions>* cache) const {
 
-        auto [oldest_st, next] = try_find_computed_accumulator<Perspective>(pos);
+        StateInfo* oldest = try_find_computed_accumulator<Perspective>(pos);
 
-        if ((oldest_st->*accPtr).computed[Perspective])
-        {
-            if (next == nullptr)
-                return;
-
-            // Now update the accumulators listed in states_to_update[], where the last element is a sentinel.
-            // Currently we update 2 accumulators.
-            //     1. for the current position
-            //     2. the next accumulator after the computed one
-            // The heuristic may change in the future.
-            if (next == pos.state())
-            {
-                StateInfo* states_to_update[1] = {next};
-
-                update_accumulator_incremental<Perspective, 1>(pos, oldest_st, states_to_update);
-            }
-            else
-            {
-                StateInfo* states_to_update[2] = {next, pos.state()};
-
-                update_accumulator_incremental<Perspective, 2>(pos, oldest_st, states_to_update);
-            }
-        }
+        if ((oldest->*accPtr).computed[Perspective] && oldest != pos.state())
+            // Start from the oldest computed accumulator, update all the
+            // accumulators up to the current position.
+            update_accumulator_incremental<Perspective, false>(pos, oldest);
         else
             update_accumulator_refresh_cache<Perspective>(pos, cache);
     }
diff --git a/src/nnue/nnue_misc.cpp b/src/nnue/nnue_misc.cpp
index bf73a58b..a2bece21 100644
--- a/src/nnue/nnue_misc.cpp
+++ b/src/nnue/nnue_misc.cpp
@@ -28,6 +28,7 @@
 #include <iostream>
 #include <sstream>
 #include <string_view>
+#include <tuple>
 
 #include "../evaluate.h"
 #include "../position.h"
@@ -45,9 +46,7 @@ constexpr std::string_view PieceToChar(" PNBRQK  pnbrqk");
 void hint_common_parent_position(const Position&    pos,
                                  const Networks&    networks,
                                  AccumulatorCaches& caches) {
-
-    int simpleEvalAbs = std::abs(simple_eval(pos, pos.side_to_move()));
-    if (simpleEvalAbs > Eval::SmallNetThreshold)
+    if (Eval::use_smallnet(pos))
         networks.small.hint_common_access(pos, &caches.small);
     else
         networks.big.hint_common_access(pos, &caches.big);
@@ -127,14 +126,15 @@ trace(Position& pos, const Eval::NNUE::Networks& networks, Eval::NNUE::Accumulat
         board[y][x] = board[y][x + 8] = board[y + 3][x + 8] = board[y + 3][x] = '+';
         if (pc != NO_PIECE)
             board[y + 1][x + 4] = PieceToChar[pc];
-        if (value != VALUE_NONE)
+        if (is_valid(value))
             format_cp_compact(value, &board[y + 2][x + 2], pos);
     };
 
     // We estimate the value of each piece by doing a differential evaluation from
     // the current base eval, simulating the removal of the piece from its square.
-    Value base = networks.big.evaluate(pos, &caches.big);
-    base       = pos.side_to_move() == WHITE ? base : -base;
+    auto [psqt, positional] = networks.big.evaluate(pos, &caches.big);
+    Value base              = psqt + positional;
+    base                    = pos.side_to_move() == WHITE ? base : -base;
 
     for (File f = FILE_A; f <= FILE_H; ++f)
         for (Rank r = RANK_1; r <= RANK_8; ++r)
@@ -150,9 +150,10 @@ trace(Position& pos, const Eval::NNUE::Networks& networks, Eval::NNUE::Accumulat
                 pos.remove_piece(sq);
                 st->accumulatorBig.computed[WHITE] = st->accumulatorBig.computed[BLACK] = false;
 
-                Value eval = networks.big.evaluate(pos, &caches.big);
-                eval       = pos.side_to_move() == WHITE ? eval : -eval;
-                v          = base - eval;
+                std::tie(psqt, positional) = networks.big.evaluate(pos, &caches.big);
+                Value eval                 = psqt + positional;
+                eval                       = pos.side_to_move() == WHITE ? eval : -eval;
+                v                          = base - eval;
 
                 pos.put_piece(pc, sq);
                 st->accumulatorBig.computed[WHITE] = st->accumulatorBig.computed[BLACK] = false;
@@ -177,16 +178,16 @@ trace(Position& pos, const Eval::NNUE::Networks& networks, Eval::NNUE::Accumulat
 
     for (std::size_t bucket = 0; bucket < LayerStacks; ++bucket)
     {
-        ss << "|  " << bucket << "        ";
-        ss << " |  ";
+        ss << "|  " << bucket << "        "  //
+           << " |  ";
         format_cp_aligned_dot(t.psqt[bucket], ss, pos);
-        ss << "  "
+        ss << "  "  //
            << " |  ";
         format_cp_aligned_dot(t.positional[bucket], ss, pos);
-        ss << "  "
+        ss << "  "  //
            << " |  ";
         format_cp_aligned_dot(t.psqt[bucket] + t.positional[bucket], ss, pos);
-        ss << "  "
+        ss << "  "  //
            << " |";
         if (bucket == t.correctBucket)
             ss << " <-- this bucket is used";
diff --git a/src/numa.h b/src/numa.h
new file mode 100644
index 00000000..1063721e
--- /dev/null
+++ b/src/numa.h
@@ -0,0 +1,1346 @@
+/*
+  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
+  Copyright (C) 2004-2024 The Stockfish developers (see AUTHORS file)
+
+  Stockfish is free software: you can redistribute it and/or modify
+  it under the terms of the GNU General Public License as published by
+  the Free Software Foundation, either version 3 of the License, or
+  (at your option) any later version.
+
+  Stockfish is distributed in the hope that it will be useful,
+  but WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+  GNU General Public License for more details.
+
+  You should have received a copy of the GNU General Public License
+  along with this program.  If not, see <http://www.gnu.org/licenses/>.
+*/
+
+#ifndef NUMA_H_INCLUDED
+#define NUMA_H_INCLUDED
+
+#include <algorithm>
+#include <atomic>
+#include <cstdint>
+#include <cstdlib>
+#include <functional>
+#include <iostream>
+#include <limits>
+#include <map>
+#include <memory>
+#include <mutex>
+#include <set>
+#include <sstream>
+#include <string>
+#include <thread>
+#include <utility>
+#include <vector>
+#include <cstring>
+
+#include "memory.h"
+
+// We support linux very well, but we explicitly do NOT support Android,
+// because there is no affected systems, not worth maintaining.
+#if defined(__linux__) && !defined(__ANDROID__)
+    #if !defined(_GNU_SOURCE)
+        #define _GNU_SOURCE
+    #endif
+    #include <sched.h>
+#elif defined(_WIN64)
+
+    #if _WIN32_WINNT < 0x0601
+        #undef _WIN32_WINNT
+        #define _WIN32_WINNT 0x0601  // Force to include needed API prototypes
+    #endif
+
+// On Windows each processor group can have up to 64 processors.
+// https://learn.microsoft.com/en-us/windows/win32/procthread/processor-groups
+static constexpr size_t WIN_PROCESSOR_GROUP_SIZE = 64;
+
+    #if !defined(NOMINMAX)
+        #define NOMINMAX
+    #endif
+    #include <windows.h>
+    #if defined small
+        #undef small
+    #endif
+
+// https://learn.microsoft.com/en-us/windows/win32/api/processthreadsapi/nf-processthreadsapi-setthreadselectedcpusetmasks
+using SetThreadSelectedCpuSetMasks_t = BOOL (*)(HANDLE, PGROUP_AFFINITY, USHORT);
+
+// https://learn.microsoft.com/en-us/windows/win32/api/processthreadsapi/nf-processthreadsapi-getthreadselectedcpusetmasks
+using GetThreadSelectedCpuSetMasks_t = BOOL (*)(HANDLE, PGROUP_AFFINITY, USHORT, PUSHORT);
+
+#endif
+
+#include "misc.h"
+
+namespace Stockfish {
+
+using CpuIndex  = size_t;
+using NumaIndex = size_t;
+
+inline CpuIndex get_hardware_concurrency() {
+    CpuIndex concurrency = std::thread::hardware_concurrency();
+
+    // Get all processors across all processor groups on windows, since
+    // hardware_concurrency() only returns the number of processors in
+    // the first group, because only these are available to std::thread.
+#ifdef _WIN64
+    concurrency = std::max<CpuIndex>(concurrency, GetActiveProcessorCount(ALL_PROCESSOR_GROUPS));
+#endif
+
+    return concurrency;
+}
+
+inline const CpuIndex SYSTEM_THREADS_NB = std::max<CpuIndex>(1, get_hardware_concurrency());
+
+#if defined(_WIN64)
+
+struct WindowsAffinity {
+    std::optional<std::set<CpuIndex>> oldApi;
+    std::optional<std::set<CpuIndex>> newApi;
+
+    // We also provide diagnostic for when the affinity is set to nullopt
+    // whether it was due to being indeterminate. If affinity is indeterminate
+    // it is best to assume it is not set at all, so consistent with the meaning
+    // of the nullopt affinity.
+    bool isNewDeterminate = true;
+    bool isOldDeterminate = true;
+
+    std::optional<std::set<CpuIndex>> get_combined() const {
+        if (!oldApi.has_value())
+            return newApi;
+        if (!newApi.has_value())
+            return oldApi;
+
+        std::set<CpuIndex> intersect;
+        std::set_intersection(oldApi->begin(), oldApi->end(), newApi->begin(), newApi->end(),
+                              std::inserter(intersect, intersect.begin()));
+        return intersect;
+    }
+
+    // Since Windows 11 and Windows Server 2022 thread affinities can span
+    // processor groups and can be set as such by a new WinAPI function. However,
+    // we may need to force using the old API if we detect that the process has
+    // affinity set by the old API already and we want to override that. Due to the
+    // limitations of the old API we cannot detect its use reliably. There will be
+    // cases where we detect not use but it has actually been used and vice versa.
+
+    bool likely_used_old_api() const { return oldApi.has_value() || !isOldDeterminate; }
+};
+
+inline std::pair<BOOL, std::vector<USHORT>> get_process_group_affinity() {
+
+    // GetProcessGroupAffinity requires the GroupArray argument to be
+    // aligned to 4 bytes instead of just 2.
+    static constexpr size_t GroupArrayMinimumAlignment = 4;
+    static_assert(GroupArrayMinimumAlignment >= alignof(USHORT));
+
+    // The function should succeed the second time, but it may fail if the group
+    // affinity has changed between GetProcessGroupAffinity calls. In such case
+    // we consider this a hard error, as we Cannot work with unstable affinities
+    // anyway.
+    static constexpr int MAX_TRIES  = 2;
+    USHORT               GroupCount = 1;
+    for (int i = 0; i < MAX_TRIES; ++i)
+    {
+        auto GroupArray = std::make_unique<USHORT[]>(
+          GroupCount + (GroupArrayMinimumAlignment / alignof(USHORT) - 1));
+
+        USHORT* GroupArrayAligned = align_ptr_up<GroupArrayMinimumAlignment>(GroupArray.get());
+
+        const BOOL status =
+          GetProcessGroupAffinity(GetCurrentProcess(), &GroupCount, GroupArrayAligned);
+
+        if (status == 0 && GetLastError() != ERROR_INSUFFICIENT_BUFFER)
+        {
+            break;
+        }
+
+        if (status != 0)
+        {
+            return std::make_pair(status,
+                                  std::vector(GroupArrayAligned, GroupArrayAligned + GroupCount));
+        }
+    }
+
+    return std::make_pair(0, std::vector<USHORT>());
+}
+
+// On Windows there are two ways to set affinity, and therefore 2 ways to get it.
+// These are not consistent, so we have to check both. In some cases it is actually
+// not possible to determine affinity. For example when two different threads have
+// affinity on different processor groups, set using SetThreadAffinityMask, we cannot
+// retrieve the actual affinities.
+// From documentation on GetProcessAffinityMask:
+//     > If the calling process contains threads in multiple groups,
+//     > the function returns zero for both affinity masks.
+// In such cases we just give up and assume we have affinity for all processors.
+// nullopt means no affinity is set, that is, all processors are allowed
+inline WindowsAffinity get_process_affinity() {
+    HMODULE k32                            = GetModuleHandle(TEXT("Kernel32.dll"));
+    auto    GetThreadSelectedCpuSetMasks_f = GetThreadSelectedCpuSetMasks_t(
+      (void (*)()) GetProcAddress(k32, "GetThreadSelectedCpuSetMasks"));
+
+    BOOL status = 0;
+
+    WindowsAffinity affinity;
+
+    if (GetThreadSelectedCpuSetMasks_f != nullptr)
+    {
+        USHORT RequiredMaskCount;
+        status = GetThreadSelectedCpuSetMasks_f(GetCurrentThread(), nullptr, 0, &RequiredMaskCount);
+
+        // We expect ERROR_INSUFFICIENT_BUFFER from GetThreadSelectedCpuSetMasks,
+        // but other failure is an actual error.
+        if (status == 0 && GetLastError() != ERROR_INSUFFICIENT_BUFFER)
+        {
+            affinity.isNewDeterminate = false;
+        }
+        else if (RequiredMaskCount > 0)
+        {
+            // If RequiredMaskCount then these affinities were never set, but it's
+            // not consistent so GetProcessAffinityMask may still return some affinity.
+            auto groupAffinities = std::make_unique<GROUP_AFFINITY[]>(RequiredMaskCount);
+
+            status = GetThreadSelectedCpuSetMasks_f(GetCurrentThread(), groupAffinities.get(),
+                                                    RequiredMaskCount, &RequiredMaskCount);
+
+            if (status == 0)
+            {
+                affinity.isNewDeterminate = false;
+            }
+            else
+            {
+                std::set<CpuIndex> cpus;
+
+                for (USHORT i = 0; i < RequiredMaskCount; ++i)
+                {
+                    const size_t procGroupIndex = groupAffinities[i].Group;
+
+                    for (size_t j = 0; j < WIN_PROCESSOR_GROUP_SIZE; ++j)
+                    {
+                        if (groupAffinities[i].Mask & (KAFFINITY(1) << j))
+                            cpus.insert(procGroupIndex * WIN_PROCESSOR_GROUP_SIZE + j);
+                    }
+                }
+
+                affinity.newApi = std::move(cpus);
+            }
+        }
+    }
+
+    // NOTE: There is no way to determine full affinity using the old API if
+    //       individual threads set affinity on different processor groups.
+
+    DWORD_PTR proc, sys;
+    status = GetProcessAffinityMask(GetCurrentProcess(), &proc, &sys);
+
+    // If proc == 0 then we cannot determine affinity because it spans processor groups.
+    // On Windows 11 and Server 2022 it will instead
+    //     > If, however, hHandle specifies a handle to the current process, the function
+    //     > always uses the calling thread's primary group (which by default is the same
+    //     > as the process' primary group) in order to set the
+    //     > lpProcessAffinityMask and lpSystemAffinityMask.
+    // So it will never be indeterminate here. We can only make assumptions later.
+    if (status == 0 || proc == 0)
+    {
+        affinity.isOldDeterminate = false;
+        return affinity;
+    }
+
+    // If SetProcessAffinityMask was never called the affinity must span
+    // all processor groups, but if it was called it must only span one.
+
+    std::vector<USHORT> groupAffinity;  // We need to capture this later and capturing
+                                        // from structured bindings requires c++20.
+
+    std::tie(status, groupAffinity) = get_process_group_affinity();
+    if (status == 0)
+    {
+        affinity.isOldDeterminate = false;
+        return affinity;
+    }
+
+    if (groupAffinity.size() == 1)
+    {
+        // We detect the case when affinity is set to all processors and correctly
+        // leave affinity.oldApi as nullopt.
+        if (GetActiveProcessorGroupCount() != 1 || proc != sys)
+        {
+            std::set<CpuIndex> cpus;
+
+            const size_t procGroupIndex = groupAffinity[0];
+
+            const uint64_t mask = static_cast<uint64_t>(proc);
+            for (size_t j = 0; j < WIN_PROCESSOR_GROUP_SIZE; ++j)
+            {
+                if (mask & (KAFFINITY(1) << j))
+                    cpus.insert(procGroupIndex * WIN_PROCESSOR_GROUP_SIZE + j);
+            }
+
+            affinity.oldApi = std::move(cpus);
+        }
+    }
+    else
+    {
+        // If we got here it means that either SetProcessAffinityMask was never set
+        // or we're on Windows 11/Server 2022.
+
+        // Since Windows 11 and Windows Server 2022 the behaviour of
+        // GetProcessAffinityMask changed:
+        //     > If, however, hHandle specifies a handle to the current process,
+        //     > the function always uses the calling thread's primary group
+        //     > (which by default is the same as the process' primary group)
+        //     > in order to set the lpProcessAffinityMask and lpSystemAffinityMask.
+        // In which case we can actually retrieve the full affinity.
+
+        if (GetThreadSelectedCpuSetMasks_f != nullptr)
+        {
+            std::thread th([&]() {
+                std::set<CpuIndex> cpus;
+                bool               isAffinityFull = true;
+
+                for (auto procGroupIndex : groupAffinity)
+                {
+                    const int numActiveProcessors =
+                      GetActiveProcessorCount(static_cast<WORD>(procGroupIndex));
+
+                    // We have to schedule to two different processors
+                    // and & the affinities we get. Otherwise our processor
+                    // choice could influence the resulting affinity.
+                    // We assume the processor IDs within the group are
+                    // filled sequentially from 0.
+                    uint64_t procCombined = std::numeric_limits<uint64_t>::max();
+                    uint64_t sysCombined  = std::numeric_limits<uint64_t>::max();
+
+                    for (int i = 0; i < std::min(numActiveProcessors, 2); ++i)
+                    {
+                        GROUP_AFFINITY GroupAffinity;
+                        std::memset(&GroupAffinity, 0, sizeof(GROUP_AFFINITY));
+                        GroupAffinity.Group = static_cast<WORD>(procGroupIndex);
+
+                        GroupAffinity.Mask = static_cast<KAFFINITY>(1) << i;
+
+                        status =
+                          SetThreadGroupAffinity(GetCurrentThread(), &GroupAffinity, nullptr);
+                        if (status == 0)
+                        {
+                            affinity.isOldDeterminate = false;
+                            return;
+                        }
+
+                        SwitchToThread();
+
+                        DWORD_PTR proc2, sys2;
+                        status = GetProcessAffinityMask(GetCurrentProcess(), &proc2, &sys2);
+                        if (status == 0)
+                        {
+                            affinity.isOldDeterminate = false;
+                            return;
+                        }
+
+                        procCombined &= static_cast<uint64_t>(proc2);
+                        sysCombined &= static_cast<uint64_t>(sys2);
+                    }
+
+                    if (procCombined != sysCombined)
+                        isAffinityFull = false;
+
+                    for (size_t j = 0; j < WIN_PROCESSOR_GROUP_SIZE; ++j)
+                    {
+                        if (procCombined & (KAFFINITY(1) << j))
+                            cpus.insert(procGroupIndex * WIN_PROCESSOR_GROUP_SIZE + j);
+                    }
+                }
+
+                // We have to detect the case where the affinity was not set,
+                // or is set to all processors so that we correctly produce as
+                // std::nullopt result.
+                if (!isAffinityFull)
+                {
+                    affinity.oldApi = std::move(cpus);
+                }
+            });
+
+            th.join();
+        }
+    }
+
+    return affinity;
+}
+
+#endif
+
+#if defined(__linux__) && !defined(__ANDROID__)
+
+inline std::set<CpuIndex> get_process_affinity() {
+
+    std::set<CpuIndex> cpus;
+
+    // For unsupported systems, or in case of a soft error, we may assume
+    // all processors are available for use.
+    [[maybe_unused]] auto set_to_all_cpus = [&]() {
+        for (CpuIndex c = 0; c < SYSTEM_THREADS_NB; ++c)
+            cpus.insert(c);
+    };
+
+    // cpu_set_t by default holds 1024 entries. This may not be enough soon,
+    // but there is no easy way to determine how many threads there actually
+    // is. In this case we just choose a reasonable upper bound.
+    static constexpr CpuIndex MaxNumCpus = 1024 * 64;
+
+    cpu_set_t* mask = CPU_ALLOC(MaxNumCpus);
+    if (mask == nullptr)
+        std::exit(EXIT_FAILURE);
+
+    const size_t masksize = CPU_ALLOC_SIZE(MaxNumCpus);
+
+    CPU_ZERO_S(masksize, mask);
+
+    const int status = sched_getaffinity(0, masksize, mask);
+
+    if (status != 0)
+    {
+        CPU_FREE(mask);
+        std::exit(EXIT_FAILURE);
+    }
+
+    for (CpuIndex c = 0; c < MaxNumCpus; ++c)
+        if (CPU_ISSET_S(c, masksize, mask))
+            cpus.insert(c);
+
+    CPU_FREE(mask);
+
+    return cpus;
+}
+
+#endif
+
+#if defined(__linux__) && !defined(__ANDROID__)
+
+inline static const auto STARTUP_PROCESSOR_AFFINITY = get_process_affinity();
+
+#elif defined(_WIN64)
+
+inline static const auto STARTUP_PROCESSOR_AFFINITY = get_process_affinity();
+inline static const auto STARTUP_USE_OLD_AFFINITY_API =
+  STARTUP_PROCESSOR_AFFINITY.likely_used_old_api();
+
+#endif
+
+// We want to abstract the purpose of storing the numa node index somewhat.
+// Whoever is using this does not need to know the specifics of the replication
+// machinery to be able to access NUMA replicated memory.
+class NumaReplicatedAccessToken {
+   public:
+    NumaReplicatedAccessToken() :
+        n(0) {}
+
+    explicit NumaReplicatedAccessToken(NumaIndex idx) :
+        n(idx) {}
+
+    NumaIndex get_numa_index() const { return n; }
+
+   private:
+    NumaIndex n;
+};
+
+// Designed as immutable, because there is no good reason to alter an already
+// existing config in a way that doesn't require recreating it completely, and
+// it would be complex and expensive to maintain class invariants.
+// The CPU (processor) numbers always correspond to the actual numbering used
+// by the system. The NUMA node numbers MAY NOT correspond to the system's
+// numbering of the NUMA nodes. In particular, empty nodes may be removed, or
+// the user may create custom nodes. It is guaranteed that NUMA nodes are NOT
+// empty: every node exposed by NumaConfig has at least one processor assigned.
+//
+// We use startup affinities so as not to modify its own behaviour in time.
+//
+// Since Stockfish doesn't support exceptions all places where an exception
+// should be thrown are replaced by std::exit.
+class NumaConfig {
+   public:
+    NumaConfig() :
+        highestCpuIndex(0),
+        customAffinity(false) {
+        const auto numCpus = SYSTEM_THREADS_NB;
+        add_cpu_range_to_node(NumaIndex{0}, CpuIndex{0}, numCpus - 1);
+    }
+
+    // This function queries the system for the mapping of processors to NUMA nodes.
+    // On Linux we read from standardized kernel sysfs, with a fallback to single NUMA
+    // node. On Windows we utilize GetNumaProcessorNodeEx, which has its quirks, see
+    // comment for Windows implementation of get_process_affinity.
+    static NumaConfig from_system([[maybe_unused]] bool respectProcessAffinity = true) {
+        NumaConfig cfg = empty();
+
+#if defined(__linux__) && !defined(__ANDROID__)
+
+        std::set<CpuIndex> allowedCpus;
+
+        if (respectProcessAffinity)
+            allowedCpus = STARTUP_PROCESSOR_AFFINITY;
+
+        auto is_cpu_allowed = [respectProcessAffinity, &allowedCpus](CpuIndex c) {
+            return !respectProcessAffinity || allowedCpus.count(c) == 1;
+        };
+
+        // On Linux things are straightforward, since there's no processor groups and
+        // any thread can be scheduled on all processors.
+        // We try to gather this information from the sysfs first
+        // https://www.kernel.org/doc/Documentation/ABI/stable/sysfs-devices-node
+
+        bool useFallback = false;
+        auto fallback    = [&]() {
+            useFallback = true;
+            cfg         = empty();
+        };
+
+        // /sys/devices/system/node/online contains information about active NUMA nodes
+        auto nodeIdsStr = read_file_to_string("/sys/devices/system/node/online");
+        if (!nodeIdsStr.has_value() || nodeIdsStr->empty())
+        {
+            fallback();
+        }
+        else
+        {
+            remove_whitespace(*nodeIdsStr);
+            for (size_t n : indices_from_shortened_string(*nodeIdsStr))
+            {
+                // /sys/devices/system/node/node.../cpulist
+                std::string path =
+                  std::string("/sys/devices/system/node/node") + std::to_string(n) + "/cpulist";
+                auto cpuIdsStr = read_file_to_string(path);
+                // Now, we only bail if the file does not exist. Some nodes may be
+                // empty, that's fine. An empty node still has a file that appears
+                // to have some whitespace, so we need to handle that.
+                if (!cpuIdsStr.has_value())
+                {
+                    fallback();
+                    break;
+                }
+                else
+                {
+                    remove_whitespace(*cpuIdsStr);
+                    for (size_t c : indices_from_shortened_string(*cpuIdsStr))
+                    {
+                        if (is_cpu_allowed(c))
+                            cfg.add_cpu_to_node(n, c);
+                    }
+                }
+            }
+        }
+
+        if (useFallback)
+        {
+            for (CpuIndex c = 0; c < SYSTEM_THREADS_NB; ++c)
+                if (is_cpu_allowed(c))
+                    cfg.add_cpu_to_node(NumaIndex{0}, c);
+        }
+
+#elif defined(_WIN64)
+
+        std::optional<std::set<CpuIndex>> allowedCpus;
+
+        if (respectProcessAffinity)
+            allowedCpus = STARTUP_PROCESSOR_AFFINITY.get_combined();
+
+        // The affinity cannot be determined in all cases on Windows,
+        // but we at least guarantee that the number of allowed processors
+        // is >= number of processors in the affinity mask. In case the user
+        // is not satisfied they must set the processor numbers explicitly.
+        auto is_cpu_allowed = [&allowedCpus](CpuIndex c) {
+            return !allowedCpus.has_value() || allowedCpus->count(c) == 1;
+        };
+
+        WORD numProcGroups = GetActiveProcessorGroupCount();
+        for (WORD procGroup = 0; procGroup < numProcGroups; ++procGroup)
+        {
+            for (BYTE number = 0; number < WIN_PROCESSOR_GROUP_SIZE; ++number)
+            {
+                PROCESSOR_NUMBER procnum;
+                procnum.Group    = procGroup;
+                procnum.Number   = number;
+                procnum.Reserved = 0;
+                USHORT nodeNumber;
+
+                const BOOL     status = GetNumaProcessorNodeEx(&procnum, &nodeNumber);
+                const CpuIndex c      = static_cast<CpuIndex>(procGroup) * WIN_PROCESSOR_GROUP_SIZE
+                                 + static_cast<CpuIndex>(number);
+                if (status != 0 && nodeNumber != std::numeric_limits<USHORT>::max()
+                    && is_cpu_allowed(c))
+                {
+                    cfg.add_cpu_to_node(nodeNumber, c);
+                }
+            }
+        }
+
+        // Split the NUMA nodes to be contained within a group if necessary.
+        // This is needed between Windows 10 Build 20348 and Windows 11, because
+        // the new NUMA allocation behaviour was introduced while there was
+        // still no way to set thread affinity spanning multiple processor groups.
+        // See https://learn.microsoft.com/en-us/windows/win32/procthread/numa-support
+        // We also do this is if need to force old API for some reason.
+        //
+        // 2024-08-26: It appears that we need to actually always force this behaviour.
+        // While Windows allows this to work now, such assignments have bad interaction
+        // with the scheduler - in particular it still prefers scheduling on the thread's
+        // "primary" node, even if it means scheduling SMT processors first.
+        // See https://github.com/official-stockfish/Stockfish/issues/5551
+        // See https://learn.microsoft.com/en-us/windows/win32/procthread/processor-groups
+        //
+        //     Each process is assigned a primary group at creation, and by default all
+        //     of its threads' primary group is the same. Each thread's ideal processor
+        //     is in the thread's primary group, so threads will preferentially be
+        //     scheduled to processors on their primary group, but they are able to
+        //     be scheduled to processors on any other group.
+        //
+        // used to be guarded by if (STARTUP_USE_OLD_AFFINITY_API)
+        {
+            NumaConfig splitCfg = empty();
+
+            NumaIndex splitNodeIndex = 0;
+            for (const auto& cpus : cfg.nodes)
+            {
+                if (cpus.empty())
+                    continue;
+
+                size_t lastProcGroupIndex = *(cpus.begin()) / WIN_PROCESSOR_GROUP_SIZE;
+                for (CpuIndex c : cpus)
+                {
+                    const size_t procGroupIndex = c / WIN_PROCESSOR_GROUP_SIZE;
+                    if (procGroupIndex != lastProcGroupIndex)
+                    {
+                        splitNodeIndex += 1;
+                        lastProcGroupIndex = procGroupIndex;
+                    }
+                    splitCfg.add_cpu_to_node(splitNodeIndex, c);
+                }
+                splitNodeIndex += 1;
+            }
+
+            cfg = std::move(splitCfg);
+        }
+
+#else
+
+        // Fallback for unsupported systems.
+        for (CpuIndex c = 0; c < SYSTEM_THREADS_NB; ++c)
+            cfg.add_cpu_to_node(NumaIndex{0}, c);
+
+#endif
+
+        // We have to ensure no empty NUMA nodes persist.
+        cfg.remove_empty_numa_nodes();
+
+        // If the user explicitly opts out from respecting the current process affinity
+        // then it may be inconsistent with the current affinity (obviously), so we
+        // consider it custom.
+        if (!respectProcessAffinity)
+            cfg.customAffinity = true;
+
+        return cfg;
+    }
+
+    // ':'-separated numa nodes
+    // ','-separated cpu indices
+    // supports "first-last" range syntax for cpu indices
+    // For example "0-15,128-143:16-31,144-159:32-47,160-175:48-63,176-191"
+    static NumaConfig from_string(const std::string& s) {
+        NumaConfig cfg = empty();
+
+        NumaIndex n = 0;
+        for (auto&& nodeStr : split(s, ":"))
+        {
+            auto indices = indices_from_shortened_string(std::string(nodeStr));
+            if (!indices.empty())
+            {
+                for (auto idx : indices)
+                {
+                    if (!cfg.add_cpu_to_node(n, CpuIndex(idx)))
+                        std::exit(EXIT_FAILURE);
+                }
+
+                n += 1;
+            }
+        }
+
+        cfg.customAffinity = true;
+
+        return cfg;
+    }
+
+    NumaConfig(const NumaConfig&)            = delete;
+    NumaConfig(NumaConfig&&)                 = default;
+    NumaConfig& operator=(const NumaConfig&) = delete;
+    NumaConfig& operator=(NumaConfig&&)      = default;
+
+    bool is_cpu_assigned(CpuIndex n) const { return nodeByCpu.count(n) == 1; }
+
+    NumaIndex num_numa_nodes() const { return nodes.size(); }
+
+    CpuIndex num_cpus_in_numa_node(NumaIndex n) const {
+        assert(n < nodes.size());
+        return nodes[n].size();
+    }
+
+    CpuIndex num_cpus() const { return nodeByCpu.size(); }
+
+    bool requires_memory_replication() const { return customAffinity || nodes.size() > 1; }
+
+    std::string to_string() const {
+        std::string str;
+
+        bool isFirstNode = true;
+        for (auto&& cpus : nodes)
+        {
+            if (!isFirstNode)
+                str += ":";
+
+            bool isFirstSet = true;
+            auto rangeStart = cpus.begin();
+            for (auto it = cpus.begin(); it != cpus.end(); ++it)
+            {
+                auto next = std::next(it);
+                if (next == cpus.end() || *next != *it + 1)
+                {
+                    // cpus[i] is at the end of the range (may be of size 1)
+                    if (!isFirstSet)
+                        str += ",";
+
+                    const CpuIndex last = *it;
+
+                    if (it != rangeStart)
+                    {
+                        const CpuIndex first = *rangeStart;
+
+                        str += std::to_string(first);
+                        str += "-";
+                        str += std::to_string(last);
+                    }
+                    else
+                        str += std::to_string(last);
+
+                    rangeStart = next;
+                    isFirstSet = false;
+                }
+            }
+
+            isFirstNode = false;
+        }
+
+        return str;
+    }
+
+    bool suggests_binding_threads(CpuIndex numThreads) const {
+        // If we can reasonably determine that the threads cannot be contained
+        // by the OS within the first NUMA node then we advise distributing
+        // and binding threads. When the threads are not bound we can only use
+        // NUMA memory replicated objects from the first node, so when the OS
+        // has to schedule on other nodes we lose performance. We also suggest
+        // binding if there's enough threads to distribute among nodes with minimal
+        // disparity. We try to ignore small nodes, in particular the empty ones.
+
+        // If the affinity set by the user does not match the affinity given by
+        // the OS then binding is necessary to ensure the threads are running on
+        // correct processors.
+        if (customAffinity)
+            return true;
+
+        // We obviously cannot distribute a single thread, so a single thread
+        // should never be bound.
+        if (numThreads <= 1)
+            return false;
+
+        size_t largestNodeSize = 0;
+        for (auto&& cpus : nodes)
+            if (cpus.size() > largestNodeSize)
+                largestNodeSize = cpus.size();
+
+        auto is_node_small = [largestNodeSize](const std::set<CpuIndex>& node) {
+            static constexpr double SmallNodeThreshold = 0.6;
+            return static_cast<double>(node.size()) / static_cast<double>(largestNodeSize)
+                <= SmallNodeThreshold;
+        };
+
+        size_t numNotSmallNodes = 0;
+        for (auto&& cpus : nodes)
+            if (!is_node_small(cpus))
+                numNotSmallNodes += 1;
+
+        return (numThreads > largestNodeSize / 2 || numThreads >= numNotSmallNodes * 4)
+            && nodes.size() > 1;
+    }
+
+    std::vector<NumaIndex> distribute_threads_among_numa_nodes(CpuIndex numThreads) const {
+        std::vector<NumaIndex> ns;
+
+        if (nodes.size() == 1)
+        {
+            // Special case for when there's no NUMA nodes. This doesn't buy us
+            // much, but let's keep the default path simple.
+            ns.resize(numThreads, NumaIndex{0});
+        }
+        else
+        {
+            std::vector<size_t> occupation(nodes.size(), 0);
+            for (CpuIndex c = 0; c < numThreads; ++c)
+            {
+                NumaIndex bestNode{0};
+                float     bestNodeFill = std::numeric_limits<float>::max();
+                for (NumaIndex n = 0; n < nodes.size(); ++n)
+                {
+                    float fill =
+                      static_cast<float>(occupation[n] + 1) / static_cast<float>(nodes[n].size());
+                    // NOTE: Do we want to perhaps fill the first available node
+                    //       up to 50% first before considering other nodes?
+                    //       Probably not, because it would interfere with running
+                    //       multiple instances. We basically shouldn't favor any
+                    //       particular node.
+                    if (fill < bestNodeFill)
+                    {
+                        bestNode     = n;
+                        bestNodeFill = fill;
+                    }
+                }
+                ns.emplace_back(bestNode);
+                occupation[bestNode] += 1;
+            }
+        }
+
+        return ns;
+    }
+
+    NumaReplicatedAccessToken bind_current_thread_to_numa_node(NumaIndex n) const {
+        if (n >= nodes.size() || nodes[n].size() == 0)
+            std::exit(EXIT_FAILURE);
+
+#if defined(__linux__) && !defined(__ANDROID__)
+
+        cpu_set_t* mask = CPU_ALLOC(highestCpuIndex + 1);
+        if (mask == nullptr)
+            std::exit(EXIT_FAILURE);
+
+        const size_t masksize = CPU_ALLOC_SIZE(highestCpuIndex + 1);
+
+        CPU_ZERO_S(masksize, mask);
+
+        for (CpuIndex c : nodes[n])
+            CPU_SET_S(c, masksize, mask);
+
+        const int status = sched_setaffinity(0, masksize, mask);
+
+        CPU_FREE(mask);
+
+        if (status != 0)
+            std::exit(EXIT_FAILURE);
+
+        // We yield this thread just to be sure it gets rescheduled.
+        // This is defensive, allowed because this code is not performance critical.
+        sched_yield();
+
+#elif defined(_WIN64)
+
+        // Requires Windows 11. No good way to set thread affinity spanning
+        // processor groups before that.
+        HMODULE k32                            = GetModuleHandle(TEXT("Kernel32.dll"));
+        auto    SetThreadSelectedCpuSetMasks_f = SetThreadSelectedCpuSetMasks_t(
+          (void (*)()) GetProcAddress(k32, "SetThreadSelectedCpuSetMasks"));
+
+        // We ALWAYS set affinity with the new API if available, because
+        // there's no downsides, and we forcibly keep it consistent with
+        // the old API should we need to use it. I.e. we always keep this
+        // as a superset of what we set with SetThreadGroupAffinity.
+        if (SetThreadSelectedCpuSetMasks_f != nullptr)
+        {
+            // Only available on Windows 11 and Windows Server 2022 onwards
+            const USHORT numProcGroups = USHORT(
+              ((highestCpuIndex + 1) + WIN_PROCESSOR_GROUP_SIZE - 1) / WIN_PROCESSOR_GROUP_SIZE);
+            auto groupAffinities = std::make_unique<GROUP_AFFINITY[]>(numProcGroups);
+            std::memset(groupAffinities.get(), 0, sizeof(GROUP_AFFINITY) * numProcGroups);
+            for (WORD i = 0; i < numProcGroups; ++i)
+                groupAffinities[i].Group = i;
+
+            for (CpuIndex c : nodes[n])
+            {
+                const size_t procGroupIndex     = c / WIN_PROCESSOR_GROUP_SIZE;
+                const size_t idxWithinProcGroup = c % WIN_PROCESSOR_GROUP_SIZE;
+                groupAffinities[procGroupIndex].Mask |= KAFFINITY(1) << idxWithinProcGroup;
+            }
+
+            HANDLE hThread = GetCurrentThread();
+
+            const BOOL status =
+              SetThreadSelectedCpuSetMasks_f(hThread, groupAffinities.get(), numProcGroups);
+            if (status == 0)
+                std::exit(EXIT_FAILURE);
+
+            // We yield this thread just to be sure it gets rescheduled.
+            // This is defensive, allowed because this code is not performance critical.
+            SwitchToThread();
+        }
+
+        // Sometimes we need to force the old API, but do not use it unless necessary.
+        if (SetThreadSelectedCpuSetMasks_f == nullptr || STARTUP_USE_OLD_AFFINITY_API)
+        {
+            // On earlier windows version (since windows 7) we cannot run a single thread
+            // on multiple processor groups, so we need to restrict the group.
+            // We assume the group of the first processor listed for this node.
+            // Processors from outside this group will not be assigned for this thread.
+            // Normally this won't be an issue because windows used to assign NUMA nodes
+            // such that they cannot span processor groups. However, since Windows 10
+            // Build 20348 the behaviour changed, so there's a small window of versions
+            // between this and Windows 11 that might exhibit problems with not all
+            // processors being utilized.
+            //
+            // We handle this in NumaConfig::from_system by manually splitting the
+            // nodes when we detect that there is no function to set affinity spanning
+            // processor nodes. This is required because otherwise our thread distribution
+            // code may produce suboptimal results.
+            //
+            // See https://learn.microsoft.com/en-us/windows/win32/procthread/numa-support
+            GROUP_AFFINITY affinity;
+            std::memset(&affinity, 0, sizeof(GROUP_AFFINITY));
+            // We use an ordered set to be sure to get the smallest cpu number here.
+            const size_t forcedProcGroupIndex = *(nodes[n].begin()) / WIN_PROCESSOR_GROUP_SIZE;
+            affinity.Group                    = static_cast<WORD>(forcedProcGroupIndex);
+            for (CpuIndex c : nodes[n])
+            {
+                const size_t procGroupIndex     = c / WIN_PROCESSOR_GROUP_SIZE;
+                const size_t idxWithinProcGroup = c % WIN_PROCESSOR_GROUP_SIZE;
+                // We skip processors that are not in the same processor group.
+                // If everything was set up correctly this will never be an issue,
+                // but we have to account for bad NUMA node specification.
+                if (procGroupIndex != forcedProcGroupIndex)
+                    continue;
+
+                affinity.Mask |= KAFFINITY(1) << idxWithinProcGroup;
+            }
+
+            HANDLE hThread = GetCurrentThread();
+
+            const BOOL status = SetThreadGroupAffinity(hThread, &affinity, nullptr);
+            if (status == 0)
+                std::exit(EXIT_FAILURE);
+
+            // We yield this thread just to be sure it gets rescheduled. This is
+            // defensive, allowed because this code is not performance critical.
+            SwitchToThread();
+        }
+
+#endif
+
+        return NumaReplicatedAccessToken(n);
+    }
+
+    template<typename FuncT>
+    void execute_on_numa_node(NumaIndex n, FuncT&& f) const {
+        std::thread th([this, &f, n]() {
+            bind_current_thread_to_numa_node(n);
+            std::forward<FuncT>(f)();
+        });
+
+        th.join();
+    }
+
+   private:
+    std::vector<std::set<CpuIndex>> nodes;
+    std::map<CpuIndex, NumaIndex>   nodeByCpu;
+    CpuIndex                        highestCpuIndex;
+
+    bool customAffinity;
+
+    static NumaConfig empty() { return NumaConfig(EmptyNodeTag{}); }
+
+    struct EmptyNodeTag {};
+
+    NumaConfig(EmptyNodeTag) :
+        highestCpuIndex(0),
+        customAffinity(false) {}
+
+    void remove_empty_numa_nodes() {
+        std::vector<std::set<CpuIndex>> newNodes;
+        for (auto&& cpus : nodes)
+            if (!cpus.empty())
+                newNodes.emplace_back(std::move(cpus));
+        nodes = std::move(newNodes);
+    }
+
+    // Returns true if successful
+    // Returns false if failed, i.e. when the cpu is already present
+    //                          strong guarantee, the structure remains unmodified
+    bool add_cpu_to_node(NumaIndex n, CpuIndex c) {
+        if (is_cpu_assigned(c))
+            return false;
+
+        while (nodes.size() <= n)
+            nodes.emplace_back();
+
+        nodes[n].insert(c);
+        nodeByCpu[c] = n;
+
+        if (c > highestCpuIndex)
+            highestCpuIndex = c;
+
+        return true;
+    }
+
+    // Returns true if successful
+    // Returns false if failed, i.e. when any of the cpus is already present
+    //                          strong guarantee, the structure remains unmodified
+    bool add_cpu_range_to_node(NumaIndex n, CpuIndex cfirst, CpuIndex clast) {
+        for (CpuIndex c = cfirst; c <= clast; ++c)
+            if (is_cpu_assigned(c))
+                return false;
+
+        while (nodes.size() <= n)
+            nodes.emplace_back();
+
+        for (CpuIndex c = cfirst; c <= clast; ++c)
+        {
+            nodes[n].insert(c);
+            nodeByCpu[c] = n;
+        }
+
+        if (clast > highestCpuIndex)
+            highestCpuIndex = clast;
+
+        return true;
+    }
+
+    static std::vector<size_t> indices_from_shortened_string(const std::string& s) {
+        std::vector<size_t> indices;
+
+        if (s.empty())
+            return indices;
+
+        for (const auto& ss : split(s, ","))
+        {
+            if (ss.empty())
+                continue;
+
+            auto parts = split(ss, "-");
+            if (parts.size() == 1)
+            {
+                const CpuIndex c = CpuIndex{str_to_size_t(std::string(parts[0]))};
+                indices.emplace_back(c);
+            }
+            else if (parts.size() == 2)
+            {
+                const CpuIndex cfirst = CpuIndex{str_to_size_t(std::string(parts[0]))};
+                const CpuIndex clast  = CpuIndex{str_to_size_t(std::string(parts[1]))};
+                for (size_t c = cfirst; c <= clast; ++c)
+                {
+                    indices.emplace_back(c);
+                }
+            }
+        }
+
+        return indices;
+    }
+};
+
+class NumaReplicationContext;
+
+// Instances of this class are tracked by the NumaReplicationContext instance.
+// NumaReplicationContext informs all tracked instances when NUMA configuration changes.
+class NumaReplicatedBase {
+   public:
+    NumaReplicatedBase(NumaReplicationContext& ctx);
+
+    NumaReplicatedBase(const NumaReplicatedBase&) = delete;
+    NumaReplicatedBase(NumaReplicatedBase&& other) noexcept;
+
+    NumaReplicatedBase& operator=(const NumaReplicatedBase&) = delete;
+    NumaReplicatedBase& operator=(NumaReplicatedBase&& other) noexcept;
+
+    virtual void on_numa_config_changed() = 0;
+    virtual ~NumaReplicatedBase();
+
+    const NumaConfig& get_numa_config() const;
+
+   private:
+    NumaReplicationContext* context;
+};
+
+// We force boxing with a unique_ptr. If this becomes an issue due to added
+// indirection we may need to add an option for a custom boxing type. When the
+// NUMA config changes the value stored at the index 0 is replicated to other nodes.
+template<typename T>
+class NumaReplicated: public NumaReplicatedBase {
+   public:
+    using ReplicatorFuncType = std::function<T(const T&)>;
+
+    NumaReplicated(NumaReplicationContext& ctx) :
+        NumaReplicatedBase(ctx) {
+        replicate_from(T{});
+    }
+
+    NumaReplicated(NumaReplicationContext& ctx, T&& source) :
+        NumaReplicatedBase(ctx) {
+        replicate_from(std::move(source));
+    }
+
+    NumaReplicated(const NumaReplicated&) = delete;
+    NumaReplicated(NumaReplicated&& other) noexcept :
+        NumaReplicatedBase(std::move(other)),
+        instances(std::exchange(other.instances, {})) {}
+
+    NumaReplicated& operator=(const NumaReplicated&) = delete;
+    NumaReplicated& operator=(NumaReplicated&& other) noexcept {
+        NumaReplicatedBase::operator=(*this, std::move(other));
+        instances = std::exchange(other.instances, {});
+
+        return *this;
+    }
+
+    NumaReplicated& operator=(T&& source) {
+        replicate_from(std::move(source));
+
+        return *this;
+    }
+
+    ~NumaReplicated() override = default;
+
+    const T& operator[](NumaReplicatedAccessToken token) const {
+        assert(token.get_numa_index() < instances.size());
+        return *(instances[token.get_numa_index()]);
+    }
+
+    const T& operator*() const { return *(instances[0]); }
+
+    const T* operator->() const { return instances[0].get(); }
+
+    template<typename FuncT>
+    void modify_and_replicate(FuncT&& f) {
+        auto source = std::move(instances[0]);
+        std::forward<FuncT>(f)(*source);
+        replicate_from(std::move(*source));
+    }
+
+    void on_numa_config_changed() override {
+        // Use the first one as the source. It doesn't matter which one we use,
+        // because they all must be identical, but the first one is guaranteed to exist.
+        auto source = std::move(instances[0]);
+        replicate_from(std::move(*source));
+    }
+
+   private:
+    std::vector<std::unique_ptr<T>> instances;
+
+    void replicate_from(T&& source) {
+        instances.clear();
+
+        const NumaConfig& cfg = get_numa_config();
+        if (cfg.requires_memory_replication())
+        {
+            for (NumaIndex n = 0; n < cfg.num_numa_nodes(); ++n)
+            {
+                cfg.execute_on_numa_node(
+                  n, [this, &source]() { instances.emplace_back(std::make_unique<T>(source)); });
+            }
+        }
+        else
+        {
+            assert(cfg.num_numa_nodes() == 1);
+            // We take advantage of the fact that replication is not required
+            // and reuse the source value, avoiding one copy operation.
+            instances.emplace_back(std::make_unique<T>(std::move(source)));
+        }
+    }
+};
+
+// We force boxing with a unique_ptr. If this becomes an issue due to added
+// indirection we may need to add an option for a custom boxing type.
+template<typename T>
+class LazyNumaReplicated: public NumaReplicatedBase {
+   public:
+    using ReplicatorFuncType = std::function<T(const T&)>;
+
+    LazyNumaReplicated(NumaReplicationContext& ctx) :
+        NumaReplicatedBase(ctx) {
+        prepare_replicate_from(T{});
+    }
+
+    LazyNumaReplicated(NumaReplicationContext& ctx, T&& source) :
+        NumaReplicatedBase(ctx) {
+        prepare_replicate_from(std::move(source));
+    }
+
+    LazyNumaReplicated(const LazyNumaReplicated&) = delete;
+    LazyNumaReplicated(LazyNumaReplicated&& other) noexcept :
+        NumaReplicatedBase(std::move(other)),
+        instances(std::exchange(other.instances, {})) {}
+
+    LazyNumaReplicated& operator=(const LazyNumaReplicated&) = delete;
+    LazyNumaReplicated& operator=(LazyNumaReplicated&& other) noexcept {
+        NumaReplicatedBase::operator=(*this, std::move(other));
+        instances = std::exchange(other.instances, {});
+
+        return *this;
+    }
+
+    LazyNumaReplicated& operator=(T&& source) {
+        prepare_replicate_from(std::move(source));
+
+        return *this;
+    }
+
+    ~LazyNumaReplicated() override = default;
+
+    const T& operator[](NumaReplicatedAccessToken token) const {
+        assert(token.get_numa_index() < instances.size());
+        ensure_present(token.get_numa_index());
+        return *(instances[token.get_numa_index()]);
+    }
+
+    const T& operator*() const { return *(instances[0]); }
+
+    const T* operator->() const { return instances[0].get(); }
+
+    template<typename FuncT>
+    void modify_and_replicate(FuncT&& f) {
+        auto source = std::move(instances[0]);
+        std::forward<FuncT>(f)(*source);
+        prepare_replicate_from(std::move(*source));
+    }
+
+    void on_numa_config_changed() override {
+        // Use the first one as the source. It doesn't matter which one we use,
+        // because they all must be identical, but the first one is guaranteed to exist.
+        auto source = std::move(instances[0]);
+        prepare_replicate_from(std::move(*source));
+    }
+
+   private:
+    mutable std::vector<std::unique_ptr<T>> instances;
+    mutable std::mutex                      mutex;
+
+    void ensure_present(NumaIndex idx) const {
+        assert(idx < instances.size());
+
+        if (instances[idx] != nullptr)
+            return;
+
+        assert(idx != 0);
+
+        std::unique_lock<std::mutex> lock(mutex);
+        // Check again for races.
+        if (instances[idx] != nullptr)
+            return;
+
+        const NumaConfig& cfg = get_numa_config();
+        cfg.execute_on_numa_node(
+          idx, [this, idx]() { instances[idx] = std::make_unique<T>(*instances[0]); });
+    }
+
+    void prepare_replicate_from(T&& source) {
+        instances.clear();
+
+        const NumaConfig& cfg = get_numa_config();
+        if (cfg.requires_memory_replication())
+        {
+            assert(cfg.num_numa_nodes() > 0);
+
+            // We just need to make sure the first instance is there.
+            // Note that we cannot move here as we need to reallocate the data
+            // on the correct NUMA node.
+            cfg.execute_on_numa_node(
+              0, [this, &source]() { instances.emplace_back(std::make_unique<T>(source)); });
+
+            // Prepare others for lazy init.
+            instances.resize(cfg.num_numa_nodes());
+        }
+        else
+        {
+            assert(cfg.num_numa_nodes() == 1);
+            // We take advantage of the fact that replication is not required
+            // and reuse the source value, avoiding one copy operation.
+            instances.emplace_back(std::make_unique<T>(std::move(source)));
+        }
+    }
+};
+
+class NumaReplicationContext {
+   public:
+    NumaReplicationContext(NumaConfig&& cfg) :
+        config(std::move(cfg)) {}
+
+    NumaReplicationContext(const NumaReplicationContext&) = delete;
+    NumaReplicationContext(NumaReplicationContext&&)      = delete;
+
+    NumaReplicationContext& operator=(const NumaReplicationContext&) = delete;
+    NumaReplicationContext& operator=(NumaReplicationContext&&)      = delete;
+
+    ~NumaReplicationContext() {
+        // The context must outlive replicated objects
+        if (!trackedReplicatedObjects.empty())
+            std::exit(EXIT_FAILURE);
+    }
+
+    void attach(NumaReplicatedBase* obj) {
+        assert(trackedReplicatedObjects.count(obj) == 0);
+        trackedReplicatedObjects.insert(obj);
+    }
+
+    void detach(NumaReplicatedBase* obj) {
+        assert(trackedReplicatedObjects.count(obj) == 1);
+        trackedReplicatedObjects.erase(obj);
+    }
+
+    // oldObj may be invalid at this point
+    void move_attached([[maybe_unused]] NumaReplicatedBase* oldObj, NumaReplicatedBase* newObj) {
+        assert(trackedReplicatedObjects.count(oldObj) == 1);
+        assert(trackedReplicatedObjects.count(newObj) == 0);
+        trackedReplicatedObjects.erase(oldObj);
+        trackedReplicatedObjects.insert(newObj);
+    }
+
+    void set_numa_config(NumaConfig&& cfg) {
+        config = std::move(cfg);
+        for (auto&& obj : trackedReplicatedObjects)
+            obj->on_numa_config_changed();
+    }
+
+    const NumaConfig& get_numa_config() const { return config; }
+
+   private:
+    NumaConfig config;
+
+    // std::set uses std::less by default, which is required for pointer comparison
+    std::set<NumaReplicatedBase*> trackedReplicatedObjects;
+};
+
+inline NumaReplicatedBase::NumaReplicatedBase(NumaReplicationContext& ctx) :
+    context(&ctx) {
+    context->attach(this);
+}
+
+inline NumaReplicatedBase::NumaReplicatedBase(NumaReplicatedBase&& other) noexcept :
+    context(std::exchange(other.context, nullptr)) {
+    context->move_attached(&other, this);
+}
+
+inline NumaReplicatedBase& NumaReplicatedBase::operator=(NumaReplicatedBase&& other) noexcept {
+    context = std::exchange(other.context, nullptr);
+
+    context->move_attached(&other, this);
+
+    return *this;
+}
+
+inline NumaReplicatedBase::~NumaReplicatedBase() {
+    if (context != nullptr)
+        context->detach(this);
+}
+
+inline const NumaConfig& NumaReplicatedBase::get_numa_config() const {
+    return context->get_numa_config();
+}
+
+}  // namespace Stockfish
+
+
+#endif  // #ifndef NUMA_H_INCLUDED
diff --git a/src/position.cpp b/src/position.cpp
index b46ba029..1b1c0269 100644
--- a/src/position.cpp
+++ b/src/position.cpp
@@ -334,8 +334,10 @@ void Position::set_check_info() const {
 // The function is only used when a new position is set up
 void Position::set_state() const {
 
-    st->key = st->materialKey  = 0;
-    st->pawnKey                = Zobrist::noPawns;
+    st->key = st->materialKey = 0;
+    st->majorPieceKey = st->minorPieceKey = 0;
+    st->nonPawnKey[WHITE] = st->nonPawnKey[BLACK] = 0;
+    st->pawnKey                                   = Zobrist::noPawns;
     st->nonPawnMaterial[WHITE] = st->nonPawnMaterial[BLACK] = VALUE_ZERO;
     st->checkersBB = attackers_to(square<KING>(sideToMove)) & pieces(~sideToMove);
 
@@ -350,8 +352,27 @@ void Position::set_state() const {
         if (type_of(pc) == PAWN)
             st->pawnKey ^= Zobrist::psq[pc][s];
 
-        else if (type_of(pc) != KING)
-            st->nonPawnMaterial[color_of(pc)] += PieceValue[pc];
+        else
+        {
+            st->nonPawnKey[color_of(pc)] ^= Zobrist::psq[pc][s];
+
+            if (type_of(pc) != KING)
+            {
+                st->nonPawnMaterial[color_of(pc)] += PieceValue[pc];
+
+                if (type_of(pc) >= ROOK)
+                    st->majorPieceKey ^= Zobrist::psq[pc][s];
+
+                else
+                    st->minorPieceKey ^= Zobrist::psq[pc][s];
+            }
+
+            else
+            {
+                st->majorPieceKey ^= Zobrist::psq[pc][s];
+                st->minorPieceKey ^= Zobrist::psq[pc][s];
+            }
+        }
     }
 
     if (st->epSquare != SQ_NONE)
@@ -671,6 +692,7 @@ void Position::do_move(Move m, StateInfo& newSt, bool givesCheck) {
     // our state pointer to point to the new (ready to be updated) state.
     std::memcpy(&newSt, st, offsetof(StateInfo, key));
     newSt.previous = st;
+    st->next       = &newSt;
     st             = &newSt;
 
     // Increment ply counters. In particular, rule50 will be reset to zero later on
@@ -706,6 +728,8 @@ void Position::do_move(Move m, StateInfo& newSt, bool givesCheck) {
         do_castling<true>(us, from, to, rfrom, rto);
 
         k ^= Zobrist::psq[captured][rfrom] ^ Zobrist::psq[captured][rto];
+        st->majorPieceKey ^= Zobrist::psq[captured][rfrom] ^ Zobrist::psq[captured][rto];
+        st->nonPawnKey[us] ^= Zobrist::psq[captured][rfrom] ^ Zobrist::psq[captured][rto];
         captured = NO_PIECE;
     }
 
@@ -731,7 +755,16 @@ void Position::do_move(Move m, StateInfo& newSt, bool givesCheck) {
             st->pawnKey ^= Zobrist::psq[captured][capsq];
         }
         else
+        {
             st->nonPawnMaterial[them] -= PieceValue[captured];
+            st->nonPawnKey[them] ^= Zobrist::psq[captured][capsq];
+
+            if (type_of(captured) >= ROOK)
+                st->majorPieceKey ^= Zobrist::psq[captured][capsq];
+
+            else
+                st->minorPieceKey ^= Zobrist::psq[captured][capsq];
+        }
 
         dp.dirty_num = 2;  // 1 piece moved, 1 piece captured
         dp.piece[1]  = captured;
@@ -789,7 +822,8 @@ void Position::do_move(Move m, StateInfo& newSt, bool givesCheck) {
 
         else if (m.type_of() == PROMOTION)
         {
-            Piece promotion = make_piece(us, m.promotion_type());
+            Piece     promotion     = make_piece(us, m.promotion_type());
+            PieceType promotionType = type_of(promotion);
 
             assert(relative_rank(us, to) == RANK_8);
             assert(type_of(promotion) >= KNIGHT && type_of(promotion) <= QUEEN);
@@ -810,6 +844,12 @@ void Position::do_move(Move m, StateInfo& newSt, bool givesCheck) {
             st->materialKey ^=
               Zobrist::psq[promotion][pieceCount[promotion] - 1] ^ Zobrist::psq[pc][pieceCount[pc]];
 
+            if (promotionType >= ROOK)
+                st->majorPieceKey ^= Zobrist::psq[promotion][to];
+
+            else
+                st->minorPieceKey ^= Zobrist::psq[promotion][to];
+
             // Update material
             st->nonPawnMaterial[us] += PieceValue[promotion];
         }
@@ -821,6 +861,23 @@ void Position::do_move(Move m, StateInfo& newSt, bool givesCheck) {
         st->rule50 = 0;
     }
 
+    else
+    {
+        st->nonPawnKey[us] ^= Zobrist::psq[pc][from] ^ Zobrist::psq[pc][to];
+
+        if (type_of(pc) == KING)
+        {
+            st->majorPieceKey ^= Zobrist::psq[pc][from] ^ Zobrist::psq[pc][to];
+            st->minorPieceKey ^= Zobrist::psq[pc][from] ^ Zobrist::psq[pc][to];
+        }
+
+        else if (type_of(pc) >= ROOK)
+            st->majorPieceKey ^= Zobrist::psq[pc][from] ^ Zobrist::psq[pc][to];
+
+        else
+            st->minorPieceKey ^= Zobrist::psq[pc][from] ^ Zobrist::psq[pc][to];
+    }
+
     // Set capture piece
     st->capturedPiece = captured;
 
@@ -963,6 +1020,7 @@ void Position::do_null_move(StateInfo& newSt, TranspositionTable& tt) {
     std::memcpy(&newSt, st, offsetof(StateInfo, accumulatorBig));
 
     newSt.previous = st;
+    st->next       = &newSt;
     st             = &newSt;
 
     st->dirtyPiece.dirty_num               = 0;
@@ -1156,9 +1214,9 @@ bool Position::has_repeated() const {
 }
 
 
-// Tests if the position has a move which draws by repetition,
-// or an earlier position has a move that directly reaches the current position.
-bool Position::has_game_cycle(int ply) const {
+// Tests if the position has a move which draws by repetition.
+// This function accurately matches the outcome of is_draw() over all legal moves.
+bool Position::upcoming_repetition(int ply) const {
 
     int j;
 
@@ -1169,10 +1227,16 @@ bool Position::has_game_cycle(int ply) const {
 
     Key        originalKey = st->key;
     StateInfo* stp         = st->previous;
+    Key        other       = originalKey ^ stp->key ^ Zobrist::side;
 
     for (int i = 3; i <= end; i += 2)
     {
-        stp = stp->previous->previous;
+        stp = stp->previous;
+        other ^= stp->key ^ stp->previous->key ^ Zobrist::side;
+        stp = stp->previous;
+
+        if (other != 0)
+            continue;
 
         Key moveKey = originalKey ^ stp->key;
         if ((j = H1(moveKey), cuckoo[j] == moveKey) || (j = H2(moveKey), cuckoo[j] == moveKey))
@@ -1188,12 +1252,6 @@ bool Position::has_game_cycle(int ply) const {
 
                 // For nodes before or at the root, check that the move is a
                 // repetition rather than a move to the current position.
-                // In the cuckoo table, both moves Rc1c5 and Rc5c1 are stored in
-                // the same location, so we have to select which square to check.
-                if (color_of(piece_on(empty(s1) ? s2 : s1)) != side_to_move())
-                    continue;
-
-                // For repetitions before or at the root, require one more
                 if (stp->repetition)
                     return true;
             }
diff --git a/src/position.h b/src/position.h
index 154ed652..888612da 100644
--- a/src/position.h
+++ b/src/position.h
@@ -43,6 +43,9 @@ struct StateInfo {
     // Copied when making a move
     Key    materialKey;
     Key    pawnKey;
+    Key    majorPieceKey;
+    Key    minorPieceKey;
+    Key    nonPawnKey[COLOR_NB];
     Value  nonPawnMaterial[COLOR_NB];
     int    castlingRights;
     int    rule50;
@@ -53,6 +56,7 @@ struct StateInfo {
     Key        key;
     Bitboard   checkersBB;
     StateInfo* previous;
+    StateInfo* next;
     Bitboard   blockersForKing[COLOR_NB];
     Bitboard   pinners[COLOR_NB];
     Bitboard   checkSquares[PIECE_TYPE_NB];
@@ -150,13 +154,16 @@ class Position {
     Key key_after(Move m) const;
     Key material_key() const;
     Key pawn_key() const;
+    Key major_piece_key() const;
+    Key minor_piece_key() const;
+    Key non_pawn_key(Color c) const;
 
     // Other properties of the position
     Color side_to_move() const;
     int   game_ply() const;
     bool  is_chess960() const;
     bool  is_draw(int ply) const;
-    bool  has_game_cycle(int ply) const;
+    bool  upcoming_repetition(int ply) const;
     bool  has_repeated() const;
     int   rule50_count() const;
     Value non_pawn_material(Color c) const;
@@ -297,6 +304,12 @@ inline Key Position::pawn_key() const { return st->pawnKey; }
 
 inline Key Position::material_key() const { return st->materialKey; }
 
+inline Key Position::major_piece_key() const { return st->majorPieceKey; }
+
+inline Key Position::minor_piece_key() const { return st->minorPieceKey; }
+
+inline Key Position::non_pawn_key(Color c) const { return st->nonPawnKey[c]; }
+
 inline Value Position::non_pawn_material(Color c) const { return st->nonPawnMaterial[c]; }
 
 inline Value Position::non_pawn_material() const {
@@ -315,8 +328,8 @@ inline bool Position::capture(Move m) const {
 }
 
 // Returns true if a move is generated from the capture stage, having also
-// queen promotions covered, i.e. consistency with the capture stage move generation
-// is needed to avoid the generation of duplicate moves.
+// queen promotions covered, i.e. consistency with the capture stage move
+// generation is needed to avoid the generation of duplicate moves.
 inline bool Position::capture_stage(Move m) const {
     assert(m.is_ok());
     return capture(m) || m.promotion_type() == QUEEN;
diff --git a/src/score.cpp b/src/score.cpp
index 292f5340..179796d2 100644
--- a/src/score.cpp
+++ b/src/score.cpp
@@ -29,7 +29,7 @@ namespace Stockfish {
 Score::Score(Value v, const Position& pos) {
     assert(-VALUE_INFINITE < v && v < VALUE_INFINITE);
 
-    if (std::abs(v) < VALUE_TB_WIN_IN_MAX_PLY)
+    if (!is_decisive(v))
     {
         score = InternalUnits{UCIEngine::to_cp(v, pos)};
     }
diff --git a/src/search.cpp b/src/search.cpp
index 6830e4b1..e352c96e 100644
--- a/src/search.cpp
+++ b/src/search.cpp
@@ -22,14 +22,19 @@
 #include <array>
 #include <atomic>
 #include <cassert>
+#include <chrono>
 #include <cmath>
 #include <cstdint>
 #include <cstdlib>
 #include <initializer_list>
+#include <iostream>
+#include <list>
+#include <ratio>
 #include <string>
 #include <utility>
 
 #include "evaluate.h"
+#include "history.h"
 #include "misc.h"
 #include "movegen.h"
 #include "movepick.h"
@@ -49,105 +54,98 @@ namespace Stockfish {
 
 namespace TB = Tablebases;
 
+void syzygy_extend_pv(const OptionsMap&            options,
+                      const Search::LimitsType&    limits,
+                      Stockfish::Position&         pos,
+                      Stockfish::Search::RootMove& rootMove,
+                      Value&                       v);
+
 using Eval::evaluate;
 using namespace Search;
 
 namespace {
 
-static constexpr double EvalLevel[10] = {0.981, 0.956, 0.895, 0.949, 0.913,
-                                         0.942, 0.933, 0.890, 0.984, 0.941};
-
 // Futility margin
 Value futility_margin(Depth d, bool noTtCutNode, bool improving, bool oppWorsening) {
-    Value futilityMult       = 126 - 46 * noTtCutNode;
-    Value improvingDeduction = 58 * improving * futilityMult / 32;
-    Value worseningDeduction = (323 + 52 * improving) * oppWorsening * futilityMult / 1024;
+    Value futilityMult       = 109 - 27 * noTtCutNode;
+    Value improvingDeduction = improving * futilityMult * 2;
+    Value worseningDeduction = oppWorsening * futilityMult / 3;
 
     return futilityMult * d - improvingDeduction - worseningDeduction;
 }
 
 constexpr int futility_move_count(bool improving, Depth depth) {
-    return improving ? (3 + depth * depth) : (3 + depth * depth) / 2;
+    return (3 + depth * depth) / (2 - improving);
 }
 
-// Add correctionHistory value to raw staticEval and guarantee evaluation does not hit the tablebase range
-Value to_corrected_static_eval(Value v, const Worker& w, const Position& pos) {
-    auto cv = w.correctionHistory[pos.side_to_move()][pawn_structure_index<Correction>(pos)];
-    v += cv * std::abs(cv) / 7350;
+// Add correctionHistory value to raw staticEval and guarantee evaluation
+// does not hit the tablebase range.
+Value to_corrected_static_eval(Value v, const Worker& w, const Position& pos, Stack* ss) {
+    const Color us    = pos.side_to_move();
+    const auto  m     = (ss - 1)->currentMove;
+    const auto  pcv   = w.pawnCorrectionHistory[us][pawn_structure_index<Correction>(pos)];
+    const auto  macv  = w.majorPieceCorrectionHistory[us][major_piece_index(pos)];
+    const auto  micv  = w.minorPieceCorrectionHistory[us][minor_piece_index(pos)];
+    const auto  wnpcv = w.nonPawnCorrectionHistory[WHITE][us][non_pawn_index<WHITE>(pos)];
+    const auto  bnpcv = w.nonPawnCorrectionHistory[BLACK][us][non_pawn_index<BLACK>(pos)];
+    int         cntcv = 1;
+
+    if (m.is_ok())
+        cntcv = int((*(ss - 2)->continuationCorrectionHistory)[pos.piece_on(m.to_sq())][m.to_sq()]);
+
+    const auto cv =
+      (6384 * pcv + 3583 * macv + 6492 * micv + 6725 * (wnpcv + bnpcv) + cntcv * 5880) / 131072;
+    v += cv;
     return std::clamp(v, VALUE_TB_LOSS_IN_MAX_PLY + 1, VALUE_TB_WIN_IN_MAX_PLY - 1);
 }
 
 // History and stats update bonus, based on depth
-int stat_bonus(Depth d) { return std::clamp(208 * d - 297, 16, 1406); }
+int stat_bonus(Depth d) { return std::min(168 * d - 100, 1718); }
 
 // History and stats update malus, based on depth
-int stat_malus(Depth d) { return (d < 4 ? 520 * d - 312 : 1479); }
+int stat_malus(Depth d) { return std::min(768 * d - 257, 2351); }
 
 // Add a small random component to draw evaluations to avoid 3-fold blindness
 Value value_draw(size_t nodes) { return VALUE_DRAW - 1 + Value(nodes & 0x2); }
-
-// Skill structure is used to implement strength limit. If we have a UCI_Elo,
-// we convert it to an appropriate skill level, anchored to the Stash engine.
-// This method is based on a fit of the Elo results for games played between
-// Stockfish at various skill levels and various versions of the Stash engine.
-// Skill 0 .. 19 now covers CCRL Blitz Elo from 1320 to 3190, approximately
-// Reference: https://github.com/vondele/Stockfish/commit/a08b8d4e9711c2
-struct Skill {
-    Skill(int skill_level, int uci_elo) {
-        if (uci_elo)
-        {
-            double e = double(uci_elo - 1320) / (3190 - 1320);
-            level = std::clamp((((37.2473 * e - 40.8525) * e + 22.2943) * e - 0.311438), 0.0, 19.0);
-        }
-        else
-            level = double(skill_level);
-    }
-    bool enabled() const { return level < 20.0; }
-    bool time_to_pick(Depth depth) const { return depth == 1 + int(level); }
-    Move pick_best(const RootMoves&, size_t multiPV);
-
-    double level;
-    Move   best = Move::none();
-};
-
 Value value_to_tt(Value v, int ply);
 Value value_from_tt(Value v, int ply, int r50c);
 void  update_pv(Move* pv, Move move, const Move* childPv);
 void  update_continuation_histories(Stack* ss, Piece pc, Square to, int bonus);
-void  update_refutations(const Position& pos, Stack* ss, Search::Worker& workerThread, Move move);
 void  update_quiet_histories(
    const Position& pos, Stack* ss, Search::Worker& workerThread, Move move, int bonus);
-void update_quiet_stats(
-  const Position& pos, Stack* ss, Search::Worker& workerThread, Move move, int bonus);
-void update_all_stats(const Position& pos,
-                      Stack*          ss,
-                      Search::Worker& workerThread,
-                      Move            bestMove,
-                      Value           bestValue,
-                      Value           beta,
-                      Square          prevSq,
-                      Move*           quietsSearched,
-                      int             quietCount,
-                      Move*           capturesSearched,
-                      int             captureCount,
-                      Depth           depth);
+void update_all_stats(const Position&      pos,
+                      Stack*               ss,
+                      Search::Worker&      workerThread,
+                      Move                 bestMove,
+                      Square               prevSq,
+                      ValueList<Move, 32>& quietsSearched,
+                      ValueList<Move, 32>& capturesSearched,
+                      Depth                depth);
 
 }  // namespace
 
 Search::Worker::Worker(SharedState&                    sharedState,
                        std::unique_ptr<ISearchManager> sm,
-                       size_t                          thread_id) :
+                       size_t                          threadId,
+                       NumaReplicatedAccessToken       token) :
     // Unpack the SharedState struct into member variables
-    thread_idx(thread_id),
+    threadIdx(threadId),
+    numaAccessToken(token),
     manager(std::move(sm)),
     options(sharedState.options),
     threads(sharedState.threads),
     tt(sharedState.tt),
     networks(sharedState.networks),
-    refreshTable(networks) {
+    refreshTable(networks[token]) {
     clear();
 }
 
+void Search::Worker::ensure_network_replicated() {
+    // Access once to force lazy initialization.
+    // We do this because we want to avoid initialization during search.
+    (void) (networks[numaAccessToken]);
+}
+
 void Search::Worker::start_searching() {
 
     // Non-main threads go directly to iterative_deepening()
@@ -157,7 +155,8 @@ void Search::Worker::start_searching() {
         return;
     }
 
-    main_manager()->tm.init(limits, rootPos.side_to_move(), rootPos.game_ply(), options);
+    main_manager()->tm.init(limits, rootPos.side_to_move(), rootPos.game_ply(), options,
+                            main_manager()->originalTimeAdjust);
     tt.new_search();
 
     if (rootMoves.empty())
@@ -181,7 +180,7 @@ void Search::Worker::start_searching() {
     {}  // Busy wait for a stop or a ponder reset
 
     // Stop the threads if not already stopped (also raise the stop if
-    // "ponderhit" just reset threads.ponder).
+    // "ponderhit" just reset threads.ponder)
     threads.stop = true;
 
     // Wait until all threads have finished
@@ -190,8 +189,8 @@ void Search::Worker::start_searching() {
     // When playing in 'nodes as time' mode, subtract the searched nodes from
     // the available ones before exiting.
     if (limits.npmsec)
-        main_manager()->tm.advance_nodes_time(limits.inc[rootPos.side_to_move()]
-                                              - threads.nodes_searched());
+        main_manager()->tm.advance_nodes_time(threads.nodes_searched()
+                                              - limits.inc[rootPos.side_to_move()]);
 
     Worker* bestThread = this;
     Skill   skill =
@@ -239,7 +238,7 @@ void Search::Worker::iterative_deepening() {
 
     // Allocate stack with extra size to allow access from (ss - 7) to (ss + 2):
     // (ss - 7) is needed for update_continuation_histories(ss - 1) which accesses (ss - 6),
-    // (ss + 2) is needed for initialization of cutOffCnt and killers.
+    // (ss + 2) is needed for initialization of cutOffCnt.
     Stack  stack[MAX_PLY + 10] = {};
     Stack* ss                  = stack + 7;
 
@@ -247,7 +246,8 @@ void Search::Worker::iterative_deepening() {
     {
         (ss - i)->continuationHistory =
           &this->continuationHistory[0][0][NO_PIECE][0];  // Use as a sentinel
-        (ss - i)->staticEval = VALUE_NONE;
+        (ss - i)->continuationCorrectionHistory = &this->continuationCorrectionHistory[NO_PIECE][0];
+        (ss - i)->staticEval                    = VALUE_NONE;
     }
 
     for (int i = 0; i <= MAX_PLY + 2; ++i)
@@ -275,6 +275,8 @@ void Search::Worker::iterative_deepening() {
 
     int searchAgainCounter = 0;
 
+    lowPlyHistory.fill(0);
+
     // Iterative deepening loop until requested to stop or the target depth is reached
     while (++rootDepth < MAX_PLY && !threads.stop
            && !(limits.depth && mainThread && rootDepth > limits.depth))
@@ -295,7 +297,7 @@ void Search::Worker::iterative_deepening() {
             searchAgainCounter++;
 
         // MultiPV loop. We perform a full root search for each PV line
-        for (pvIdx = 0; pvIdx < multiPV && !threads.stop; ++pvIdx)
+        for (pvIdx = 0; pvIdx < multiPV; ++pvIdx)
         {
             if (pvIdx == pvLast)
             {
@@ -309,13 +311,13 @@ void Search::Worker::iterative_deepening() {
             selDepth = 0;
 
             // Reset aspiration window starting size
+            delta     = 5 + std::abs(rootMoves[pvIdx].meanSquaredScore) / 13461;
             Value avg = rootMoves[pvIdx].averageScore;
-            delta     = 10 + avg * avg / 9530;
             alpha     = std::max(avg - delta, -VALUE_INFINITE);
             beta      = std::min(avg + delta, VALUE_INFINITE);
 
             // Adjust optimism based on root move's averageScore (~4 Elo)
-            optimism[us]  = 119 * avg / (std::abs(avg) + 88);
+            optimism[us]  = 150 * avg / (std::abs(avg) + 85);
             optimism[~us] = -optimism[us];
 
             // Start with a small aspiration window and, in the case of a fail
@@ -324,10 +326,11 @@ void Search::Worker::iterative_deepening() {
             int failedHighCnt = 0;
             while (true)
             {
-                // Adjust the effective depth searched, but ensure at least one effective increment
-                // for every four searchAgain steps (see issue #2717).
+                // Adjust the effective depth searched, but ensure at least one
+                // effective increment for every four searchAgain steps (see issue #2717).
                 Depth adjustedDepth =
                   std::max(1, rootDepth - failedHighCnt - 3 * (searchAgainCounter + 1) / 4);
+                rootDelta = beta - alpha;
                 bestValue = search<Root>(rootPos, ss, alpha, beta, adjustedDepth, false);
 
                 // Bring the best move to the front. It is critical that sorting
@@ -344,14 +347,15 @@ void Search::Worker::iterative_deepening() {
                 if (threads.stop)
                     break;
 
-                // When failing high/low give some update (without cluttering
-                // the UI) before a re-search.
+                // When failing high/low give some update before a re-search. To avoid
+                // excessive output that could hang GUIs like Fritz 19, only start
+                // at nodes > 10M (rather than depth N, which can be reached quickly)
                 if (mainThread && multiPV == 1 && (bestValue <= alpha || bestValue >= beta)
-                    && elapsed() > 3000)
+                    && nodes > 10000000)
                     main_manager()->pv(*this, threads, tt, rootDepth);
 
-                // In case of failing low/high increase aspiration window and
-                // re-search, otherwise exit the loop.
+                // In case of failing low/high increase aspiration window and re-search,
+                // otherwise exit the loop.
                 if (bestValue <= alpha)
                 {
                     beta  = (alpha + beta) / 2;
@@ -378,13 +382,17 @@ void Search::Worker::iterative_deepening() {
             std::stable_sort(rootMoves.begin() + pvFirst, rootMoves.begin() + pvIdx + 1);
 
             if (mainThread
-                && (threads.stop || pvIdx + 1 == multiPV || elapsed() > 3000)
-                // A thread that aborted search can have mated-in/TB-loss PV and score
-                // that cannot be trusted, i.e. it can be delayed or refuted if we would have
-                // had time to fully search other root-moves. Thus we suppress this output and
-                // below pick a proven score/PV for this thread (from the previous iteration).
-                && !(threads.abortedSearch && rootMoves[0].uciScore <= VALUE_TB_LOSS_IN_MAX_PLY))
+                && (threads.stop || pvIdx + 1 == multiPV || nodes > 10000000)
+                // A thread that aborted search can have mated-in/TB-loss PV and
+                // score that cannot be trusted, i.e. it can be delayed or refuted
+                // if we would have had time to fully search other root-moves. Thus
+                // we suppress this output and below pick a proven score/PV for this
+                // thread (from the previous iteration).
+                && !(threads.abortedSearch && is_loss(rootMoves[0].uciScore)))
                 main_manager()->pv(*this, threads, tt, rootDepth);
+
+            if (threads.stop)
+                break;
         }
 
         if (!threads.stop)
@@ -393,7 +401,7 @@ void Search::Worker::iterative_deepening() {
         // We make sure not to pick an unproven mated-in score,
         // in case this thread prematurely stopped search (aborted-search).
         if (threads.abortedSearch && rootMoves[0].score != -VALUE_INFINITE
-            && rootMoves[0].score <= VALUE_TB_LOSS_IN_MAX_PLY)
+            && is_loss(rootMoves[0].score))
         {
             // Bring the last best move to the front for best thread selection.
             Utility::move_to_front(rootMoves, [&lastBestPV = std::as_const(lastBestPV)](
@@ -425,7 +433,7 @@ void Search::Worker::iterative_deepening() {
             skill.pick_best(rootMoves, multiPV);
 
         // Use part of the gained time from a previous stable move for the current move
-        for (Thread* th : threads)
+        for (auto&& th : threads)
         {
             totBestMoveChanges += th->worker->bestMoveChanges;
             th->worker->bestMoveChanges = 0;
@@ -436,20 +444,19 @@ void Search::Worker::iterative_deepening() {
         {
             int nodesEffort = rootMoves[0].effort * 100 / std::max(size_t(1), size_t(nodes));
 
-            double fallingEval = (1067 + 223 * (mainThread->bestPreviousAverageScore - bestValue)
-                                  + 97 * (mainThread->iterValue[iterIdx] - bestValue))
-                               / 10000.0;
+            double fallingEval = (11 + 2 * (mainThread->bestPreviousAverageScore - bestValue)
+                                  + (mainThread->iterValue[iterIdx] - bestValue))
+                               / 100.0;
             fallingEval = std::clamp(fallingEval, 0.580, 1.667);
 
             // If the bestMove is stable over several iterations, reduce time accordingly
             timeReduction    = lastBestMoveDepth + 8 < completedDepth ? 1.495 : 0.687;
             double reduction = (1.48 + mainThread->previousTimeReduction) / (2.17 * timeReduction);
             double bestMoveInstability = 1 + 1.88 * totBestMoveChanges / threads.size();
-            int    el                  = std::clamp((bestValue + 750) / 150, 0, 9);
             double recapture           = limits.capSq == rootMoves[0].pv[0].to_sq() ? 0.955 : 1.005;
 
-            double totalTime = mainThread->tm.optimum() * fallingEval * reduction
-                             * bestMoveInstability * EvalLevel[el] * recapture;
+            double totalTime =
+              mainThread->tm.optimum() * fallingEval * reduction * bestMoveInstability * recapture;
 
             // Cap used time in case of a single legal move for a better viewer experience
             if (rootMoves.size() == 1)
@@ -491,41 +498,53 @@ void Search::Worker::iterative_deepening() {
                              skill.best ? skill.best : skill.pick_best(rootMoves, multiPV)));
 }
 
+// Reset histories, usually before a new game
 void Search::Worker::clear() {
-    counterMoves.fill(Move::none());
     mainHistory.fill(0);
-    captureHistory.fill(0);
-    pawnHistory.fill(0);
-    correctionHistory.fill(0);
+    lowPlyHistory.fill(0);
+    captureHistory.fill(-758);
+    pawnHistory.fill(-1158);
+    pawnCorrectionHistory.fill(0);
+    majorPieceCorrectionHistory.fill(0);
+    minorPieceCorrectionHistory.fill(0);
+    nonPawnCorrectionHistory[WHITE].fill(0);
+    nonPawnCorrectionHistory[BLACK].fill(0);
+
+    for (auto& to : continuationCorrectionHistory)
+        for (auto& h : to)
+            h->fill(0);
 
     for (bool inCheck : {false, true})
         for (StatsType c : {NoCaptures, Captures})
             for (auto& to : continuationHistory[inCheck][c])
                 for (auto& h : to)
-                    h->fill(-60);
+                    h->fill(-645);
 
     for (size_t i = 1; i < reductions.size(); ++i)
-        reductions[i] = int((18.93 + std::log(size_t(options["Threads"])) / 2) * std::log(i));
+        reductions[i] = int((19.43 + std::log(size_t(options["Threads"])) / 2) * std::log(i));
 
-    refreshTable.clear(networks);
+    refreshTable.clear(networks[numaAccessToken]);
 }
 
 
-// Main search function for both PV and non-PV nodes.
+// Main search function for both PV and non-PV nodes
 template<NodeType nodeType>
 Value Search::Worker::search(
   Position& pos, Stack* ss, Value alpha, Value beta, Depth depth, bool cutNode) {
 
     constexpr bool PvNode   = nodeType != NonPV;
     constexpr bool rootNode = nodeType == Root;
+    const bool     allNode  = !(PvNode || cutNode);
 
     // Dive into quiescence search when the depth reaches zero
     if (depth <= 0)
         return qsearch < PvNode ? PV : NonPV > (pos, ss, alpha, beta);
 
-    // Check if we have an upcoming move that draws by repetition, or
-    // if the opponent had an alternative move earlier to this position.
-    if (!rootNode && alpha < VALUE_DRAW && pos.has_game_cycle(ss->ply))
+    // Limit the depth if extensions made it too large
+    depth = std::min(depth, MAX_PLY - 1);
+
+    // Check if we have an upcoming move that draws by repetition
+    if (!rootNode && alpha < VALUE_DRAW && pos.upcoming_repetition(ss->ply))
     {
         alpha = value_draw(this->nodes);
         if (alpha >= beta)
@@ -537,28 +556,29 @@ Value Search::Worker::search(
     assert(0 < depth && depth < MAX_PLY);
     assert(!(PvNode && cutNode));
 
-    Move      pv[MAX_PLY + 1], capturesSearched[32], quietsSearched[32];
+    Move      pv[MAX_PLY + 1];
     StateInfo st;
     ASSERT_ALIGNED(&st, Eval::NNUE::CacheLineSize);
 
-    TTEntry* tte;
-    Key      posKey;
-    Move     ttMove, move, excludedMove, bestMove;
-    Depth    extension, newDepth;
-    Value    bestValue, value, ttValue, eval, maxValue, probCutBeta;
-    bool     givesCheck, improving, priorCapture, opponentWorsening;
-    bool     capture, moveCountPruning, ttCapture;
-    Piece    movedPiece;
-    int      moveCount, captureCount, quietCount;
+    Key   posKey;
+    Move  move, excludedMove, bestMove;
+    Depth extension, newDepth;
+    Value bestValue, value, eval, maxValue, probCutBeta;
+    bool  givesCheck, improving, priorCapture, opponentWorsening;
+    bool  capture, ttCapture;
+    Piece movedPiece;
+
+    ValueList<Move, 32> capturesSearched;
+    ValueList<Move, 32> quietsSearched;
 
     // Step 1. Initialize node
     Worker* thisThread = this;
     ss->inCheck        = pos.checkers();
     priorCapture       = pos.captured_piece();
     Color us           = pos.side_to_move();
-    moveCount = captureCount = quietCount = ss->moveCount = 0;
-    bestValue                                             = -VALUE_INFINITE;
-    maxValue                                              = VALUE_INFINITE;
+    ss->moveCount      = 0;
+    bestValue          = -VALUE_INFINITE;
+    maxValue           = VALUE_INFINITE;
 
     // Check for the available remaining time
     if (is_mainthread())
@@ -574,7 +594,8 @@ Value Search::Worker::search(
         if (threads.stop.load(std::memory_order_relaxed) || pos.is_draw(ss->ply)
             || ss->ply >= MAX_PLY)
             return (ss->ply >= MAX_PLY && !ss->inCheck)
-                   ? evaluate(networks, pos, refreshTable, thisThread->optimism[us])
+                   ? evaluate(networks[numaAccessToken], pos, refreshTable,
+                              thisThread->optimism[us])
                    : value_draw(thisThread->nodes);
 
         // Step 3. Mate distance pruning. Even if we mate at the next move our score
@@ -588,43 +609,42 @@ Value Search::Worker::search(
         if (alpha >= beta)
             return alpha;
     }
-    else
-        thisThread->rootDelta = beta - alpha;
 
     assert(0 <= ss->ply && ss->ply < MAX_PLY);
 
-    bestMove             = Move::none();
-    (ss + 2)->killers[0] = (ss + 2)->killers[1] = Move::none();
-    (ss + 2)->cutoffCnt                         = 0;
+    bestMove            = Move::none();
+    (ss + 2)->cutoffCnt = 0;
     Square prevSq = ((ss - 1)->currentMove).is_ok() ? ((ss - 1)->currentMove).to_sq() : SQ_NONE;
     ss->statScore = 0;
 
-    // Step 4. Transposition table lookup.
-    excludedMove = ss->excludedMove;
-    posKey       = pos.key();
-    tte          = tt.probe(posKey, ss->ttHit);
-    ttValue   = ss->ttHit ? value_from_tt(tte->value(), ss->ply, pos.rule50_count()) : VALUE_NONE;
-    ttMove    = rootNode  ? thisThread->rootMoves[thisThread->pvIdx].pv[0]
-              : ss->ttHit ? tte->move()
-                          : Move::none();
-    ttCapture = ttMove && pos.capture_stage(ttMove);
+    // Step 4. Transposition table lookup
+    excludedMove                   = ss->excludedMove;
+    posKey                         = pos.key();
+    auto [ttHit, ttData, ttWriter] = tt.probe(posKey);
+    // Need further processing of the saved data
+    ss->ttHit    = ttHit;
+    ttData.move  = rootNode ? thisThread->rootMoves[thisThread->pvIdx].pv[0]
+                 : ttHit    ? ttData.move
+                            : Move::none();
+    ttData.value = ttHit ? value_from_tt(ttData.value, ss->ply, pos.rule50_count()) : VALUE_NONE;
+    ss->ttPv     = excludedMove ? ss->ttPv : PvNode || (ttHit && ttData.is_pv);
+    ttCapture    = ttData.move && pos.capture_stage(ttData.move);
 
     // At this point, if excluded, skip straight to step 6, static eval. However,
     // to save indentation, we list the condition in all code between here and there.
-    if (!excludedMove)
-        ss->ttPv = PvNode || (ss->ttHit && tte->is_pv());
 
     // At non-PV nodes we check for an early TT cutoff
-    if (!PvNode && !excludedMove && tte->depth() > depth
-        && ttValue != VALUE_NONE  // Possible in case of TT access race or if !ttHit
-        && (tte->bound() & (ttValue >= beta ? BOUND_LOWER : BOUND_UPPER)))
+    if (!PvNode && !excludedMove && ttData.depth > depth - (ttData.value <= beta)
+        && is_valid(ttData.value)  // Can happen when !ttHit or when access race in probe()
+        && (ttData.bound & (ttData.value >= beta ? BOUND_LOWER : BOUND_UPPER))
+        && (cutNode == (ttData.value >= beta) || depth > 8))
     {
         // If ttMove is quiet, update move sorting heuristics on TT hit (~2 Elo)
-        if (ttMove && ttValue >= beta)
+        if (ttData.move && ttData.value >= beta)
         {
             // Bonus for a quiet ttMove that fails high (~2 Elo)
             if (!ttCapture)
-                update_quiet_stats(pos, ss, *this, ttMove, stat_bonus(depth));
+                update_quiet_histories(pos, ss, *this, ttData.move, stat_bonus(depth));
 
             // Extra penalty for early quiet moves of
             // the previous ply (~1 Elo on STC, ~2 Elo on LTC)
@@ -636,9 +656,7 @@ Value Search::Worker::search(
         // Partial workaround for the graph history interaction problem
         // For high rule50 counts don't produce transposition table cutoffs.
         if (pos.rule50_count() < 90)
-            return ttValue >= beta && std::abs(ttValue) < VALUE_TB_WIN_IN_MAX_PLY
-                   ? (ttValue * 3 + beta) / 4
-                   : ttValue;
+            return ttData.value;
     }
 
     // Step 5. Tablebases probe
@@ -665,7 +683,7 @@ Value Search::Worker::search(
 
                 Value tbValue = VALUE_TB - ss->ply;
 
-                // use the range VALUE_TB to VALUE_TB_WIN_IN_MAX_PLY to score
+                // Use the range VALUE_TB to VALUE_TB_WIN_IN_MAX_PLY to score
                 value = wdl < -drawScore ? -tbValue
                       : wdl > drawScore  ? tbValue
                                          : VALUE_DRAW + 2 * wdl * drawScore;
@@ -676,9 +694,9 @@ Value Search::Worker::search(
 
                 if (b == BOUND_EXACT || (b == BOUND_LOWER ? value >= beta : value <= alpha))
                 {
-                    tte->save(posKey, value_to_tt(value, ss->ply), ss->ttPv, b,
-                              std::min(MAX_PLY - 1, depth + 6), Move::none(), VALUE_NONE,
-                              tt.generation());
+                    ttWriter.write(posKey, value_to_tt(value, ss->ply), ss->ttPv, b,
+                                   std::min(MAX_PLY - 1, depth + 6), Move::none(), VALUE_NONE,
+                                   tt.generation());
 
                     return value;
                 }
@@ -699,7 +717,7 @@ Value Search::Worker::search(
     if (ss->inCheck)
     {
         // Skip early pruning when in check
-        ss->staticEval = eval = VALUE_NONE;
+        ss->staticEval = eval = (ss - 2)->staticEval;
         improving             = false;
         goto moves_loop;
     }
@@ -707,98 +725,100 @@ Value Search::Worker::search(
     {
         // Providing the hint that this node's accumulator will be used often
         // brings significant Elo gain (~13 Elo).
-        Eval::NNUE::hint_common_parent_position(pos, networks, refreshTable);
+        Eval::NNUE::hint_common_parent_position(pos, networks[numaAccessToken], refreshTable);
         unadjustedStaticEval = eval = ss->staticEval;
     }
     else if (ss->ttHit)
     {
         // Never assume anything about values stored in TT
-        unadjustedStaticEval = tte->eval();
-        if (unadjustedStaticEval == VALUE_NONE)
-            unadjustedStaticEval = evaluate(networks, pos, refreshTable, thisThread->optimism[us]);
+        unadjustedStaticEval = ttData.eval;
+        if (!is_valid(unadjustedStaticEval))
+            unadjustedStaticEval =
+              evaluate(networks[numaAccessToken], pos, refreshTable, thisThread->optimism[us]);
         else if (PvNode)
-            Eval::NNUE::hint_common_parent_position(pos, networks, refreshTable);
+            Eval::NNUE::hint_common_parent_position(pos, networks[numaAccessToken], refreshTable);
 
-        ss->staticEval = eval = to_corrected_static_eval(unadjustedStaticEval, *thisThread, pos);
+        ss->staticEval = eval =
+          to_corrected_static_eval(unadjustedStaticEval, *thisThread, pos, ss);
 
         // ttValue can be used as a better position evaluation (~7 Elo)
-        if (ttValue != VALUE_NONE && (tte->bound() & (ttValue > eval ? BOUND_LOWER : BOUND_UPPER)))
-            eval = ttValue;
+        if (is_valid(ttData.value)
+            && (ttData.bound & (ttData.value > eval ? BOUND_LOWER : BOUND_UPPER)))
+            eval = ttData.value;
     }
     else
     {
-        unadjustedStaticEval = evaluate(networks, pos, refreshTable, thisThread->optimism[us]);
-        ss->staticEval = eval = to_corrected_static_eval(unadjustedStaticEval, *thisThread, pos);
+        unadjustedStaticEval =
+          evaluate(networks[numaAccessToken], pos, refreshTable, thisThread->optimism[us]);
+        ss->staticEval = eval =
+          to_corrected_static_eval(unadjustedStaticEval, *thisThread, pos, ss);
 
         // Static evaluation is saved as it was before adjustment by correction history
-        tte->save(posKey, VALUE_NONE, ss->ttPv, BOUND_NONE, DEPTH_NONE, Move::none(),
-                  unadjustedStaticEval, tt.generation());
+        ttWriter.write(posKey, VALUE_NONE, ss->ttPv, BOUND_NONE, DEPTH_UNSEARCHED, Move::none(),
+                       unadjustedStaticEval, tt.generation());
     }
 
     // Use static evaluation difference to improve quiet move ordering (~9 Elo)
     if (((ss - 1)->currentMove).is_ok() && !(ss - 1)->inCheck && !priorCapture)
     {
-        int bonus = std::clamp(-13 * int((ss - 1)->staticEval + ss->staticEval), -1796, 1526);
-        bonus     = bonus > 0 ? 2 * bonus : bonus / 2;
+        int bonus = std::clamp(-10 * int((ss - 1)->staticEval + ss->staticEval), -1831, 1428) + 623;
         thisThread->mainHistory[~us][((ss - 1)->currentMove).from_to()] << bonus;
         if (type_of(pos.piece_on(prevSq)) != PAWN && ((ss - 1)->currentMove).type_of() != PROMOTION)
             thisThread->pawnHistory[pawn_structure_index(pos)][pos.piece_on(prevSq)][prevSq]
-              << bonus / 2;
+              << bonus;
     }
 
     // Set up the improving flag, which is true if current static evaluation is
     // bigger than the previous static evaluation at our turn (if we were in
-    // check at our previous move we look at static evaluation at move prior to it
-    // and if we were in check at move prior to it flag is set to true) and is
+    // check at our previous move we go back until we weren't in check) and is
     // false otherwise. The improving flag is used in various pruning heuristics.
-    improving = (ss - 2)->staticEval != VALUE_NONE
-                ? ss->staticEval > (ss - 2)->staticEval
-                : (ss - 4)->staticEval != VALUE_NONE && ss->staticEval > (ss - 4)->staticEval;
+    improving = ss->staticEval > (ss - 2)->staticEval;
 
     opponentWorsening = ss->staticEval + (ss - 1)->staticEval > 2;
 
     // Step 7. Razoring (~1 Elo)
-    // If eval is really low check with qsearch if it can exceed alpha, if it can't,
-    // return a fail low.
-    // Adjust razor margin according to cutoffCnt. (~1 Elo)
-    if (eval < alpha - 433 - (302 - 141 * ((ss + 1)->cutoffCnt > 3)) * depth * depth)
+    // If eval is really low, check with qsearch if we can exceed alpha. If the
+    // search suggests we cannot exceed alpha, return a speculative fail low.
+    if (eval < alpha - 469 - 307 * depth * depth)
     {
         value = qsearch<NonPV>(pos, ss, alpha - 1, alpha);
-        if (value < alpha)
+        if (value < alpha && !is_decisive(value))
             return value;
     }
 
     // Step 8. Futility pruning: child node (~40 Elo)
     // The depth condition is important for mate finding.
-    if (!ss->ttPv && depth < 11
+    if (!ss->ttPv && depth < 14
         && eval - futility_margin(depth, cutNode && !ss->ttHit, improving, opponentWorsening)
-               - (ss - 1)->statScore / 254
+               - (ss - 1)->statScore / 290
              >= beta
-        && eval >= beta && eval < VALUE_TB_WIN_IN_MAX_PLY && (!ttMove || ttCapture))
-        return beta > VALUE_TB_LOSS_IN_MAX_PLY ? (eval + beta) / 2 : eval;
+        && eval >= beta && (!ttData.move || ttCapture) && !is_loss(beta) && !is_win(eval))
+        return beta + (eval - beta) / 3;
+
+    improving |= ss->staticEval >= beta + 100;
 
     // Step 9. Null move search with verification search (~35 Elo)
-    if (!PvNode && (ss - 1)->currentMove != Move::null() && (ss - 1)->statScore < 16993
-        && eval >= beta && ss->staticEval >= beta - 19 * depth + 326 && !excludedMove
-        && pos.non_pawn_material(us) && ss->ply >= thisThread->nmpMinPly
-        && beta > VALUE_TB_LOSS_IN_MAX_PLY)
+    if (cutNode && (ss - 1)->currentMove != Move::null() && eval >= beta
+        && ss->staticEval >= beta - 21 * depth + 421 && !excludedMove && pos.non_pawn_material(us)
+        && ss->ply >= thisThread->nmpMinPly && !is_loss(beta))
     {
         assert(eval - beta >= 0);
 
         // Null move dynamic reduction based on depth and eval
-        Depth R = std::min(int(eval - beta) / 134, 6) + depth / 3 + 4;
+        Depth R = std::min(int(eval - beta) / 235, 7) + depth / 3 + 5;
 
-        ss->currentMove         = Move::null();
-        ss->continuationHistory = &thisThread->continuationHistory[0][0][NO_PIECE][0];
+        ss->currentMove                   = Move::null();
+        ss->continuationHistory           = &thisThread->continuationHistory[0][0][NO_PIECE][0];
+        ss->continuationCorrectionHistory = &thisThread->continuationCorrectionHistory[NO_PIECE][0];
 
         pos.do_null_move(st, tt);
 
-        Value nullValue = -search<NonPV>(pos, ss + 1, -beta, -beta + 1, depth - R, !cutNode);
+        Value nullValue = -search<NonPV>(pos, ss + 1, -beta, -beta + 1, depth - R, false);
 
         pos.undo_null_move();
 
         // Do not return unproven mate or TB scores
-        if (nullValue >= beta && nullValue < VALUE_TB_WIN_IN_MAX_PLY)
+        if (nullValue >= beta && !is_win(nullValue))
         {
             if (thisThread->nmpMinPly || depth < 16)
                 return nullValue;
@@ -819,81 +839,94 @@ Value Search::Worker::search(
     }
 
     // Step 10. Internal iterative reductions (~9 Elo)
-    // For PV nodes without a ttMove, we decrease depth by 3.
-    if (PvNode && !ttMove)
+    // For PV nodes without a ttMove, we decrease depth.
+    if (PvNode && !ttData.move)
         depth -= 3;
 
-    // Use qsearch if depth <= 0.
+    // Use qsearch if depth <= 0
     if (depth <= 0)
         return qsearch<PV>(pos, ss, alpha, beta);
 
-    // For cutNodes without a ttMove, we decrease depth by 2 if depth is high enough.
-    if (cutNode && depth >= 8 && !ttMove)
-        depth -= 2;
+    // For cutNodes, if depth is high enough, decrease depth by 2 if there is no ttMove,
+    // or by 1 if there is a ttMove with an upper bound.
+    if (cutNode && depth >= 7 && (!ttData.move || ttData.bound == BOUND_UPPER))
+        depth -= 1 + !ttData.move;
 
     // Step 11. ProbCut (~10 Elo)
-    // If we have a good enough capture (or queen promotion) and a reduced search returns a value
-    // much above beta, we can (almost) safely prune the previous move.
-    probCutBeta = beta + 159 - 66 * improving;
-    if (
-      !PvNode && depth > 3
-      && std::abs(beta) < VALUE_TB_WIN_IN_MAX_PLY
-      // If value from transposition table is lower than probCutBeta, don't attempt probCut
-      // there and in further interactions with transposition table cutoff depth is set to depth - 3
-      // because probCut search has depth set to depth - 4 but we also do a move before it
-      // So effective depth is equal to depth - 3
-      && !(tte->depth() >= depth - 3 && ttValue != VALUE_NONE && ttValue < probCutBeta))
+    // If we have a good enough capture (or queen promotion) and a reduced search
+    // returns a value much above beta, we can (almost) safely prune the previous move.
+    probCutBeta = beta + 187 - 56 * improving;
+    if (!PvNode && depth > 3
+        && !is_decisive(beta)
+        // If value from transposition table is lower than probCutBeta, don't attempt
+        // probCut there and in further interactions with transposition table cutoff
+        // depth is set to depth - 3 because probCut search has depth set to depth - 4
+        // but we also do a move before it. So effective depth is equal to depth - 3.
+        && !(ttData.depth >= depth - 3 && is_valid(ttData.value) && ttData.value < probCutBeta))
     {
         assert(probCutBeta < VALUE_INFINITE && probCutBeta > beta);
 
-        MovePicker mp(pos, ttMove, probCutBeta - ss->staticEval, &thisThread->captureHistory);
+        MovePicker mp(pos, ttData.move, probCutBeta - ss->staticEval, &thisThread->captureHistory);
+        Piece      captured;
 
         while ((move = mp.next_move()) != Move::none())
-            if (move != excludedMove && pos.legal(move))
+        {
+            assert(move.is_ok());
+
+            if (move == excludedMove)
+                continue;
+
+            if (!pos.legal(move))
+                continue;
+
+            assert(pos.capture_stage(move));
+
+            movedPiece = pos.moved_piece(move);
+            captured   = pos.piece_on(move.to_sq());
+
+
+            // Prefetch the TT entry for the resulting position
+            prefetch(tt.first_entry(pos.key_after(move)));
+
+            ss->currentMove = move;
+            ss->continuationHistory =
+              &this->continuationHistory[ss->inCheck][true][pos.moved_piece(move)][move.to_sq()];
+            ss->continuationCorrectionHistory =
+              &this->continuationCorrectionHistory[pos.moved_piece(move)][move.to_sq()];
+
+            thisThread->nodes.fetch_add(1, std::memory_order_relaxed);
+            pos.do_move(move, st);
+
+            // Perform a preliminary qsearch to verify that the move holds
+            value = -qsearch<NonPV>(pos, ss + 1, -probCutBeta, -probCutBeta + 1);
+
+            // If the qsearch held, perform the regular search
+            if (value >= probCutBeta)
+                value =
+                  -search<NonPV>(pos, ss + 1, -probCutBeta, -probCutBeta + 1, depth - 4, !cutNode);
+
+            pos.undo_move(move);
+
+            if (value >= probCutBeta)
             {
-                assert(pos.capture_stage(move));
+                thisThread->captureHistory[movedPiece][move.to_sq()][type_of(captured)] << 1300;
 
-                // Prefetch the TT entry for the resulting position
-                prefetch(tt.first_entry(pos.key_after(move)));
-
-                ss->currentMove = move;
-                ss->continuationHistory =
-                  &this
-                     ->continuationHistory[ss->inCheck][true][pos.moved_piece(move)][move.to_sq()];
-
-                thisThread->nodes.fetch_add(1, std::memory_order_relaxed);
-                pos.do_move(move, st);
-
-                // Perform a preliminary qsearch to verify that the move holds
-                value = -qsearch<NonPV>(pos, ss + 1, -probCutBeta, -probCutBeta + 1);
-
-                // If the qsearch held, perform the regular search
-                if (value >= probCutBeta)
-                    value = -search<NonPV>(pos, ss + 1, -probCutBeta, -probCutBeta + 1, depth - 4,
-                                           !cutNode);
-
-                pos.undo_move(move);
-
-                if (value >= probCutBeta)
-                {
-                    // Save ProbCut data into transposition table
-                    tte->save(posKey, value_to_tt(value, ss->ply), ss->ttPv, BOUND_LOWER, depth - 3,
-                              move, unadjustedStaticEval, tt.generation());
-                    return std::abs(value) < VALUE_TB_WIN_IN_MAX_PLY ? value - (probCutBeta - beta)
-                                                                     : value;
-                }
+                // Save ProbCut data into transposition table
+                ttWriter.write(posKey, value_to_tt(value, ss->ply), ss->ttPv, BOUND_LOWER,
+                               depth - 3, move, unadjustedStaticEval, tt.generation());
+                return is_decisive(value) ? value : value - (probCutBeta - beta);
             }
+        }
 
-        Eval::NNUE::hint_common_parent_position(pos, networks, refreshTable);
+        Eval::NNUE::hint_common_parent_position(pos, networks[numaAccessToken], refreshTable);
     }
 
 moves_loop:  // When in check, search starts here
 
-    // Step 12. A small Probcut idea, when we are in check (~4 Elo)
-    probCutBeta = beta + 420;
-    if (ss->inCheck && !PvNode && ttCapture && (tte->bound() & BOUND_LOWER)
-        && tte->depth() >= depth - 4 && ttValue >= probCutBeta
-        && std::abs(ttValue) < VALUE_TB_WIN_IN_MAX_PLY && std::abs(beta) < VALUE_TB_WIN_IN_MAX_PLY)
+    // Step 12. A small Probcut idea (~4 Elo)
+    probCutBeta = beta + 417;
+    if ((ttData.bound & BOUND_LOWER) && ttData.depth >= depth - 4 && ttData.value >= probCutBeta
+        && !is_decisive(beta) && is_valid(ttData.value) && !is_decisive(ttData.value))
         return probCutBeta;
 
     const PieceToHistory* contHist[] = {(ss - 1)->continuationHistory,
@@ -903,18 +936,17 @@ moves_loop:  // When in check, search starts here
                                         nullptr,
                                         (ss - 6)->continuationHistory};
 
-    Move countermove =
-      prevSq != SQ_NONE ? thisThread->counterMoves[pos.piece_on(prevSq)][prevSq] : Move::none();
 
-    MovePicker mp(pos, ttMove, depth, &thisThread->mainHistory, &thisThread->captureHistory,
-                  contHist, &thisThread->pawnHistory, countermove, ss->killers);
+    MovePicker mp(pos, ttData.move, depth, &thisThread->mainHistory, &thisThread->lowPlyHistory,
+                  &thisThread->captureHistory, contHist, &thisThread->pawnHistory, ss->ply);
 
-    value            = bestValue;
-    moveCountPruning = false;
+    value = bestValue;
+
+    int moveCount = 0;
 
     // Step 13. Loop through all pseudo-legal moves until no moves remain
     // or a beta cutoff occurs.
-    while ((move = mp.next_move(moveCountPruning)) != Move::none())
+    while ((move = mp.next_move()) != Move::none())
     {
         assert(move.is_ok());
 
@@ -935,7 +967,7 @@ moves_loop:  // When in check, search starts here
 
         ss->moveCount = ++moveCount;
 
-        if (rootNode && is_mainthread() && elapsed() > 3000)
+        if (rootNode && is_mainthread() && nodes > 10000000)
         {
             main_manager()->updates.onIter(
               {depth, UCIEngine::move(move, pos.is_chess960()), moveCount + thisThread->pvIdx});
@@ -957,14 +989,14 @@ moves_loop:  // When in check, search starts here
 
         // Step 14. Pruning at shallow depth (~120 Elo).
         // Depth conditions are important for mate finding.
-        if (!rootNode && pos.non_pawn_material(us) && bestValue > VALUE_TB_LOSS_IN_MAX_PLY)
+        if (!rootNode && pos.non_pawn_material(us) && !is_loss(bestValue))
         {
             // Skip quiet moves if movecount exceeds our FutilityMoveCount threshold (~8 Elo)
-            if (!moveCountPruning)
-                moveCountPruning = moveCount >= futility_move_count(improving, depth);
+            if (moveCount >= futility_move_count(improving, depth))
+                mp.skip_quiet_moves();
 
             // Reduced depth of the next LMR search
-            int lmrDepth = newDepth - r;
+            int lmrDepth = newDepth - r / 1024;
 
             if (capture || givesCheck)
             {
@@ -975,15 +1007,15 @@ moves_loop:  // When in check, search starts here
                 // Futility pruning for captures (~2 Elo)
                 if (!givesCheck && lmrDepth < 7 && !ss->inCheck)
                 {
-                    Value futilityValue = ss->staticEval + 295 + 280 * lmrDepth
+                    Value futilityValue = ss->staticEval + 287 + 253 * lmrDepth
                                         + PieceValue[capturedPiece] + captHist / 7;
                     if (futilityValue <= alpha)
                         continue;
                 }
 
                 // SEE based pruning for captures and checks (~11 Elo)
-                int seeHist = std::clamp(captHist / 32, -197 * depth, 196 * depth);
-                if (!pos.see_ge(move, -186 * depth - seeHist))
+                int seeHist = std::clamp(captHist / 33, -161 * depth, 156 * depth);
+                if (!pos.see_ge(move, -162 * depth - seeHist))
                     continue;
             }
             else
@@ -991,33 +1023,32 @@ moves_loop:  // When in check, search starts here
                 int history =
                   (*contHist[0])[movedPiece][move.to_sq()]
                   + (*contHist[1])[movedPiece][move.to_sq()]
-                  + (*contHist[3])[movedPiece][move.to_sq()] / 2
                   + thisThread->pawnHistory[pawn_structure_index(pos)][movedPiece][move.to_sq()];
 
                 // Continuation history based pruning (~2 Elo)
-                if (lmrDepth < 6 && history < -4081 * depth)
+                if (history < -3884 * depth)
                     continue;
 
                 history += 2 * thisThread->mainHistory[us][move.from_to()];
 
-                lmrDepth += history / 4768;
+                lmrDepth += history / 3609;
 
                 Value futilityValue =
-                  ss->staticEval + (bestValue < ss->staticEval - 52 ? 134 : 54) + 142 * lmrDepth;
+                  ss->staticEval + (bestValue < ss->staticEval - 45 ? 140 : 43) + 141 * lmrDepth;
 
                 // Futility pruning: parent node (~13 Elo)
-                if (!ss->inCheck && lmrDepth < 13 && futilityValue <= alpha)
+                if (!ss->inCheck && lmrDepth < 12 && futilityValue <= alpha)
                 {
-                    if (bestValue <= futilityValue && std::abs(bestValue) < VALUE_TB_WIN_IN_MAX_PLY
-                        && futilityValue < VALUE_TB_WIN_IN_MAX_PLY)
-                        bestValue = (bestValue + futilityValue * 3) / 4;
+                    if (bestValue <= futilityValue && !is_decisive(bestValue)
+                        && !is_win(futilityValue))
+                        bestValue = futilityValue;
                     continue;
                 }
 
                 lmrDepth = std::max(lmrDepth, 0);
 
                 // Prune moves with negative SEE (~4 Elo)
-                if (!pos.see_ge(move, -28 * lmrDepth * lmrDepth))
+                if (!pos.see_ge(move, -25 * lmrDepth * lmrDepth))
                     continue;
             }
         }
@@ -1026,22 +1057,25 @@ moves_loop:  // When in check, search starts here
         // We take care to not overdo to avoid search getting stuck.
         if (ss->ply < thisThread->rootDepth * 2)
         {
-            // Singular extension search (~94 Elo). If all moves but one fail low on a
-            // search of (alpha-s, beta-s), and just one fails high on (alpha, beta),
-            // then that move is singular and should be extended. To verify this we do
-            // a reduced search on the position excluding the ttMove and if the result
-            // is lower than ttValue minus a margin, then we will extend the ttMove.
+            // Singular extension search (~76 Elo, ~170 nElo). If all moves but one
+            // fail low on a search of (alpha-s, beta-s), and just one fails high on
+            // (alpha, beta), then that move is singular and should be extended. To
+            // verify this we do a reduced search on the position excluding the ttMove
+            // and if the result is lower than ttValue minus a margin, then we will
+            // extend the ttMove. Recursive singular search is avoided.
 
-            // Note: the depth margin and singularBeta margin are known for having non-linear
-            // scaling. Their values are optimized to time controls of 180+1.8 and longer
-            // so changing them requires tests at these types of time controls.
-            // Recursive singular search is avoided.
-            if (!rootNode && move == ttMove && !excludedMove
-                && depth >= 4 - (thisThread->completedDepth > 32) + ss->ttPv
-                && std::abs(ttValue) < VALUE_TB_WIN_IN_MAX_PLY && (tte->bound() & BOUND_LOWER)
-                && tte->depth() >= depth - 3)
+            // Note: the depth margin and singularBeta margin are known for having
+            // non-linear scaling. Their values are optimized to time controls of
+            // 180+1.8 and longer so changing them requires tests at these types of
+            // time controls. Generally, higher singularBeta (i.e closer to ttValue)
+            // and lower extension margins scale well.
+
+            if (!rootNode && move == ttData.move && !excludedMove
+                && depth >= 4 - (thisThread->completedDepth > 33) + ss->ttPv
+                && is_valid(ttData.value) && !is_decisive(ttData.value)
+                && (ttData.bound & BOUND_LOWER) && ttData.depth >= depth - 3)
             {
-                Value singularBeta  = ttValue - (65 + 52 * (ss->ttPv && !PvNode)) * depth / 63;
+                Value singularBeta  = ttData.value - (56 + 79 * (ss->ttPv && !PvNode)) * depth / 64;
                 Depth singularDepth = newDepth / 2;
 
                 ss->excludedMove = move;
@@ -1051,55 +1085,46 @@ moves_loop:  // When in check, search starts here
 
                 if (value < singularBeta)
                 {
-                    int doubleMargin = 251 * PvNode - 241 * !ttCapture;
-                    int tripleMargin =
-                      135 + 234 * PvNode - 248 * !ttCapture + 124 * (ss->ttPv || !ttCapture);
-                    int quadMargin = 447 + 354 * PvNode - 300 * !ttCapture + 206 * ss->ttPv;
+                    int doubleMargin = 249 * PvNode - 194 * !ttCapture;
+                    int tripleMargin = 94 + 287 * PvNode - 249 * !ttCapture + 99 * ss->ttPv;
 
                     extension = 1 + (value < singularBeta - doubleMargin)
-                              + (value < singularBeta - tripleMargin)
-                              + (value < singularBeta - quadMargin);
+                              + (value < singularBeta - tripleMargin);
 
                     depth += ((!PvNode) && (depth < 14));
                 }
 
                 // Multi-cut pruning
                 // Our ttMove is assumed to fail high based on the bound of the TT entry,
-                // and if after excluding the ttMove with a reduced search we fail high over the original beta,
-                // we assume this expected cut-node is not singular (multiple moves fail high),
-                // and we can prune the whole subtree by returning a softbound.
-                else if (singularBeta >= beta)
-                {
-                    if (!ttCapture)
-                        update_quiet_histories(pos, ss, *this, ttMove, -stat_malus(depth));
-
-                    return singularBeta;
-                }
+                // and if after excluding the ttMove with a reduced search we fail high
+                // over the original beta, we assume this expected cut-node is not
+                // singular (multiple moves fail high), and we can prune the whole
+                // subtree by returning a softbound.
+                else if (value >= beta && !is_decisive(value))
+                    return value;
 
                 // Negative extensions
-                // If other moves failed high over (ttValue - margin) without the ttMove on a reduced search,
-                // but we cannot do multi-cut because (ttValue - margin) is lower than the original beta,
-                // we do not know if the ttMove is singular or can do a multi-cut,
-                // so we reduce the ttMove in favor of other moves based on some conditions:
+                // If other moves failed high over (ttValue - margin) without the
+                // ttMove on a reduced search, but we cannot do multi-cut because
+                // (ttValue - margin) is lower than the original beta, we do not know
+                // if the ttMove is singular or can do a multi-cut, so we reduce the
+                // ttMove in favor of other moves based on some conditions:
 
                 // If the ttMove is assumed to fail high over current beta (~7 Elo)
-                else if (ttValue >= beta)
+                else if (ttData.value >= beta)
                     extension = -3;
 
-                // If we are on a cutNode but the ttMove is not assumed to fail high over current beta (~1 Elo)
+                // If we are on a cutNode but the ttMove is not assumed to fail high
+                // over current beta (~1 Elo)
                 else if (cutNode)
                     extension = -2;
-
-                // If the ttMove is assumed to fail low over the value of the reduced search (~1 Elo)
-                else if (ttValue <= value)
-                    extension = -1;
             }
 
-            // Extension for capturing the previous moved piece (~0 Elo on STC, ~1 Elo on LTC)
-            else if (PvNode && move == ttMove && move.to_sq() == prevSq
+            // Extension for capturing the previous moved piece (~1 Elo at LTC)
+            else if (PvNode && move.to_sq() == prevSq
                      && thisThread->captureHistory[movedPiece][move.to_sq()]
                                                   [type_of(pos.piece_on(move.to_sq()))]
-                          > 4016)
+                          > 4321)
                 extension = 1;
         }
 
@@ -1113,68 +1138,77 @@ moves_loop:  // When in check, search starts here
         ss->currentMove = move;
         ss->continuationHistory =
           &thisThread->continuationHistory[ss->inCheck][capture][movedPiece][move.to_sq()];
-
+        ss->continuationCorrectionHistory =
+          &thisThread->continuationCorrectionHistory[movedPiece][move.to_sq()];
         uint64_t nodeCount = rootNode ? uint64_t(nodes) : 0;
 
         // Step 16. Make the move
         thisThread->nodes.fetch_add(1, std::memory_order_relaxed);
         pos.do_move(move, st, givesCheck);
 
+        // These reduction adjustments have proven non-linear scaling.
+        // They are optimized to time controls of 180 + 1.8 and longer,
+        // so changing them or adding conditions that are similar requires
+        // tests at these types of time controls.
+
         // Decrease reduction if position is or has been on the PV (~7 Elo)
         if (ss->ttPv)
-            r -= 1 + (ttValue > alpha) + (tte->depth() >= depth);
-
-        else if (cutNode && move != ttMove && move != ss->killers[0])
-            r++;
-
-        // Increase reduction for cut nodes (~4 Elo)
-        if (cutNode)
-            r += 2 - (tte->depth() >= depth && ss->ttPv);
-
-        // Increase reduction if ttMove is a capture (~3 Elo)
-        if (ttCapture)
-            r++;
+            r -= 1024 + (ttData.value > alpha) * 1024 + (ttData.depth >= depth) * 1024;
 
         // Decrease reduction for PvNodes (~0 Elo on STC, ~2 Elo on LTC)
         if (PvNode)
-            r--;
+            r -= 1024;
+
+        // These reduction adjustments have no proven non-linear scaling
+
+        // Increase reduction for cut nodes (~4 Elo)
+        if (cutNode)
+            r += 2518 - (ttData.depth >= depth && ss->ttPv) * 991;
+
+        // Increase reduction if ttMove is a capture but the current move is not a capture (~3 Elo)
+        if (ttCapture && !capture)
+            r += 1043 + (depth < 8) * 999;
 
         // Increase reduction if next ply has a lot of fail high (~5 Elo)
         if ((ss + 1)->cutoffCnt > 3)
-            r++;
+            r += 938 + allNode * 960;
 
-        // Set reduction to 0 for first picked move (ttMove) (~2 Elo)
-        // Nullifies all previous reduction adjustments to ttMove and leaves only history to do them
-        else if (move == ttMove)
-            r = 0;
+        // For first picked move (ttMove) reduce reduction (~3 Elo)
+        else if (move == ttData.move)
+            r -= 1879;
 
-        ss->statScore = 2 * thisThread->mainHistory[us][move.from_to()]
-                      + (*contHist[0])[movedPiece][move.to_sq()]
-                      + (*contHist[1])[movedPiece][move.to_sq()]
-                      + (*contHist[3])[movedPiece][move.to_sq()] - 5078;
+        if (capture)
+            ss->statScore =
+              7 * int(PieceValue[pos.captured_piece()])
+              + thisThread->captureHistory[movedPiece][move.to_sq()][type_of(pos.captured_piece())]
+              - 5000;
+        else
+            ss->statScore = 2 * thisThread->mainHistory[us][move.from_to()]
+                          + (*contHist[0])[movedPiece][move.to_sq()]
+                          + (*contHist[1])[movedPiece][move.to_sq()] - 3996;
 
         // Decrease/increase reduction for moves with a good/bad history (~8 Elo)
-        r -= ss->statScore / 12076;
+        r -= ss->statScore * 1287 / 16384;
 
         // Step 17. Late moves reduction / extension (LMR, ~117 Elo)
-        if (depth >= 2 && moveCount > 1 + rootNode)
+        if (depth >= 2 && moveCount > 1)
         {
             // In general we want to cap the LMR depth search at newDepth, but when
             // reduction is negative, we allow this move a limited search extension
             // beyond the first move depth.
             // To prevent problems when the max value is less than the min value,
             // std::clamp has been replaced by a more robust implementation.
-            Depth d = std::max(1, std::min(newDepth - r, newDepth + 1));
+            Depth d = std::max(1, std::min(newDepth - r / 1024, newDepth + !allNode));
 
             value = -search<NonPV>(pos, ss + 1, -(alpha + 1), -alpha, d, true);
 
             // Do a full-depth search when reduced LMR search fails high
             if (value > alpha && d < newDepth)
             {
-                // Adjust full-depth search based on LMR results - if the result
-                // was good enough search deeper, if it was bad enough search shallower.
-                const bool doDeeperSearch    = value > (bestValue + 40 + 2 * newDepth);  // (~1 Elo)
-                const bool doShallowerSearch = value < bestValue + newDepth;             // (~2 Elo)
+                // Adjust full-depth search based on LMR results - if the result was
+                // good enough search deeper, if it was bad enough search shallower.
+                const bool doDeeperSearch    = value > (bestValue + 42 + 2 * newDepth);  // (~1 Elo)
+                const bool doShallowerSearch = value < bestValue + 10;                   // (~2 Elo)
 
                 newDepth += doDeeperSearch - doShallowerSearch;
 
@@ -1182,10 +1216,7 @@ moves_loop:  // When in check, search starts here
                     value = -search<NonPV>(pos, ss + 1, -(alpha + 1), -alpha, newDepth, !cutNode);
 
                 // Post LMR continuation history updates (~1 Elo)
-                int bonus = value <= alpha ? -stat_malus(newDepth)
-                          : value >= beta  ? stat_bonus(newDepth)
-                                           : 0;
-
+                int bonus = 2 * (value >= beta) * stat_bonus(newDepth);
                 update_continuation_histories(ss, movedPiece, move.to_sq(), bonus);
             }
         }
@@ -1194,11 +1225,12 @@ moves_loop:  // When in check, search starts here
         else if (!PvNode || moveCount > 1)
         {
             // Increase reduction if ttMove is not present (~6 Elo)
-            if (!ttMove)
-                r += 2;
+            if (!ttData.move)
+                r += 2037;
 
             // Note that if expected reduction is high, we reduce search depth by 1 here (~9 Elo)
-            value = -search<NonPV>(pos, ss + 1, -(alpha + 1), -alpha, newDepth - (r > 3), !cutNode);
+            value =
+              -search<NonPV>(pos, ss + 1, -(alpha + 1), -alpha, newDepth - (r > 2983), !cutNode);
         }
 
         // For PV nodes only, do a full PV search on the first move or after a fail high,
@@ -1208,6 +1240,10 @@ moves_loop:  // When in check, search starts here
             (ss + 1)->pv    = pv;
             (ss + 1)->pv[0] = Move::none();
 
+            // Extend move from transposition table if we are about to dive into qsearch.
+            if (move == ttData.move && ss->ply <= thisThread->rootDepth * 2)
+                newDepth = std::max(newDepth, 1);
+
             value = -search<PV>(pos, ss + 1, -beta, -alpha, newDepth, false);
         }
 
@@ -1218,8 +1254,8 @@ moves_loop:  // When in check, search starts here
 
         // Step 20. Check for a new best move
         // Finished searching the move. If a stop occurred, the return value of
-        // the search cannot be trusted, and we return immediately without
-        // updating best move, PV and TT.
+        // the search cannot be trusted, and we return immediately without updating
+        // best move, principal variation nor transposition table.
         if (threads.stop.load(std::memory_order_relaxed))
             return VALUE_ZERO;
 
@@ -1231,7 +1267,11 @@ moves_loop:  // When in check, search starts here
             rm.effort += nodes - nodeCount;
 
             rm.averageScore =
-              rm.averageScore != -VALUE_INFINITE ? (2 * value + rm.averageScore) / 3 : value;
+              rm.averageScore != -VALUE_INFINITE ? (value + rm.averageScore) / 2 : value;
+
+            rm.meanSquaredScore = rm.meanSquaredScore != -VALUE_INFINITE * VALUE_INFINITE
+                                  ? (value * std::abs(value) + rm.meanSquaredScore) / 2
+                                  : value * std::abs(value);
 
             // PV move or new best move?
             if (moveCount == 1 || value > alpha)
@@ -1271,11 +1311,16 @@ moves_loop:  // When in check, search starts here
                 rm.score = -VALUE_INFINITE;
         }
 
-        if (value > bestValue)
+        // In case we have an alternative move equal in eval to the current bestmove,
+        // promote it to bestmove by pretending it just exceeds alpha (but not beta).
+        int inc = (value == bestValue && ss->ply + 2 >= thisThread->rootDepth
+                   && (int(nodes) & 15) == 0 && !is_win(std::abs(value) + 1));
+
+        if (value + inc > bestValue)
         {
             bestValue = value;
 
-            if (value > alpha)
+            if (value + inc > alpha)
             {
                 bestMove = move;
 
@@ -1284,14 +1329,14 @@ moves_loop:  // When in check, search starts here
 
                 if (value >= beta)
                 {
-                    ss->cutoffCnt += 1 + !ttMove;
+                    ss->cutoffCnt += !ttData.move + (extension < 2);
                     assert(value >= beta);  // Fail high
                     break;
                 }
                 else
                 {
                     // Reduce other moves if we have found at least one score improvement (~2 Elo)
-                    if (depth > 2 && depth < 13 && beta < 15868 && value > -14630)
+                    if (depth > 2 && depth < 14 && !is_decisive(value))
                         depth -= 2;
 
                     assert(depth > 0);
@@ -1305,9 +1350,9 @@ moves_loop:  // When in check, search starts here
         if (move != bestMove && moveCount <= 32)
         {
             if (capture)
-                capturesSearched[captureCount++] = move;
+                capturesSearched.push_back(move);
             else
-                quietsSearched[quietCount++] = move;
+                quietsSearched.push_back(move);
         }
     }
 
@@ -1319,30 +1364,54 @@ moves_loop:  // When in check, search starts here
     assert(moveCount || !ss->inCheck || excludedMove || !MoveList<LEGAL>(pos).size());
 
     // Adjust best value for fail high cases at non-pv nodes
-    if (!PvNode && bestValue >= beta && std::abs(bestValue) < VALUE_TB_WIN_IN_MAX_PLY
-        && std::abs(beta) < VALUE_TB_WIN_IN_MAX_PLY && std::abs(alpha) < VALUE_TB_WIN_IN_MAX_PLY)
+    if (!PvNode && bestValue >= beta && !is_decisive(bestValue) && !is_decisive(beta)
+        && !is_decisive(alpha))
         bestValue = (bestValue * depth + beta) / (depth + 1);
 
     if (!moveCount)
         bestValue = excludedMove ? alpha : ss->inCheck ? mated_in(ss->ply) : VALUE_DRAW;
 
-    // If there is a move that produces search value greater than alpha we update the stats of searched moves
+    // If there is a move that produces search value greater than alpha,
+    // we update the stats of searched moves.
     else if (bestMove)
-        update_all_stats(pos, ss, *this, bestMove, bestValue, beta, prevSq, quietsSearched,
-                         quietCount, capturesSearched, captureCount, depth);
+        update_all_stats(pos, ss, *this, bestMove, prevSq, quietsSearched, capturesSearched, depth);
 
     // Bonus for prior countermove that caused the fail low
     else if (!priorCapture && prevSq != SQ_NONE)
     {
-        int bonus = (depth > 5) + (PvNode || cutNode) + ((ss - 1)->statScore < -14455)
-                  + ((ss - 1)->moveCount > 10) + (!ss->inCheck && bestValue <= ss->staticEval - 130)
-                  + (!(ss - 1)->inCheck && bestValue <= -(ss - 1)->staticEval - 77);
+        int bonus = (117 * (depth > 5) + 39 * !allNode + 168 * ((ss - 1)->moveCount > 8)
+                     + 115 * (!ss->inCheck && bestValue <= ss->staticEval - 108)
+                     + 119 * (!(ss - 1)->inCheck && bestValue <= -(ss - 1)->staticEval - 83));
+
+        // Proportional to "how much damage we have to undo"
+        bonus += std::min(-(ss - 1)->statScore / 113, 300);
+
+        bonus = std::max(bonus, 0);
+
         update_continuation_histories(ss - 1, pos.piece_on(prevSq), prevSq,
-                                      stat_bonus(depth) * bonus);
+                                      stat_bonus(depth) * bonus / 93);
         thisThread->mainHistory[~us][((ss - 1)->currentMove).from_to()]
-          << stat_bonus(depth) * bonus / 2;
+          << stat_bonus(depth) * bonus / 179;
+
+
+        if (type_of(pos.piece_on(prevSq)) != PAWN && ((ss - 1)->currentMove).type_of() != PROMOTION)
+            thisThread->pawnHistory[pawn_structure_index(pos)][pos.piece_on(prevSq)][prevSq]
+              << stat_bonus(depth) * bonus / 24;
     }
 
+    else if (priorCapture && prevSq != SQ_NONE)
+    {
+        // bonus for prior countermoves that caused the fail low
+        Piece capturedPiece = pos.captured_piece();
+        assert(capturedPiece != NO_PIECE);
+        thisThread->captureHistory[pos.piece_on(prevSq)][prevSq][type_of(capturedPiece)]
+          << stat_bonus(depth) * 2;
+    }
+
+    // Bonus when search fails low and there is a TT move
+    else if (ttData.move && !allNode)
+        thisThread->mainHistory[us][ttData.move.from_to()] << stat_bonus(depth) * 23 / 100;
+
     if (PvNode)
         bestValue = std::min(bestValue, maxValue);
 
@@ -1351,23 +1420,36 @@ moves_loop:  // When in check, search starts here
     if (bestValue <= alpha)
         ss->ttPv = ss->ttPv || ((ss - 1)->ttPv && depth > 3);
 
-    // Write gathered information in transposition table
-    // Static evaluation is saved as it was before correction history
+    // Write gathered information in transposition table. Note that the
+    // static evaluation is saved as it was before correction history.
     if (!excludedMove && !(rootNode && thisThread->pvIdx))
-        tte->save(posKey, value_to_tt(bestValue, ss->ply), ss->ttPv,
-                  bestValue >= beta    ? BOUND_LOWER
-                  : PvNode && bestMove ? BOUND_EXACT
-                                       : BOUND_UPPER,
-                  depth, bestMove, unadjustedStaticEval, tt.generation());
+        ttWriter.write(posKey, value_to_tt(bestValue, ss->ply), ss->ttPv,
+                       bestValue >= beta    ? BOUND_LOWER
+                       : PvNode && bestMove ? BOUND_EXACT
+                                            : BOUND_UPPER,
+                       depth, bestMove, unadjustedStaticEval, tt.generation());
 
     // Adjust correction history
-    if (!ss->inCheck && (!bestMove || !pos.capture(bestMove))
-        && !(bestValue >= beta && bestValue <= ss->staticEval)
-        && !(!bestMove && bestValue >= ss->staticEval))
+    if (!ss->inCheck && !(bestMove && pos.capture(bestMove))
+        && ((bestValue < ss->staticEval && bestValue < beta)  // negative correction & no fail high
+            || (bestValue > ss->staticEval && bestMove)))     // positive correction & no fail low
     {
+        const auto       m             = (ss - 1)->currentMove;
+        static const int nonPawnWeight = 154;
+
         auto bonus = std::clamp(int(bestValue - ss->staticEval) * depth / 8,
                                 -CORRECTION_HISTORY_LIMIT / 4, CORRECTION_HISTORY_LIMIT / 4);
-        thisThread->correctionHistory[us][pawn_structure_index<Correction>(pos)] << bonus;
+        thisThread->pawnCorrectionHistory[us][pawn_structure_index<Correction>(pos)]
+          << bonus * 107 / 128;
+        thisThread->majorPieceCorrectionHistory[us][major_piece_index(pos)] << bonus * 162 / 128;
+        thisThread->minorPieceCorrectionHistory[us][minor_piece_index(pos)] << bonus * 148 / 128;
+        thisThread->nonPawnCorrectionHistory[WHITE][us][non_pawn_index<WHITE>(pos)]
+          << bonus * nonPawnWeight / 128;
+        thisThread->nonPawnCorrectionHistory[BLACK][us][non_pawn_index<BLACK>(pos)]
+          << bonus * nonPawnWeight / 128;
+
+        if (m.is_ok())
+            (*(ss - 2)->continuationCorrectionHistory)[pos.piece_on(m.to_sq())][m.to_sq()] << bonus;
     }
 
     assert(bestValue > -VALUE_INFINITE && bestValue < VALUE_INFINITE);
@@ -1376,22 +1458,23 @@ moves_loop:  // When in check, search starts here
 }
 
 
-// Quiescence search function, which is called by the main search
-// function with zero depth, or recursively with further decreasing depth per call.
-// (~155 Elo)
+// Quiescence search function, which is called by the main search function with
+// depth zero, or recursively with further decreasing depth. With depth <= 0, we
+// "should" be using static eval only, but tactical moves may confuse the static eval.
+// To fight this horizon effect, we implement this qsearch of tactical moves (~155 Elo).
+// See https://www.chessprogramming.org/Horizon_Effect
+// and https://www.chessprogramming.org/Quiescence_Search
 template<NodeType nodeType>
-Value Search::Worker::qsearch(Position& pos, Stack* ss, Value alpha, Value beta, Depth depth) {
+Value Search::Worker::qsearch(Position& pos, Stack* ss, Value alpha, Value beta) {
 
     static_assert(nodeType != Root);
     constexpr bool PvNode = nodeType == PV;
 
     assert(alpha >= -VALUE_INFINITE && alpha < beta && beta <= VALUE_INFINITE);
     assert(PvNode || (alpha == beta - 1));
-    assert(depth <= 0);
 
-    // Check if we have an upcoming move that draws by repetition, or if
-    // the opponent had an alternative move earlier to this position. (~1 Elo)
-    if (alpha < VALUE_DRAW && pos.has_game_cycle(ss->ply))
+    // Check if we have an upcoming move that draws by repetition (~1 Elo)
+    if (alpha < VALUE_DRAW && pos.upcoming_repetition(ss->ply))
     {
         alpha = value_draw(this->nodes);
         if (alpha >= beta)
@@ -1402,14 +1485,12 @@ Value Search::Worker::qsearch(Position& pos, Stack* ss, Value alpha, Value beta,
     StateInfo st;
     ASSERT_ALIGNED(&st, Eval::NNUE::CacheLineSize);
 
-    TTEntry* tte;
-    Key      posKey;
-    Move     ttMove, move, bestMove;
-    Depth    ttDepth;
-    Value    bestValue, value, ttValue, futilityBase;
-    bool     pvHit, givesCheck, capture;
-    int      moveCount;
-    Color    us = pos.side_to_move();
+    Key   posKey;
+    Move  move, bestMove;
+    Value bestValue, value, futilityBase;
+    bool  pvHit, givesCheck, capture;
+    int   moveCount;
+    Color us = pos.side_to_move();
 
     // Step 1. Initialize node
     if (PvNode)
@@ -1430,26 +1511,25 @@ Value Search::Worker::qsearch(Position& pos, Stack* ss, Value alpha, Value beta,
     // Step 2. Check for an immediate draw or maximum ply reached
     if (pos.is_draw(ss->ply) || ss->ply >= MAX_PLY)
         return (ss->ply >= MAX_PLY && !ss->inCheck)
-               ? evaluate(networks, pos, refreshTable, thisThread->optimism[us])
+               ? evaluate(networks[numaAccessToken], pos, refreshTable, thisThread->optimism[us])
                : VALUE_DRAW;
 
     assert(0 <= ss->ply && ss->ply < MAX_PLY);
 
-    // Decide the replacement and cutoff priority of the qsearch TT entries
-    ttDepth = ss->inCheck || depth >= DEPTH_QS_CHECKS ? DEPTH_QS_CHECKS : DEPTH_QS_NO_CHECKS;
-
     // Step 3. Transposition table lookup
-    posKey  = pos.key();
-    tte     = tt.probe(posKey, ss->ttHit);
-    ttValue = ss->ttHit ? value_from_tt(tte->value(), ss->ply, pos.rule50_count()) : VALUE_NONE;
-    ttMove  = ss->ttHit ? tte->move() : Move::none();
-    pvHit   = ss->ttHit && tte->is_pv();
+    posKey                         = pos.key();
+    auto [ttHit, ttData, ttWriter] = tt.probe(posKey);
+    // Need further processing of the saved data
+    ss->ttHit    = ttHit;
+    ttData.move  = ttHit ? ttData.move : Move::none();
+    ttData.value = ttHit ? value_from_tt(ttData.value, ss->ply, pos.rule50_count()) : VALUE_NONE;
+    pvHit        = ttHit && ttData.is_pv;
 
     // At non-PV nodes we check for an early TT cutoff
-    if (!PvNode && tte->depth() >= ttDepth
-        && ttValue != VALUE_NONE  // Only in case of TT access race or if !ttHit
-        && (tte->bound() & (ttValue >= beta ? BOUND_LOWER : BOUND_UPPER)))
-        return ttValue;
+    if (!PvNode && ttData.depth >= DEPTH_QS
+        && is_valid(ttData.value)  // Can happen when !ttHit or when access race in probe()
+        && (ttData.bound & (ttData.value >= beta ? BOUND_LOWER : BOUND_UPPER)))
+        return ttData.value;
 
     // Step 4. Static evaluation of the position
     Value unadjustedStaticEval = VALUE_NONE;
@@ -1460,64 +1540,64 @@ Value Search::Worker::qsearch(Position& pos, Stack* ss, Value alpha, Value beta,
         if (ss->ttHit)
         {
             // Never assume anything about values stored in TT
-            unadjustedStaticEval = tte->eval();
-            if (unadjustedStaticEval == VALUE_NONE)
+            unadjustedStaticEval = ttData.eval;
+            if (!is_valid(unadjustedStaticEval))
                 unadjustedStaticEval =
-                  evaluate(networks, pos, refreshTable, thisThread->optimism[us]);
+                  evaluate(networks[numaAccessToken], pos, refreshTable, thisThread->optimism[us]);
             ss->staticEval = bestValue =
-              to_corrected_static_eval(unadjustedStaticEval, *thisThread, pos);
+              to_corrected_static_eval(unadjustedStaticEval, *thisThread, pos, ss);
 
             // ttValue can be used as a better position evaluation (~13 Elo)
-            if (ttValue != VALUE_NONE
-                && (tte->bound() & (ttValue > bestValue ? BOUND_LOWER : BOUND_UPPER)))
-                bestValue = ttValue;
+            if (is_valid(ttData.value) && !is_decisive(ttData.value)
+                && (ttData.bound & (ttData.value > bestValue ? BOUND_LOWER : BOUND_UPPER)))
+                bestValue = ttData.value;
         }
         else
         {
-            // In case of null move search, use previous static eval with a different sign
-            unadjustedStaticEval = (ss - 1)->currentMove != Move::null()
-                                   ? evaluate(networks, pos, refreshTable, thisThread->optimism[us])
-                                   : -(ss - 1)->staticEval;
-            ss->staticEval       = bestValue =
-              to_corrected_static_eval(unadjustedStaticEval, *thisThread, pos);
+            // In case of null move search, use previous static eval with opposite sign
+            unadjustedStaticEval =
+              (ss - 1)->currentMove != Move::null()
+                ? evaluate(networks[numaAccessToken], pos, refreshTable, thisThread->optimism[us])
+                : -(ss - 1)->staticEval;
+            ss->staticEval = bestValue =
+              to_corrected_static_eval(unadjustedStaticEval, *thisThread, pos, ss);
         }
 
         // Stand pat. Return immediately if static value is at least beta
         if (bestValue >= beta)
         {
+            if (!is_decisive(bestValue))
+                bestValue = (bestValue + beta) / 2;
             if (!ss->ttHit)
-                tte->save(posKey, value_to_tt(bestValue, ss->ply), false, BOUND_LOWER, DEPTH_NONE,
-                          Move::none(), unadjustedStaticEval, tt.generation());
-
+                ttWriter.write(posKey, value_to_tt(bestValue, ss->ply), false, BOUND_LOWER,
+                               DEPTH_UNSEARCHED, Move::none(), unadjustedStaticEval,
+                               tt.generation());
             return bestValue;
         }
 
         if (bestValue > alpha)
             alpha = bestValue;
 
-        futilityBase = ss->staticEval + 270;
+        futilityBase = ss->staticEval + 306;
     }
 
     const PieceToHistory* contHist[] = {(ss - 1)->continuationHistory,
                                         (ss - 2)->continuationHistory};
 
-    // Initialize a MovePicker object for the current position, and prepare
-    // to search the moves. Because the depth is <= 0 here, only captures,
-    // queen promotions, and other checks (only if depth >= DEPTH_QS_CHECKS)
-    // will be generated.
-    Square     prevSq = ((ss - 1)->currentMove).is_ok() ? ((ss - 1)->currentMove).to_sq() : SQ_NONE;
-    MovePicker mp(pos, ttMove, depth, &thisThread->mainHistory, &thisThread->captureHistory,
-                  contHist, &thisThread->pawnHistory);
+    Square prevSq = ((ss - 1)->currentMove).is_ok() ? ((ss - 1)->currentMove).to_sq() : SQ_NONE;
 
-    int quietCheckEvasions = 0;
+    // Initialize a MovePicker object for the current position, and prepare to search
+    // the moves. We presently use two stages of move generator in quiescence search:
+    // captures, or evasions only when in check.
+    MovePicker mp(pos, ttData.move, DEPTH_QS, &thisThread->mainHistory, &thisThread->lowPlyHistory,
+                  &thisThread->captureHistory, contHist, &thisThread->pawnHistory, ss->ply);
 
-    // Step 5. Loop through all pseudo-legal moves until no moves remain
-    // or a beta cutoff occurs.
+    // Step 5. Loop through all pseudo-legal moves until no moves remain or a beta
+    // cutoff occurs.
     while ((move = mp.next_move()) != Move::none())
     {
         assert(move.is_ok());
 
-        // Check for legality
         if (!pos.legal(move))
             continue;
 
@@ -1527,10 +1607,10 @@ Value Search::Worker::qsearch(Position& pos, Stack* ss, Value alpha, Value beta,
         moveCount++;
 
         // Step 6. Pruning
-        if (bestValue > VALUE_TB_LOSS_IN_MAX_PLY && pos.non_pawn_material(us))
+        if (!is_loss(bestValue) && pos.non_pawn_material(us))
         {
             // Futility pruning and moveCount pruning (~10 Elo)
-            if (!givesCheck && move.to_sq() != prevSq && futilityBase > VALUE_TB_LOSS_IN_MAX_PLY
+            if (!givesCheck && move.to_sq() != prevSq && !is_loss(futilityBase)
                 && move.type_of() != PROMOTION)
             {
                 if (moveCount > 2)
@@ -1538,44 +1618,34 @@ Value Search::Worker::qsearch(Position& pos, Stack* ss, Value alpha, Value beta,
 
                 Value futilityValue = futilityBase + PieceValue[pos.piece_on(move.to_sq())];
 
-                // If static eval + value of piece we are going to capture is much lower
-                // than alpha we can prune this move. (~2 Elo)
+                // If static eval + value of piece we are going to capture is
+                // much lower than alpha, we can prune this move. (~2 Elo)
                 if (futilityValue <= alpha)
                 {
                     bestValue = std::max(bestValue, futilityValue);
                     continue;
                 }
 
-                // If static eval is much lower than alpha and move is not winning material
+                // If static exchange evaluation is low enough
                 // we can prune this move. (~2 Elo)
-                if (futilityBase <= alpha && !pos.see_ge(move, 1))
+                if (!pos.see_ge(move, alpha - futilityBase))
                 {
-                    bestValue = std::max(bestValue, futilityBase);
-                    continue;
-                }
-
-                // If static exchange evaluation is much worse than what is needed to not
-                // fall below alpha we can prune this move.
-                if (futilityBase > alpha && !pos.see_ge(move, (alpha - futilityBase) * 4))
-                {
-                    bestValue = alpha;
+                    bestValue = std::min(alpha, futilityBase);
                     continue;
                 }
             }
 
-            // We prune after the second quiet check evasion move, where being 'in check' is
-            // implicitly checked through the counter, and being a 'quiet move' apart from
-            // being a tt move is assumed after an increment because captures are pushed ahead.
-            if (quietCheckEvasions > 1)
-                break;
-
             // Continuation history based pruning (~3 Elo)
-            if (!capture && (*contHist[0])[pos.moved_piece(move)][move.to_sq()] < 0
-                && (*contHist[1])[pos.moved_piece(move)][move.to_sq()] < 0)
+            if (!capture
+                && (*contHist[0])[pos.moved_piece(move)][move.to_sq()]
+                       + (*contHist[1])[pos.moved_piece(move)][move.to_sq()]
+                       + thisThread->pawnHistory[pawn_structure_index(pos)][pos.moved_piece(move)]
+                                                [move.to_sq()]
+                     <= 5095)
                 continue;
 
             // Do not search moves with bad enough SEE values (~5 Elo)
-            if (!pos.see_ge(move, -69))
+            if (!pos.see_ge(move, -83))
                 continue;
         }
 
@@ -1587,13 +1657,13 @@ Value Search::Worker::qsearch(Position& pos, Stack* ss, Value alpha, Value beta,
         ss->continuationHistory =
           &thisThread
              ->continuationHistory[ss->inCheck][capture][pos.moved_piece(move)][move.to_sq()];
-
-        quietCheckEvasions += !capture && ss->inCheck;
+        ss->continuationCorrectionHistory =
+          &thisThread->continuationCorrectionHistory[pos.moved_piece(move)][move.to_sq()];
 
         // Step 7. Make and search the move
         thisThread->nodes.fetch_add(1, std::memory_order_relaxed);
         pos.do_move(move, st, givesCheck);
-        value = -qsearch<nodeType>(pos, ss + 1, -beta, -alpha, depth - 1);
+        value = -qsearch<nodeType>(pos, ss + 1, -beta, -alpha);
         pos.undo_move(move);
 
         assert(value > -VALUE_INFINITE && value < VALUE_INFINITE);
@@ -1619,61 +1689,67 @@ Value Search::Worker::qsearch(Position& pos, Stack* ss, Value alpha, Value beta,
     }
 
     // Step 9. Check for mate
-    // All legal moves have been searched. A special case: if we're in check
-    // and no legal moves were found, it is checkmate.
+    // All legal moves have been searched. A special case: if we are
+    // in check and no legal moves were found, it is checkmate.
     if (ss->inCheck && bestValue == -VALUE_INFINITE)
     {
         assert(!MoveList<LEGAL>(pos).size());
         return mated_in(ss->ply);  // Plies to mate from the root
     }
 
-    if (std::abs(bestValue) < VALUE_TB_WIN_IN_MAX_PLY && bestValue >= beta)
+    if (!is_decisive(bestValue) && bestValue >= beta)
         bestValue = (3 * bestValue + beta) / 4;
 
-    // Save gathered info in transposition table
-    // Static evaluation is saved as it was before adjustment by correction history
-    tte->save(posKey, value_to_tt(bestValue, ss->ply), pvHit,
-              bestValue >= beta ? BOUND_LOWER : BOUND_UPPER, ttDepth, bestMove,
-              unadjustedStaticEval, tt.generation());
+    // Save gathered info in transposition table. The static evaluation
+    // is saved as it was before adjustment by correction history.
+    ttWriter.write(posKey, value_to_tt(bestValue, ss->ply), pvHit,
+                   bestValue >= beta ? BOUND_LOWER : BOUND_UPPER, DEPTH_QS, bestMove,
+                   unadjustedStaticEval, tt.generation());
 
     assert(bestValue > -VALUE_INFINITE && bestValue < VALUE_INFINITE);
 
     return bestValue;
 }
 
-Depth Search::Worker::reduction(bool i, Depth d, int mn, int delta) {
+Depth Search::Worker::reduction(bool i, Depth d, int mn, int delta) const {
     int reductionScale = reductions[d] * reductions[mn];
-    return (reductionScale + 1318 - delta * 760 / rootDelta) / 1024 + (!i && reductionScale > 1066);
+    return (reductionScale + 1304 - delta * 814 / rootDelta) + (!i && reductionScale > 1423) * 1135;
 }
 
+// elapsed() returns the time elapsed since the search started. If the
+// 'nodestime' option is enabled, it will return the count of nodes searched
+// instead. This function is called to check whether the search should be
+// stopped based on predefined thresholds like time limits or nodes searched.
+//
+// elapsed_time() returns the actual time elapsed since the start of the search.
+// This function is intended for use only when printing PV outputs, and not used
+// for making decisions within the search algorithm itself.
 TimePoint Search::Worker::elapsed() const {
     return main_manager()->tm.elapsed([this]() { return threads.nodes_searched(); });
 }
 
+TimePoint Search::Worker::elapsed_time() const { return main_manager()->tm.elapsed_time(); }
+
 
 namespace {
-// Adjusts a mate or TB score from "plies to mate from the root"
-// to "plies to mate from the current position". Standard scores are unchanged.
+// Adjusts a mate or TB score from "plies to mate from the root" to
+// "plies to mate from the current position". Standard scores are unchanged.
 // The function is called before storing a value in the transposition table.
-Value value_to_tt(Value v, int ply) {
-
-    assert(v != VALUE_NONE);
-    return v >= VALUE_TB_WIN_IN_MAX_PLY ? v + ply : v <= VALUE_TB_LOSS_IN_MAX_PLY ? v - ply : v;
-}
+Value value_to_tt(Value v, int ply) { return is_win(v) ? v + ply : is_loss(v) ? v - ply : v; }
 
 
-// Inverse of value_to_tt(): it adjusts a mate or TB score
-// from the transposition table (which refers to the plies to mate/be mated from
-// current position) to "plies to mate/be mated (TB win/loss) from the root".
-// However, to avoid potentially false mate or TB scores related to the 50 moves rule
-// and the graph history interaction, we return the highest non-TB score instead.
+// Inverse of value_to_tt(): it adjusts a mate or TB score from the transposition
+// table (which refers to the plies to mate/be mated from current position) to
+// "plies to mate/be mated (TB win/loss) from the root". However, to avoid
+// potentially false mate or TB scores related to the 50 moves rule and the
+// graph history interaction, we return the highest non-TB score instead.
 Value value_from_tt(Value v, int ply, int r50c) {
 
-    if (v == VALUE_NONE)
+    if (!is_valid(v))
         return VALUE_NONE;
 
     // handle TB win or better
-    if (v >= VALUE_TB_WIN_IN_MAX_PLY)
+    if (is_win(v))
     {
         // Downgrade a potentially false mate score
         if (v >= VALUE_MATE_IN_MAX_PLY && VALUE_MATE - v > 100 - r50c)
@@ -1687,7 +1763,7 @@ Value value_from_tt(Value v, int ply, int r50c) {
     }
 
     // handle TB loss or worse
-    if (v <= VALUE_TB_LOSS_IN_MAX_PLY)
+    if (is_loss(v))
     {
         // Downgrade a potentially false mate score.
         if (v <= VALUE_MATED_IN_MAX_PLY && VALUE_MATE + v > 100 - r50c)
@@ -1714,118 +1790,88 @@ void update_pv(Move* pv, Move move, const Move* childPv) {
 
 
 // Updates stats at the end of search() when a bestMove is found
-void update_all_stats(const Position& pos,
-                      Stack*          ss,
-                      Search::Worker& workerThread,
-                      Move            bestMove,
-                      Value           bestValue,
-                      Value           beta,
-                      Square          prevSq,
-                      Move*           quietsSearched,
-                      int             quietCount,
-                      Move*           capturesSearched,
-                      int             captureCount,
-                      Depth           depth) {
+void update_all_stats(const Position&      pos,
+                      Stack*               ss,
+                      Search::Worker&      workerThread,
+                      Move                 bestMove,
+                      Square               prevSq,
+                      ValueList<Move, 32>& quietsSearched,
+                      ValueList<Move, 32>& capturesSearched,
+                      Depth                depth) {
 
     CapturePieceToHistory& captureHistory = workerThread.captureHistory;
     Piece                  moved_piece    = pos.moved_piece(bestMove);
     PieceType              captured;
 
-    int quietMoveBonus = stat_bonus(depth + 1);
-    int quietMoveMalus = stat_malus(depth);
+    int bonus = stat_bonus(depth);
+    int malus = stat_malus(depth);
 
     if (!pos.capture_stage(bestMove))
     {
-        int bestMoveBonus = bestValue > beta + 165 ? quietMoveBonus      // larger bonus
-                                                   : stat_bonus(depth);  // smaller bonus
-
-        update_quiet_stats(pos, ss, workerThread, bestMove, bestMoveBonus);
+        update_quiet_histories(pos, ss, workerThread, bestMove, bonus);
 
         // Decrease stats for all non-best quiet moves
-        for (int i = 0; i < quietCount; ++i)
-            update_quiet_histories(pos, ss, workerThread, quietsSearched[i], -quietMoveMalus);
+        for (Move move : quietsSearched)
+            update_quiet_histories(pos, ss, workerThread, move, -malus);
     }
     else
     {
         // Increase stats for the best move in case it was a capture move
         captured = type_of(pos.piece_on(bestMove.to_sq()));
-        captureHistory[moved_piece][bestMove.to_sq()][captured] << quietMoveBonus;
+        captureHistory[moved_piece][bestMove.to_sq()][captured] << bonus;
     }
 
-    // Extra penalty for a quiet early move that was not a TT move or
-    // main killer move in previous ply when it gets refuted.
-    if (prevSq != SQ_NONE
-        && ((ss - 1)->moveCount == 1 + (ss - 1)->ttHit
-            || ((ss - 1)->currentMove == (ss - 1)->killers[0]))
-        && !pos.captured_piece())
-        update_continuation_histories(ss - 1, pos.piece_on(prevSq), prevSq, -quietMoveMalus);
+    // Extra penalty for a quiet early move that was not a TT move in
+    // previous ply when it gets refuted.
+    if (prevSq != SQ_NONE && ((ss - 1)->moveCount == 1 + (ss - 1)->ttHit) && !pos.captured_piece())
+        update_continuation_histories(ss - 1, pos.piece_on(prevSq), prevSq, -malus);
 
     // Decrease stats for all non-best capture moves
-    for (int i = 0; i < captureCount; ++i)
+    for (Move move : capturesSearched)
     {
-        moved_piece = pos.moved_piece(capturesSearched[i]);
-        captured    = type_of(pos.piece_on(capturesSearched[i].to_sq()));
-        captureHistory[moved_piece][capturesSearched[i].to_sq()][captured] << -quietMoveMalus;
+        moved_piece = pos.moved_piece(move);
+        captured    = type_of(pos.piece_on(move.to_sq()));
+        captureHistory[moved_piece][move.to_sq()][captured] << -malus;
     }
 }
 
 
-// Updates histories of the move pairs formed
-// by moves at ply -1, -2, -3, -4, and -6 with current move.
+// Updates histories of the move pairs formed by moves
+// at ply -1, -2, -3, -4, and -6 with current move.
 void update_continuation_histories(Stack* ss, Piece pc, Square to, int bonus) {
 
+    bonus = bonus * 50 / 64;
+
     for (int i : {1, 2, 3, 4, 6})
     {
         // Only update the first 2 continuation histories if we are in check
         if (ss->inCheck && i > 2)
             break;
         if (((ss - i)->currentMove).is_ok())
-            (*(ss - i)->continuationHistory)[pc][to] << bonus / (1 + 3 * (i == 3));
+            (*(ss - i)->continuationHistory)[pc][to] << bonus / (1 + (i == 3));
     }
 }
 
 // Updates move sorting heuristics
-void update_refutations(const Position& pos, Stack* ss, Search::Worker& workerThread, Move move) {
-
-    // Update killers
-    if (ss->killers[0] != move)
-    {
-        ss->killers[1] = ss->killers[0];
-        ss->killers[0] = move;
-    }
-
-    // Update countermove history
-    if (((ss - 1)->currentMove).is_ok())
-    {
-        Square prevSq                                           = ((ss - 1)->currentMove).to_sq();
-        workerThread.counterMoves[pos.piece_on(prevSq)][prevSq] = move;
-    }
-}
 
 void update_quiet_histories(
   const Position& pos, Stack* ss, Search::Worker& workerThread, Move move, int bonus) {
 
     Color us = pos.side_to_move();
     workerThread.mainHistory[us][move.from_to()] << bonus;
+    if (ss->ply < LOW_PLY_HISTORY_SIZE)
+        workerThread.lowPlyHistory[ss->ply][move.from_to()] << bonus;
 
     update_continuation_histories(ss, pos.moved_piece(move), move.to_sq(), bonus);
 
     int pIndex = pawn_structure_index(pos);
-    workerThread.pawnHistory[pIndex][pos.moved_piece(move)][move.to_sq()] << bonus;
-}
-
-// Updates move sorting heuristics
-void update_quiet_stats(
-  const Position& pos, Stack* ss, Search::Worker& workerThread, Move move, int bonus) {
-
-    update_refutations(pos, ss, workerThread, move);
-    update_quiet_histories(pos, ss, workerThread, move, bonus);
+    workerThread.pawnHistory[pIndex][pos.moved_piece(move)][move.to_sq()] << bonus / 2;
 }
 
 }
 
-// When playing with strength handicap, choose the best move among a set of RootMoves
-// using a statistical rule dependent on 'level'. Idea by Heinz van Saanen.
+// When playing with strength handicap, choose the best move among a set of
+// RootMoves using a statistical rule dependent on 'level'. Idea by Heinz van Saanen.
 Move Skill::pick_best(const RootMoves& rootMoves, size_t multiPV) {
     static PRNG rng(now());  // PRNG sequence should be non-deterministic
 
@@ -1856,8 +1902,8 @@ Move Skill::pick_best(const RootMoves& rootMoves, size_t multiPV) {
 }
 
 
-// Used to print debug info and, more importantly,
-// to detect when we are out of available time and thus stop the search.
+// Used to print debug info and, more importantly, to detect
+// when we are out of available time and thus stop the search.
 void SearchManager::check_time(Search::Worker& worker) {
     if (--callsCnt > 0)
         return;
@@ -1890,18 +1936,152 @@ void SearchManager::check_time(Search::Worker& worker) {
         worker.threads.stop = worker.threads.abortedSearch = true;
 }
 
-void SearchManager::pv(const Search::Worker&     worker,
+// Used to correct and extend PVs for moves that have a TB (but not a mate) score.
+// Keeps the search based PV for as long as it is verified to maintain the game
+// outcome, truncates afterwards. Finally, extends to mate the PV, providing a
+// possible continuation (but not a proven mating line).
+void syzygy_extend_pv(const OptionsMap&         options,
+                      const Search::LimitsType& limits,
+                      Position&                 pos,
+                      RootMove&                 rootMove,
+                      Value&                    v) {
+
+    auto t_start      = std::chrono::steady_clock::now();
+    int  moveOverhead = int(options["Move Overhead"]);
+
+    // Do not use more than moveOverhead / 2 time, if time management is active
+    auto time_abort = [&t_start, &moveOverhead, &limits]() -> bool {
+        auto t_end = std::chrono::steady_clock::now();
+        return limits.use_time_management()
+            && 2 * std::chrono::duration<double, std::milli>(t_end - t_start).count()
+                 > moveOverhead;
+    };
+
+    std::list<StateInfo> sts;
+
+    // Step 0, do the rootMove, no correction allowed, as needed for MultiPV in TB.
+    auto& stRoot = sts.emplace_back();
+    pos.do_move(rootMove.pv[0], stRoot);
+    int ply = 1;
+
+    // Step 1, walk the PV to the last position in TB with correct decisive score
+    while (size_t(ply) < rootMove.pv.size())
+    {
+        Move& pvMove = rootMove.pv[ply];
+
+        RootMoves legalMoves;
+        for (const auto& m : MoveList<LEGAL>(pos))
+            legalMoves.emplace_back(m);
+
+        Tablebases::Config config = Tablebases::rank_root_moves(options, pos, legalMoves);
+        RootMove&          rm     = *std::find(legalMoves.begin(), legalMoves.end(), pvMove);
+
+        if (legalMoves[0].tbRank != rm.tbRank)
+            break;
+
+        ply++;
+
+        auto& st = sts.emplace_back();
+        pos.do_move(pvMove, st);
+
+        // Do not allow for repetitions or drawing moves along the PV in TB regime
+        if (config.rootInTB && pos.is_draw(ply))
+        {
+            pos.undo_move(pvMove);
+            ply--;
+            break;
+        }
+
+        // Full PV shown will thus be validated and end in TB.
+        // If we cannot validate the full PV in time, we do not show it.
+        if (config.rootInTB && time_abort())
+            break;
+    }
+
+    // Resize the PV to the correct part
+    rootMove.pv.resize(ply);
+
+    // Step 2, now extend the PV to mate, as if the user explored syzygy-tables.info
+    // using top ranked moves (minimal DTZ), which gives optimal mates only for simple
+    // endgames e.g. KRvK.
+    while (!pos.is_draw(0))
+    {
+        if (time_abort())
+            break;
+
+        RootMoves legalMoves;
+        for (const auto& m : MoveList<LEGAL>(pos))
+        {
+            auto&     rm = legalMoves.emplace_back(m);
+            StateInfo tmpSI;
+            pos.do_move(m, tmpSI);
+            // Give a score of each move to break DTZ ties restricting opponent mobility,
+            // but not giving the opponent a capture.
+            for (const auto& mOpp : MoveList<LEGAL>(pos))
+                rm.tbRank -= pos.capture(mOpp) ? 100 : 1;
+            pos.undo_move(m);
+        }
+
+        // Mate found
+        if (legalMoves.size() == 0)
+            break;
+
+        // Sort moves according to their above assigned rank.
+        // This will break ties for moves with equal DTZ in rank_root_moves.
+        std::stable_sort(
+          legalMoves.begin(), legalMoves.end(),
+          [](const Search::RootMove& a, const Search::RootMove& b) { return a.tbRank > b.tbRank; });
+
+        // The winning side tries to minimize DTZ, the losing side maximizes it
+        Tablebases::Config config = Tablebases::rank_root_moves(options, pos, legalMoves, true);
+
+        // If DTZ is not available we might not find a mate, so we bail out
+        if (!config.rootInTB || config.cardinality > 0)
+            break;
+
+        ply++;
+
+        Move& pvMove = legalMoves[0].pv[0];
+        rootMove.pv.push_back(pvMove);
+        auto& st = sts.emplace_back();
+        pos.do_move(pvMove, st);
+    }
+
+    // Finding a draw in this function is an exceptional case, that cannot happen
+    // during engine game play, since we have a winning score, and play correctly
+    // with TB support. However, it can be that a position is draw due to the 50 move
+    // rule if it has been been reached on the board with a non-optimal 50 move counter
+    // (e.g. 8/8/6k1/3B4/3K4/4N3/8/8 w - - 54 106 ) which TB with dtz counter rounding
+    // cannot always correctly rank. See also
+    // https://github.com/official-stockfish/Stockfish/issues/5175#issuecomment-2058893495
+    // We adjust the score to match the found PV. Note that a TB loss score can be
+    // displayed if the engine did not find a drawing move yet, but eventually search
+    // will figure it out (e.g. 1kq5/q2r4/5K2/8/8/8/8/7Q w - - 96 1 )
+    if (pos.is_draw(0))
+        v = VALUE_DRAW;
+
+    // Undo the PV moves
+    for (auto it = rootMove.pv.rbegin(); it != rootMove.pv.rend(); ++it)
+        pos.undo_move(*it);
+
+    // Inform if we couldn't get a full extension in time
+    if (time_abort())
+        sync_cout
+          << "info string Syzygy based PV extension requires more time, increase Move Overhead as needed."
+          << sync_endl;
+}
+
+void SearchManager::pv(Search::Worker&           worker,
                        const ThreadPool&         threads,
                        const TranspositionTable& tt,
-                       Depth                     depth) const {
+                       Depth                     depth) {
 
-    const auto  nodes     = threads.nodes_searched();
-    const auto& rootMoves = worker.rootMoves;
-    const auto& pos       = worker.rootPos;
-    size_t      pvIdx     = worker.pvIdx;
-    TimePoint   time      = tm.elapsed([nodes]() { return nodes; }) + 1;
-    size_t      multiPV   = std::min(size_t(worker.options["MultiPV"]), rootMoves.size());
-    uint64_t    tbHits    = threads.tb_hits() + (worker.tbConfig.rootInTB ? rootMoves.size() : 0);
+    const auto nodes     = threads.nodes_searched();
+    auto&      rootMoves = worker.rootMoves;
+    auto&      pos       = worker.rootPos;
+    size_t     pvIdx     = worker.pvIdx;
+    size_t     multiPV   = std::min(size_t(worker.options["MultiPV"]), rootMoves.size());
+    uint64_t   tbHits    = threads.tb_hits() + (worker.tbConfig.rootInTB ? rootMoves.size() : 0);
 
     for (size_t i = 0; i < multiPV; ++i)
     {
@@ -1919,11 +2099,18 @@ void SearchManager::pv(const Search::Worker&     worker,
         bool tb = worker.tbConfig.rootInTB && std::abs(v) <= VALUE_TB;
         v       = tb ? rootMoves[i].tbScore : v;
 
+        bool isExact = i != pvIdx || tb || !updated;  // tablebase- and previous-scores are exact
+
+        // Potentially correct and extend the PV, and in exceptional cases v
+        if (is_decisive(v) && std::abs(v) < VALUE_MATE_IN_MAX_PLY
+            && ((!rootMoves[i].scoreLowerbound && !rootMoves[i].scoreUpperbound) || isExact))
+            syzygy_extend_pv(worker.options, worker.limits, pos, rootMoves[i], v);
+
         std::string pv;
         for (Move m : rootMoves[i].pv)
             pv += UCIEngine::move(m, pos.is_chess960()) + " ";
 
-        // remove last whitespace
+        // Remove last whitespace
         if (!pv.empty())
             pv.pop_back();
 
@@ -1940,15 +2127,16 @@ void SearchManager::pv(const Search::Worker&     worker,
         info.score    = {v, pos};
         info.wdl      = wdl;
 
-        if (i == pvIdx && !tb && updated)  // tablebase- and previous-scores are exact
+        if (!isExact)
             info.bound = bound;
 
-        info.timeMs   = time;
-        info.nodes    = nodes;
-        info.nps      = nodes * 1000 / time;
-        info.tbHits   = tbHits;
-        info.pv       = pv;
-        info.hashfull = tt.hashfull();
+        TimePoint time = tm.elapsed_time() + 1;
+        info.timeMs    = time;
+        info.nodes     = nodes;
+        info.nps       = nodes * 1000 / time;
+        info.tbHits    = tbHits;
+        info.pv        = pv;
+        info.hashfull  = tt.hashfull();
 
         updates.onUpdateFull(info);
     }
@@ -1963,20 +2151,17 @@ bool RootMove::extract_ponder_from_tt(const TranspositionTable& tt, Position& po
     StateInfo st;
     ASSERT_ALIGNED(&st, Eval::NNUE::CacheLineSize);
 
-    bool ttHit;
-
     assert(pv.size() == 1);
     if (pv[0] == Move::none())
         return false;
 
     pos.do_move(pv[0], st);
-    TTEntry* tte = tt.probe(pos.key(), ttHit);
 
+    auto [ttHit, ttData, ttWriter] = tt.probe(pos.key());
     if (ttHit)
     {
-        Move m = tte->move();  // Local copy to be SMP safe
-        if (MoveList<LEGAL>(pos).contains(m))
-            pv.push_back(m);
+        if (MoveList<LEGAL>(pos).contains(ttData.move))
+            pv.push_back(ttData.move);
     }
 
     pos.undo_move(pv[0]);
diff --git a/src/search.h b/src/search.h
index cb73a5af..b618855b 100644
--- a/src/search.h
+++ b/src/search.h
@@ -19,6 +19,7 @@
 #ifndef SEARCH_H_INCLUDED
 #define SEARCH_H_INCLUDED
 
+#include <algorithm>
 #include <array>
 #include <atomic>
 #include <cassert>
@@ -30,21 +31,19 @@
 #include <string_view>
 #include <vector>
 
+#include "history.h"
 #include "misc.h"
-#include "movepick.h"
+#include "nnue/network.h"
+#include "nnue/nnue_accumulator.h"
+#include "numa.h"
 #include "position.h"
 #include "score.h"
 #include "syzygy/tbprobe.h"
 #include "timeman.h"
 #include "types.h"
-#include "nnue/nnue_accumulator.h"
 
 namespace Stockfish {
 
-namespace Eval::NNUE {
-struct Networks;
-}
-
 // Different node types, used as a template parameter
 enum NodeType {
     NonPV,
@@ -62,19 +61,19 @@ namespace Search {
 // shallower and deeper in the tree during the search. Each search thread has
 // its own array of Stack objects, indexed by the current ply.
 struct Stack {
-    Move*           pv;
-    PieceToHistory* continuationHistory;
-    int             ply;
-    Move            currentMove;
-    Move            excludedMove;
-    Move            killers[2];
-    Value           staticEval;
-    int             statScore;
-    int             moveCount;
-    bool            inCheck;
-    bool            ttPv;
-    bool            ttHit;
-    int             cutoffCnt;
+    Move*                       pv;
+    PieceToHistory*             continuationHistory;
+    CorrectionHistory<PieceTo>* continuationCorrectionHistory;
+    int                         ply;
+    Move                        currentMove;
+    Move                        excludedMove;
+    Value                       staticEval;
+    int                         statScore;
+    int                         moveCount;
+    bool                        inCheck;
+    bool                        ttPv;
+    bool                        ttHit;
+    int                         cutoffCnt;
 };
 
 
@@ -92,15 +91,16 @@ struct RootMove {
         return m.score != score ? m.score < score : m.previousScore < previousScore;
     }
 
-    uint64_t          effort          = 0;
-    Value             score           = -VALUE_INFINITE;
-    Value             previousScore   = -VALUE_INFINITE;
-    Value             averageScore    = -VALUE_INFINITE;
-    Value             uciScore        = -VALUE_INFINITE;
-    bool              scoreLowerbound = false;
-    bool              scoreUpperbound = false;
-    int               selDepth        = 0;
-    int               tbRank          = 0;
+    uint64_t          effort           = 0;
+    Value             score            = -VALUE_INFINITE;
+    Value             previousScore    = -VALUE_INFINITE;
+    Value             averageScore     = -VALUE_INFINITE;
+    Value             meanSquaredScore = -VALUE_INFINITE * VALUE_INFINITE;
+    Value             uciScore         = -VALUE_INFINITE;
+    bool              scoreLowerbound  = false;
+    bool              scoreUpperbound  = false;
+    int               selDepth         = 0;
+    int               tbRank           = 0;
     Value             tbScore;
     std::vector<Move> pv;
 };
@@ -133,19 +133,19 @@ struct LimitsType {
 // The UCI stores the uci options, thread pool, and transposition table.
 // This struct is used to easily forward data to the Search::Worker class.
 struct SharedState {
-    SharedState(const OptionsMap&           optionsMap,
-                ThreadPool&                 threadPool,
-                TranspositionTable&         transpositionTable,
-                const Eval::NNUE::Networks& nets) :
+    SharedState(const OptionsMap&                               optionsMap,
+                ThreadPool&                                     threadPool,
+                TranspositionTable&                             transpositionTable,
+                const LazyNumaReplicated<Eval::NNUE::Networks>& nets) :
         options(optionsMap),
         threads(threadPool),
         tt(transpositionTable),
         networks(nets) {}
 
-    const OptionsMap&           options;
-    ThreadPool&                 threads;
-    TranspositionTable&         tt;
-    const Eval::NNUE::Networks& networks;
+    const OptionsMap&                               options;
+    ThreadPool&                                     threads;
+    TranspositionTable&                             tt;
+    const LazyNumaReplicated<Eval::NNUE::Networks>& networks;
 };
 
 class Worker;
@@ -182,6 +182,34 @@ struct InfoIteration {
     size_t           currmovenumber;
 };
 
+// Skill structure is used to implement strength limit. If we have a UCI_Elo,
+// we convert it to an appropriate skill level, anchored to the Stash engine.
+// This method is based on a fit of the Elo results for games played between
+// Stockfish at various skill levels and various versions of the Stash engine.
+// Skill 0 .. 19 now covers CCRL Blitz Elo from 1320 to 3190, approximately
+// Reference: https://github.com/vondele/Stockfish/commit/a08b8d4e9711c2
+struct Skill {
+    // Lowest and highest Elo ratings used in the skill level calculation
+    constexpr static int LowestElo  = 1320;
+    constexpr static int HighestElo = 3190;
+
+    Skill(int skill_level, int uci_elo) {
+        if (uci_elo)
+        {
+            double e = double(uci_elo - LowestElo) / (HighestElo - LowestElo);
+            level = std::clamp((((37.2473 * e - 40.8525) * e + 22.2943) * e - 0.311438), 0.0, 19.0);
+        }
+        else
+            level = double(skill_level);
+    }
+    bool enabled() const { return level < 20.0; }
+    bool time_to_pick(Depth depth) const { return depth == 1 + int(level); }
+    Move pick_best(const RootMoves&, size_t multiPV);
+
+    double level;
+    Move   best = Move::none();
+};
+
 // SearchManager manages the search from the main thread. It is responsible for
 // keeping track of the time, and storing data strictly related to the main thread.
 class SearchManager: public ISearchManager {
@@ -204,12 +232,13 @@ class SearchManager: public ISearchManager {
 
     void check_time(Search::Worker& worker) override;
 
-    void pv(const Search::Worker&     worker,
+    void pv(Search::Worker&           worker,
             const ThreadPool&         threads,
             const TranspositionTable& tt,
-            Depth                     depth) const;
+            Depth                     depth);
 
     Stockfish::TimeManagement tm;
+    double                    originalTimeAdjust;
     int                       callsCnt;
     std::atomic_bool          ponder;
 
@@ -235,47 +264,55 @@ class NullSearchManager: public ISearchManager {
 // of the search history, and storing data required for the search.
 class Worker {
    public:
-    Worker(SharedState&, std::unique_ptr<ISearchManager>, size_t);
+    Worker(SharedState&, std::unique_ptr<ISearchManager>, size_t, NumaReplicatedAccessToken);
 
-    // Called at instantiation to initialize Reductions tables
-    // Reset histories, usually before a new game
+    // Called at instantiation to initialize reductions tables.
+    // Reset histories, usually before a new game.
     void clear();
 
     // Called when the program receives the UCI 'go' command.
     // It searches from the root position and outputs the "bestmove".
     void start_searching();
 
-    bool is_mainthread() const { return thread_idx == 0; }
+    bool is_mainthread() const { return threadIdx == 0; }
+
+    void ensure_network_replicated();
 
     // Public because they need to be updatable by the stats
-    CounterMoveHistory    counterMoves;
-    ButterflyHistory      mainHistory;
+    ButterflyHistory mainHistory;
+    LowPlyHistory    lowPlyHistory;
+
     CapturePieceToHistory captureHistory;
     ContinuationHistory   continuationHistory[2][2];
     PawnHistory           pawnHistory;
-    CorrectionHistory     correctionHistory;
+
+    CorrectionHistory<Pawn>         pawnCorrectionHistory;
+    CorrectionHistory<Major>        majorPieceCorrectionHistory;
+    CorrectionHistory<Minor>        minorPieceCorrectionHistory;
+    CorrectionHistory<NonPawn>      nonPawnCorrectionHistory[COLOR_NB];
+    CorrectionHistory<Continuation> continuationCorrectionHistory;
 
    private:
     void iterative_deepening();
 
-    // Main search function for both PV and non-PV nodes
+    // This is the main search function, for both PV and non-PV nodes
     template<NodeType nodeType>
     Value search(Position& pos, Stack* ss, Value alpha, Value beta, Depth depth, bool cutNode);
 
     // Quiescence search function, which is called by the main search
     template<NodeType nodeType>
-    Value qsearch(Position& pos, Stack* ss, Value alpha, Value beta, Depth depth = 0);
+    Value qsearch(Position& pos, Stack* ss, Value alpha, Value beta);
 
-    Depth reduction(bool i, Depth d, int mn, int delta);
+    Depth reduction(bool i, Depth d, int mn, int delta) const;
 
-    // Get a pointer to the search manager, only allowed to be called by the
-    // main thread.
+    // Pointer to the search manager, only allowed to be called by the main thread
     SearchManager* main_manager() const {
-        assert(thread_idx == 0);
+        assert(threadIdx == 0);
         return static_cast<SearchManager*>(manager.get());
     }
 
     TimePoint elapsed() const;
+    TimePoint elapsed_time() const;
 
     LimitsType limits;
 
@@ -291,7 +328,8 @@ class Worker {
     Depth     rootDepth, completedDepth;
     Value     rootDelta;
 
-    size_t thread_idx;
+    size_t                    threadIdx;
+    NumaReplicatedAccessToken numaAccessToken;
 
     // Reductions lookup table initialized at startup
     std::array<int, MAX_MOVES> reductions;  // [depth or moveNumber]
@@ -301,10 +339,10 @@ class Worker {
 
     Tablebases::Config tbConfig;
 
-    const OptionsMap&           options;
-    ThreadPool&                 threads;
-    TranspositionTable&         tt;
-    const Eval::NNUE::Networks& networks;
+    const OptionsMap&                               options;
+    ThreadPool&                                     threads;
+    TranspositionTable&                             tt;
+    const LazyNumaReplicated<Eval::NNUE::Networks>& networks;
 
     // Used by NNUE
     Eval::NNUE::AccumulatorCaches refreshTable;
diff --git a/src/syzygy/tbprobe.cpp b/src/syzygy/tbprobe.cpp
index 722dc9d3..9b24e700 100644
--- a/src/syzygy/tbprobe.cpp
+++ b/src/syzygy/tbprobe.cpp
@@ -66,7 +66,7 @@ namespace {
 
 constexpr int TBPIECES = 7;  // Max number of supported pieces
 constexpr int MAX_DTZ =
-  1 << 18;  // Max DTZ supported, large enough to deal with the syzygy TB limit.
+  1 << 18;  // Max DTZ supported times 2, large enough to deal with the syzygy TB limit.
 
 enum {
     BigEndian,
@@ -443,6 +443,8 @@ class TBTables {
 
     std::deque<TBTable<WDL>> wdlTable;
     std::deque<TBTable<DTZ>> dtzTable;
+    size_t                   foundDTZFiles = 0;
+    size_t                   foundWDLFiles = 0;
 
     void insert(Key key, TBTable<WDL>* wdl, TBTable<DTZ>* dtz) {
         uint32_t homeBucket = uint32_t(key) & (Size - 1);
@@ -486,9 +488,16 @@ class TBTables {
         memset(hashTable, 0, sizeof(hashTable));
         wdlTable.clear();
         dtzTable.clear();
+        foundDTZFiles = 0;
+        foundWDLFiles = 0;
     }
-    size_t size() const { return wdlTable.size(); }
-    void   add(const std::vector<PieceType>& pieces);
+
+    void info() const {
+        sync_cout << "info string Found " << foundWDLFiles << " WDL and " << foundDTZFiles
+                  << " DTZ tablebase files (up to " << MaxCardinality << "-man)." << sync_endl;
+    }
+
+    void add(const std::vector<PieceType>& pieces);
 };
 
 TBTables TBTables;
@@ -501,13 +510,22 @@ void TBTables::add(const std::vector<PieceType>& pieces) {
 
     for (PieceType pt : pieces)
         code += PieceToChar[pt];
+    code.insert(code.find('K', 1), "v");
 
-    TBFile file(code.insert(code.find('K', 1), "v") + ".rtbw");  // KRK -> KRvK
+    TBFile file_dtz(code + ".rtbz");  // KRK -> KRvK
+    if (file_dtz.is_open())
+    {
+        file_dtz.close();
+        foundDTZFiles++;
+    }
+
+    TBFile file(code + ".rtbw");  // KRK -> KRvK
 
     if (!file.is_open())  // Only WDL file is checked
         return;
 
     file.close();
+    foundWDLFiles++;
 
     MaxCardinality = std::max(int(pieces.size()), MaxCardinality);
 
@@ -1326,7 +1344,7 @@ void Tablebases::init(const std::string& paths) {
     MaxCardinality = 0;
     TBFile::Paths  = paths;
 
-    if (paths.empty() || paths == "<empty>")
+    if (paths.empty())
         return;
 
     // MapB1H1H7[] encodes a square below a1-h8 diagonal to 0..27
@@ -1466,7 +1484,7 @@ void Tablebases::init(const std::string& paths) {
         }
     }
 
-    sync_cout << "info string Found " << TBTables.size() << " tablebases" << sync_endl;
+    TBTables.info();
 }
 
 // Probe the WDL table for a particular position.
@@ -1574,7 +1592,10 @@ int Tablebases::probe_dtz(Position& pos, ProbeState* result) {
 // Use the DTZ tables to rank root moves.
 //
 // A return value false indicates that not all probes were successful.
-bool Tablebases::root_probe(Position& pos, Search::RootMoves& rootMoves, bool rule50) {
+bool Tablebases::root_probe(Position&          pos,
+                            Search::RootMoves& rootMoves,
+                            bool               rule50,
+                            bool               rankDTZ) {
 
     ProbeState result = OK;
     StateInfo  st;
@@ -1585,7 +1606,7 @@ bool Tablebases::root_probe(Position& pos, Search::RootMoves& rootMoves, bool ru
     // Check whether a position was repeated since the last zeroing move.
     bool rep = pos.has_repeated();
 
-    int dtz, bound = rule50 ? (MAX_DTZ - 100) : 1;
+    int dtz, bound = rule50 ? (MAX_DTZ / 2 - 100) : 1;
 
     // Probe and rank each move
     for (auto& m : rootMoves)
@@ -1624,8 +1645,10 @@ bool Tablebases::root_probe(Position& pos, Search::RootMoves& rootMoves, bool ru
 
         // Better moves are ranked higher. Certain wins are ranked equally.
         // Losing moves are ranked equally unless a 50-move draw is in sight.
-        int r    = dtz > 0 ? (dtz + cnt50 <= 99 && !rep ? MAX_DTZ : MAX_DTZ - (dtz + cnt50))
-                 : dtz < 0 ? (-dtz * 2 + cnt50 < 100 ? -MAX_DTZ : -MAX_DTZ + (-dtz + cnt50))
+        int r    = dtz > 0 ? (dtz + cnt50 <= 99 && !rep ? MAX_DTZ - (rankDTZ ? dtz : 0)
+                                                        : MAX_DTZ / 2 - (dtz + cnt50))
+                 : dtz < 0 ? (-dtz * 2 + cnt50 < 100 ? -MAX_DTZ - (rankDTZ ? dtz : 0)
+                                                     : -MAX_DTZ / 2 + (-dtz + cnt50))
                            : 0;
         m.tbRank = r;
 
@@ -1633,10 +1656,11 @@ bool Tablebases::root_probe(Position& pos, Search::RootMoves& rootMoves, bool ru
         // 1 cp to cursed wins and let it grow to 49 cp as the positions gets
         // closer to a real win.
         m.tbScore = r >= bound ? VALUE_MATE - MAX_PLY - 1
-                  : r > 0      ? Value((std::max(3, r - (MAX_DTZ - 200)) * int(PawnValue)) / 200)
-                  : r == 0     ? VALUE_DRAW
-                  : r > -bound ? Value((std::min(-3, r + (MAX_DTZ - 200)) * int(PawnValue)) / 200)
-                               : -VALUE_MATE + MAX_PLY + 1;
+                  : r > 0  ? Value((std::max(3, r - (MAX_DTZ / 2 - 200)) * int(PawnValue)) / 200)
+                  : r == 0 ? VALUE_DRAW
+                  : r > -bound
+                    ? Value((std::min(-3, r + (MAX_DTZ / 2 - 200)) * int(PawnValue)) / 200)
+                    : -VALUE_MATE + MAX_PLY + 1;
     }
 
     return true;
@@ -1683,7 +1707,8 @@ bool Tablebases::root_probe_wdl(Position& pos, Search::RootMoves& rootMoves, boo
 
 Config Tablebases::rank_root_moves(const OptionsMap&  options,
                                    Position&          pos,
-                                   Search::RootMoves& rootMoves) {
+                                   Search::RootMoves& rootMoves,
+                                   bool               rankDTZ) {
     Config config;
 
     if (rootMoves.empty())
@@ -1707,7 +1732,7 @@ Config Tablebases::rank_root_moves(const OptionsMap&  options,
     if (config.cardinality >= popcount(pos.pieces()) && !pos.can_castle(ANY_CASTLING))
     {
         // Rank moves using DTZ tables
-        config.rootInTB = root_probe(pos, rootMoves, options["Syzygy50MoveRule"]);
+        config.rootInTB = root_probe(pos, rootMoves, options["Syzygy50MoveRule"], rankDTZ);
 
         if (!config.rootInTB)
         {
diff --git a/src/syzygy/tbprobe.h b/src/syzygy/tbprobe.h
index e10950f4..75a18585 100644
--- a/src/syzygy/tbprobe.h
+++ b/src/syzygy/tbprobe.h
@@ -66,9 +66,12 @@ extern int MaxCardinality;
 void     init(const std::string& paths);
 WDLScore probe_wdl(Position& pos, ProbeState* result);
 int      probe_dtz(Position& pos, ProbeState* result);
-bool     root_probe(Position& pos, Search::RootMoves& rootMoves, bool rule50);
+bool     root_probe(Position& pos, Search::RootMoves& rootMoves, bool rule50, bool rankDTZ);
 bool     root_probe_wdl(Position& pos, Search::RootMoves& rootMoves, bool rule50);
-Config   rank_root_moves(const OptionsMap& options, Position& pos, Search::RootMoves& rootMoves);
+Config   rank_root_moves(const OptionsMap&  options,
+                         Position&          pos,
+                         Search::RootMoves& rootMoves,
+                         bool               rankDTZ = false);
 
 }  // namespace Stockfish::Tablebases
 
diff --git a/src/thread.cpp b/src/thread.cpp
index 9052654b..5f73771e 100644
--- a/src/thread.cpp
+++ b/src/thread.cpp
@@ -22,19 +22,17 @@
 #include <cassert>
 #include <deque>
 #include <memory>
+#include <string>
 #include <unordered_map>
 #include <utility>
-#include <string>
 
-#include "misc.h"
 #include "movegen.h"
 #include "search.h"
 #include "syzygy/tbprobe.h"
 #include "timeman.h"
-#include "tt.h"
 #include "types.h"
-#include "ucioption.h"
 #include "uci.h"
+#include "ucioption.h"
 
 namespace Stockfish {
 
@@ -42,13 +40,24 @@ namespace Stockfish {
 // in idle_loop(). Note that 'searching' and 'exit' should be already set.
 Thread::Thread(Search::SharedState&                    sharedState,
                std::unique_ptr<Search::ISearchManager> sm,
-               size_t                                  n) :
-    worker(std::make_unique<Search::Worker>(sharedState, std::move(sm), n)),
+               size_t                                  n,
+               OptionalThreadToNumaNodeBinder          binder) :
     idx(n),
     nthreads(sharedState.options["Threads"]),
     stdThread(&Thread::idle_loop, this) {
 
     wait_for_search_finished();
+
+    run_custom_job([this, &binder, &sharedState, &sm, n]() {
+        // Use the binder to [maybe] bind the threads to a NUMA node before doing
+        // the Worker allocation. Ideally we would also allocate the SearchManager
+        // here, but that's minor.
+        this->numaAccessToken = binder();
+        this->worker =
+          std::make_unique<Search::Worker>(sharedState, std::move(sm), n, this->numaAccessToken);
+    });
+
+    wait_for_search_finished();
 }
 
 
@@ -63,38 +72,42 @@ Thread::~Thread() {
     stdThread.join();
 }
 
-
 // Wakes up the thread that will start the search
 void Thread::start_searching() {
-    mutex.lock();
-    searching = true;
-    mutex.unlock();   // Unlock before notifying saves a few CPU-cycles
-    cv.notify_one();  // Wake up the thread in idle_loop()
+    assert(worker != nullptr);
+    run_custom_job([this]() { worker->start_searching(); });
 }
 
+// Clears the histories for the thread worker (usually before a new game)
+void Thread::clear_worker() {
+    assert(worker != nullptr);
+    run_custom_job([this]() { worker->clear(); });
+}
 
-// Blocks on the condition variable
-// until the thread has finished searching.
+// Blocks on the condition variable until the thread has finished searching
 void Thread::wait_for_search_finished() {
 
     std::unique_lock<std::mutex> lk(mutex);
     cv.wait(lk, [&] { return !searching; });
 }
 
+// Launching a function in the thread
+void Thread::run_custom_job(std::function<void()> f) {
+    {
+        std::unique_lock<std::mutex> lk(mutex);
+        cv.wait(lk, [&] { return !searching; });
+        jobFunc   = std::move(f);
+        searching = true;
+    }
+    cv.notify_one();
+}
 
-// Thread gets parked here, blocked on the
-// condition variable, when it has no work to do.
+void Thread::ensure_network_replicated() { worker->ensure_network_replicated(); }
+
+// Thread gets parked here, blocked on the condition variable
+// when the thread has no work to do.
 
 void Thread::idle_loop() {
-
-    // If OS already scheduled us on a different group than 0 then don't overwrite
-    // the choice, eventually we are one of many one-threaded processes running on
-    // some Windows NUMA hardware, for instance in fishtest. To make it simple,
-    // just check if running threads are below a threshold, in this case, all this
-    // NUMA machinery is not needed.
-    if (nthreads > 8)
-        WinProcGroup::bind_this_thread(idx);
-
     while (true)
     {
         std::unique_lock<std::mutex> lk(mutex);
@@ -105,15 +118,17 @@ void Thread::idle_loop() {
         if (exit)
             return;
 
+        std::function<void()> job = std::move(jobFunc);
+        jobFunc                   = nullptr;
+
         lk.unlock();
 
-        worker->start_searching();
+        if (job)
+            job();
     }
 }
 
-Search::SearchManager* ThreadPool::main_manager() {
-    return static_cast<Search::SearchManager*>(main_thread()->worker.get()->manager.get());
-}
+Search::SearchManager* ThreadPool::main_manager() { return main_thread()->worker->main_manager(); }
 
 uint64_t ThreadPool::nodes_searched() const { return accumulate(&Search::Worker::nodes); }
 uint64_t ThreadPool::tb_hits() const { return accumulate(&Search::Worker::tbHits); }
@@ -121,59 +136,107 @@ uint64_t ThreadPool::tb_hits() const { return accumulate(&Search::Worker::tbHits
 // Creates/destroys threads to match the requested number.
 // Created and launched threads will immediately go to sleep in idle_loop.
 // Upon resizing, threads are recreated to allow for binding if necessary.
-void ThreadPool::set(Search::SharedState                         sharedState,
+void ThreadPool::set(const NumaConfig&                           numaConfig,
+                     Search::SharedState                         sharedState,
                      const Search::SearchManager::UpdateContext& updateContext) {
 
     if (threads.size() > 0)  // destroy any existing thread(s)
     {
         main_thread()->wait_for_search_finished();
 
-        while (threads.size() > 0)
-            delete threads.back(), threads.pop_back();
+        threads.clear();
+
+        boundThreadToNumaNode.clear();
     }
 
     const size_t requested = sharedState.options["Threads"];
 
     if (requested > 0)  // create new thread(s)
     {
-        auto manager = std::make_unique<Search::SearchManager>(updateContext);
-        threads.push_back(new Thread(sharedState, std::move(manager), 0));
+        // Binding threads may be problematic when there's multiple NUMA nodes and
+        // multiple Stockfish instances running. In particular, if each instance
+        // runs a single thread then they would all be mapped to the first NUMA node.
+        // This is undesirable, and so the default behaviour (i.e. when the user does not
+        // change the NumaConfig UCI setting) is to not bind the threads to processors
+        // unless we know for sure that we span NUMA nodes and replication is required.
+        const std::string numaPolicy(sharedState.options["NumaPolicy"]);
+        const bool        doBindThreads = [&]() {
+            if (numaPolicy == "none")
+                return false;
+
+            if (numaPolicy == "auto")
+                return numaConfig.suggests_binding_threads(requested);
+
+            // numaPolicy == "system", or explicitly set by the user
+            return true;
+        }();
+
+        boundThreadToNumaNode = doBindThreads
+                                ? numaConfig.distribute_threads_among_numa_nodes(requested)
+                                : std::vector<NumaIndex>{};
 
         while (threads.size() < requested)
         {
-            auto null_manager = std::make_unique<Search::NullSearchManager>();
-            threads.push_back(new Thread(sharedState, std::move(null_manager), threads.size()));
+            const size_t    threadId = threads.size();
+            const NumaIndex numaId   = doBindThreads ? boundThreadToNumaNode[threadId] : 0;
+            auto            manager  = threadId == 0 ? std::unique_ptr<Search::ISearchManager>(
+                                             std::make_unique<Search::SearchManager>(updateContext))
+                                                     : std::make_unique<Search::NullSearchManager>();
+
+            // When not binding threads we want to force all access to happen
+            // from the same NUMA node, because in case of NUMA replicated memory
+            // accesses we don't want to trash cache in case the threads get scheduled
+            // on the same NUMA node.
+            auto binder = doBindThreads ? OptionalThreadToNumaNodeBinder(numaConfig, numaId)
+                                        : OptionalThreadToNumaNodeBinder(numaId);
+
+            threads.emplace_back(
+              std::make_unique<Thread>(sharedState, std::move(manager), threadId, binder));
         }
 
         clear();
 
         main_thread()->wait_for_search_finished();
-
-        // Reallocate the hash with the new threadpool size
-        sharedState.tt.resize(sharedState.options["Hash"], requested);
     }
 }
 
 
 // Sets threadPool data to initial values
 void ThreadPool::clear() {
-
-    for (Thread* th : threads)
-        th->worker->clear();
-
     if (threads.size() == 0)
         return;
 
-    main_manager()->callsCnt                 = 0;
-    main_manager()->bestPreviousScore        = VALUE_INFINITE;
+    for (auto&& th : threads)
+        th->clear_worker();
+
+    for (auto&& th : threads)
+        th->wait_for_search_finished();
+
+    // These two affect the time taken on the first move of a game:
     main_manager()->bestPreviousAverageScore = VALUE_INFINITE;
-    main_manager()->previousTimeReduction    = 1.0;
+    main_manager()->previousTimeReduction    = 0.85;
+
+    main_manager()->callsCnt           = 0;
+    main_manager()->bestPreviousScore  = VALUE_INFINITE;
+    main_manager()->originalTimeAdjust = -1;
     main_manager()->tm.clear();
 }
 
+void ThreadPool::run_on_thread(size_t threadId, std::function<void()> f) {
+    assert(threads.size() > threadId);
+    threads[threadId]->run_custom_job(std::move(f));
+}
 
-// Wakes up main thread waiting in idle_loop() and
-// returns immediately. Main thread will wake up other threads and start the search.
+void ThreadPool::wait_on_thread(size_t threadId) {
+    assert(threads.size() > threadId);
+    threads[threadId]->wait_for_search_finished();
+}
+
+size_t ThreadPool::num_threads() const { return threads.size(); }
+
+
+// Wakes up main thread waiting in idle_loop() and returns immediately.
+// Main thread will wake up other threads and start the search.
 void ThreadPool::start_thinking(const OptionsMap&  options,
                                 Position&          pos,
                                 StateListPtr&      states,
@@ -213,33 +276,38 @@ void ThreadPool::start_thinking(const OptionsMap&  options,
     // We use Position::set() to set root position across threads. But there are
     // some StateInfo fields (previous, pliesFromNull, capturedPiece) that cannot
     // be deduced from a fen string, so set() clears them and they are set from
-    // setupStates->back() later. The rootState is per thread, earlier states are shared
-    // since they are read-only.
-    for (Thread* th : threads)
+    // setupStates->back() later. The rootState is per thread, earlier states are
+    // shared since they are read-only.
+    for (auto&& th : threads)
     {
-        th->worker->limits = limits;
-        th->worker->nodes = th->worker->tbHits = th->worker->nmpMinPly =
-          th->worker->bestMoveChanges          = 0;
-        th->worker->rootDepth = th->worker->completedDepth = 0;
-        th->worker->rootMoves                              = rootMoves;
-        th->worker->rootPos.set(pos.fen(), pos.is_chess960(), &th->worker->rootState);
-        th->worker->rootState = setupStates->back();
-        th->worker->tbConfig  = tbConfig;
+        th->run_custom_job([&]() {
+            th->worker->limits = limits;
+            th->worker->nodes = th->worker->tbHits = th->worker->nmpMinPly =
+              th->worker->bestMoveChanges          = 0;
+            th->worker->rootDepth = th->worker->completedDepth = 0;
+            th->worker->rootMoves                              = rootMoves;
+            th->worker->rootPos.set(pos.fen(), pos.is_chess960(), &th->worker->rootState);
+            th->worker->rootState = setupStates->back();
+            th->worker->tbConfig  = tbConfig;
+        });
     }
 
+    for (auto&& th : threads)
+        th->wait_for_search_finished();
+
     main_thread()->start_searching();
 }
 
 Thread* ThreadPool::get_best_thread() const {
 
-    Thread* bestThread = threads.front();
+    Thread* bestThread = threads.front().get();
     Value   minScore   = VALUE_NONE;
 
     std::unordered_map<Move, int64_t, Move::MoveHash> votes(
       2 * std::min(size(), bestThread->worker->rootMoves.size()));
 
     // Find the minimum score of all threads
-    for (Thread* th : threads)
+    for (auto&& th : threads)
         minScore = std::min(minScore, th->worker->rootMoves[0].score);
 
     // Vote according to score and depth, and select the best thread
@@ -247,10 +315,10 @@ Thread* ThreadPool::get_best_thread() const {
         return (th->worker->rootMoves[0].score - minScore + 14) * int(th->worker->completedDepth);
     };
 
-    for (Thread* th : threads)
-        votes[th->worker->rootMoves[0].pv[0]] += thread_voting_value(th);
+    for (auto&& th : threads)
+        votes[th->worker->rootMoves[0].pv[0]] += thread_voting_value(th.get());
 
-    for (Thread* th : threads)
+    for (auto&& th : threads)
     {
         const auto bestThreadScore = bestThread->worker->rootMoves[0].score;
         const auto newThreadScore  = th->worker->rootMoves[0].score;
@@ -261,59 +329,82 @@ Thread* ThreadPool::get_best_thread() const {
         const auto bestThreadMoveVote = votes[bestThreadPV[0]];
         const auto newThreadMoveVote  = votes[newThreadPV[0]];
 
-        const bool bestThreadInProvenWin = bestThreadScore >= VALUE_TB_WIN_IN_MAX_PLY;
-        const bool newThreadInProvenWin  = newThreadScore >= VALUE_TB_WIN_IN_MAX_PLY;
+        const bool bestThreadInProvenWin = is_win(bestThreadScore);
+        const bool newThreadInProvenWin  = is_win(newThreadScore);
 
         const bool bestThreadInProvenLoss =
-          bestThreadScore != -VALUE_INFINITE && bestThreadScore <= VALUE_TB_LOSS_IN_MAX_PLY;
+          bestThreadScore != -VALUE_INFINITE && is_loss(bestThreadScore);
         const bool newThreadInProvenLoss =
-          newThreadScore != -VALUE_INFINITE && newThreadScore <= VALUE_TB_LOSS_IN_MAX_PLY;
+          newThreadScore != -VALUE_INFINITE && is_loss(newThreadScore);
 
-        // Note that we make sure not to pick a thread with truncated-PV for better viewer experience.
+        // We make sure not to pick a thread with truncated principal variation
         const bool betterVotingValue =
-          thread_voting_value(th) * int(newThreadPV.size() > 2)
+          thread_voting_value(th.get()) * int(newThreadPV.size() > 2)
           > thread_voting_value(bestThread) * int(bestThreadPV.size() > 2);
 
         if (bestThreadInProvenWin)
         {
             // Make sure we pick the shortest mate / TB conversion
             if (newThreadScore > bestThreadScore)
-                bestThread = th;
+                bestThread = th.get();
         }
         else if (bestThreadInProvenLoss)
         {
             // Make sure we pick the shortest mated / TB conversion
             if (newThreadInProvenLoss && newThreadScore < bestThreadScore)
-                bestThread = th;
+                bestThread = th.get();
         }
         else if (newThreadInProvenWin || newThreadInProvenLoss
-                 || (newThreadScore > VALUE_TB_LOSS_IN_MAX_PLY
+                 || (!is_loss(newThreadScore)
                      && (newThreadMoveVote > bestThreadMoveVote
                          || (newThreadMoveVote == bestThreadMoveVote && betterVotingValue))))
-            bestThread = th;
+            bestThread = th.get();
     }
 
     return bestThread;
 }
 
 
-// Start non-main threads
-// Will be invoked by main thread after it has started searching
+// Start non-main threads.
+// Will be invoked by main thread after it has started searching.
 void ThreadPool::start_searching() {
 
-    for (Thread* th : threads)
+    for (auto&& th : threads)
         if (th != threads.front())
             th->start_searching();
 }
 
 
 // Wait for non-main threads
-
 void ThreadPool::wait_for_search_finished() const {
 
-    for (Thread* th : threads)
+    for (auto&& th : threads)
         if (th != threads.front())
             th->wait_for_search_finished();
 }
 
+std::vector<size_t> ThreadPool::get_bound_thread_count_by_numa_node() const {
+    std::vector<size_t> counts;
+
+    if (!boundThreadToNumaNode.empty())
+    {
+        NumaIndex highestNumaNode = 0;
+        for (NumaIndex n : boundThreadToNumaNode)
+            if (n > highestNumaNode)
+                highestNumaNode = n;
+
+        counts.resize(highestNumaNode + 1, 0);
+
+        for (NumaIndex n : boundThreadToNumaNode)
+            counts[n] += 1;
+    }
+
+    return counts;
+}
+
+void ThreadPool::ensure_network_replicated() {
+    for (auto&& th : threads)
+        th->ensure_network_replicated();
+}
+
 }  // namespace Stockfish
diff --git a/src/thread.h b/src/thread.h
index 223652ae..43e2e142 100644
--- a/src/thread.h
+++ b/src/thread.h
@@ -23,10 +23,12 @@
 #include <condition_variable>
 #include <cstddef>
 #include <cstdint>
+#include <functional>
 #include <memory>
 #include <mutex>
 #include <vector>
 
+#include "numa.h"
 #include "position.h"
 #include "search.h"
 #include "thread_win32_osx.h"
@@ -37,6 +39,32 @@ namespace Stockfish {
 class OptionsMap;
 using Value = int;
 
+// Sometimes we don't want to actually bind the threads, but the recipient still
+// needs to think it runs on *some* NUMA node, such that it can access structures
+// that rely on NUMA node knowledge. This class encapsulates this optional process
+// such that the recipient does not need to know whether the binding happened or not.
+class OptionalThreadToNumaNodeBinder {
+   public:
+    OptionalThreadToNumaNodeBinder(NumaIndex n) :
+        numaConfig(nullptr),
+        numaId(n) {}
+
+    OptionalThreadToNumaNodeBinder(const NumaConfig& cfg, NumaIndex n) :
+        numaConfig(&cfg),
+        numaId(n) {}
+
+    NumaReplicatedAccessToken operator()() const {
+        if (numaConfig != nullptr)
+            return numaConfig->bind_current_thread_to_numa_node(numaId);
+        else
+            return NumaReplicatedAccessToken(numaId);
+    }
+
+   private:
+    const NumaConfig* numaConfig;
+    NumaIndex         numaId;
+};
+
 // Abstraction of a thread. It contains a pointer to the worker and a native thread.
 // After construction, the native thread is started with idle_loop()
 // waiting for a signal to start searching.
@@ -44,22 +72,37 @@ using Value = int;
 // the search is finished, it goes back to idle_loop() waiting for a new signal.
 class Thread {
    public:
-    Thread(Search::SharedState&, std::unique_ptr<Search::ISearchManager>, size_t);
+    Thread(Search::SharedState&,
+           std::unique_ptr<Search::ISearchManager>,
+           size_t,
+           OptionalThreadToNumaNodeBinder);
     virtual ~Thread();
 
-    void   idle_loop();
-    void   start_searching();
+    void idle_loop();
+    void start_searching();
+    void clear_worker();
+    void run_custom_job(std::function<void()> f);
+
+    void ensure_network_replicated();
+
+    // Thread has been slightly altered to allow running custom jobs, so
+    // this name is no longer correct. However, this class (and ThreadPool)
+    // require further work to make them properly generic while maintaining
+    // appropriate specificity regarding search, from the point of view of an
+    // outside user, so renaming of this function is left for whenever that happens.
     void   wait_for_search_finished();
     size_t id() const { return idx; }
 
     std::unique_ptr<Search::Worker> worker;
+    std::function<void()>           jobFunc;
 
    private:
-    std::mutex              mutex;
-    std::condition_variable cv;
-    size_t                  idx, nthreads;
-    bool                    exit = false, searching = true;  // Set before starting std::thread
-    NativeThread            stdThread;
+    std::mutex                mutex;
+    std::condition_variable   cv;
+    size_t                    idx, nthreads;
+    bool                      exit = false, searching = true;  // Set before starting std::thread
+    NativeThread              stdThread;
+    NumaReplicatedAccessToken numaAccessToken;
 };
 
 
@@ -67,31 +110,46 @@ class Thread {
 // parking and, most importantly, launching a thread. All the access to threads
 // is done through this class.
 class ThreadPool {
-
    public:
+    ThreadPool() {}
+
     ~ThreadPool() {
         // destroy any existing thread(s)
         if (threads.size() > 0)
         {
             main_thread()->wait_for_search_finished();
 
-            while (threads.size() > 0)
-                delete threads.back(), threads.pop_back();
+            threads.clear();
         }
     }
 
-    void start_thinking(const OptionsMap&, Position&, StateListPtr&, Search::LimitsType);
-    void clear();
-    void set(Search::SharedState, const Search::SearchManager::UpdateContext&);
+    ThreadPool(const ThreadPool&) = delete;
+    ThreadPool(ThreadPool&&)      = delete;
+
+    ThreadPool& operator=(const ThreadPool&) = delete;
+    ThreadPool& operator=(ThreadPool&&)      = delete;
+
+    void   start_thinking(const OptionsMap&, Position&, StateListPtr&, Search::LimitsType);
+    void   run_on_thread(size_t threadId, std::function<void()> f);
+    void   wait_on_thread(size_t threadId);
+    size_t num_threads() const;
+    void   clear();
+    void   set(const NumaConfig& numaConfig,
+               Search::SharedState,
+               const Search::SearchManager::UpdateContext&);
 
     Search::SearchManager* main_manager();
-    Thread*                main_thread() const { return threads.front(); }
+    Thread*                main_thread() const { return threads.front().get(); }
     uint64_t               nodes_searched() const;
     uint64_t               tb_hits() const;
     Thread*                get_best_thread() const;
     void                   start_searching();
     void                   wait_for_search_finished() const;
 
+    std::vector<size_t> get_bound_thread_count_by_numa_node() const;
+
+    void ensure_network_replicated();
+
     std::atomic_bool stop, abortedSearch, increaseDepth;
 
     auto cbegin() const noexcept { return threads.cbegin(); }
@@ -102,13 +160,14 @@ class ThreadPool {
     auto empty() const noexcept { return threads.empty(); }
 
    private:
-    StateListPtr         setupStates;
-    std::vector<Thread*> threads;
+    StateListPtr                         setupStates;
+    std::vector<std::unique_ptr<Thread>> threads;
+    std::vector<NumaIndex>               boundThreadToNumaNode;
 
     uint64_t accumulate(std::atomic<uint64_t> Search::Worker::*member) const {
 
         uint64_t sum = 0;
-        for (Thread* th : threads)
+        for (auto&& th : threads)
             sum += (th->worker.get()->*member).load(std::memory_order_relaxed);
         return sum;
     }
diff --git a/src/timeman.cpp b/src/timeman.cpp
index c651745f..9de70fdc 100644
--- a/src/timeman.cpp
+++ b/src/timeman.cpp
@@ -32,12 +32,12 @@ TimePoint TimeManagement::optimum() const { return optimumTime; }
 TimePoint TimeManagement::maximum() const { return maximumTime; }
 
 void TimeManagement::clear() {
-    availableNodes = 0;  // When in 'nodes as time' mode
+    availableNodes = -1;  // When in 'nodes as time' mode
 }
 
 void TimeManagement::advance_nodes_time(std::int64_t nodes) {
     assert(useNodesTime);
-    availableNodes += nodes;
+    availableNodes = std::max(int64_t(0), availableNodes - nodes);
 }
 
 // Called at the beginning of the search and calculates
@@ -47,15 +47,19 @@ void TimeManagement::advance_nodes_time(std::int64_t nodes) {
 void TimeManagement::init(Search::LimitsType& limits,
                           Color               us,
                           int                 ply,
-                          const OptionsMap&   options) {
-    // If we have no time, no need to initialize TM, except for the start time,
-    // which is used by movetime.
-    startTime = limits.startTime;
+                          const OptionsMap&   options,
+                          double&             originalTimeAdjust) {
+    TimePoint npmsec = TimePoint(options["nodestime"]);
+
+    // If we have no time, we don't need to fully initialize TM.
+    // startTime is used by movetime and useNodesTime is used in elapsed calls.
+    startTime    = limits.startTime;
+    useNodesTime = npmsec != 0;
+
     if (limits.time[us] == 0)
         return;
 
     TimePoint moveOverhead = TimePoint(options["Move Overhead"]);
-    TimePoint npmsec       = TimePoint(options["nodestime"]);
 
     // optScale is a percentage of available time to use for the current move.
     // maxScale is a multiplier applied to optimumTime.
@@ -65,26 +69,31 @@ void TimeManagement::init(Search::LimitsType& limits,
     // to nodes, and use resulting values in time management formulas.
     // WARNING: to avoid time losses, the given npmsec (nodes per millisecond)
     // must be much lower than the real engine speed.
-    if (npmsec)
+    if (useNodesTime)
     {
-        useNodesTime = true;
-
-        if (!availableNodes)                            // Only once at game start
+        if (availableNodes == -1)                       // Only once at game start
             availableNodes = npmsec * limits.time[us];  // Time is in msec
 
         // Convert from milliseconds to nodes
         limits.time[us] = TimePoint(availableNodes);
         limits.inc[us] *= npmsec;
         limits.npmsec = npmsec;
+        moveOverhead *= npmsec;
     }
 
+    // These numbers are used where multiplications, divisions or comparisons
+    // with constants are involved.
+    const int64_t   scaleFactor = useNodesTime ? npmsec : 1;
+    const TimePoint scaledTime  = limits.time[us] / scaleFactor;
+    const TimePoint scaledInc   = limits.inc[us] / scaleFactor;
+
     // Maximum move horizon of 50 moves
     int mtg = limits.movestogo ? std::min(limits.movestogo, 50) : 50;
 
-    // if less than one second, gradually reduce mtg
-    if (limits.time[us] < 1000 && (double(mtg) / limits.time[us] > 0.05))
+    // If less than one second, gradually reduce mtg
+    if (scaledTime < 1000 && double(mtg) / scaledInc > 0.05)
     {
-        mtg = limits.time[us] * 0.05;
+        mtg = scaledTime * 0.05;
     }
 
     // Make sure timeLeft is > 0 since we may use it as a divisor
@@ -96,24 +105,26 @@ void TimeManagement::init(Search::LimitsType& limits,
     // game time for the current move, so also cap to a percentage of available game time.
     if (limits.movestogo == 0)
     {
-        // Use extra time with larger increments
-        double optExtra = limits.inc[us] < 500 ? 1.0 : 1.13;
+        // Extra time according to timeLeft
+        if (originalTimeAdjust < 0)
+            originalTimeAdjust = 0.3285 * std::log10(timeLeft) - 0.4830;
 
         // Calculate time constants based on current time left.
-        double optConstant =
-          std::min(0.00308 + 0.000319 * std::log10(limits.time[us] / 1000.0), 0.00506);
-        double maxConstant = std::max(3.39 + 3.01 * std::log10(limits.time[us] / 1000.0), 2.93);
+        double logTimeInSec = std::log10(scaledTime / 1000.0);
+        double optConstant  = std::min(0.00308 + 0.000319 * logTimeInSec, 0.00506);
+        double maxConstant  = std::max(3.39 + 3.01 * logTimeInSec, 2.93);
 
         optScale = std::min(0.0122 + std::pow(ply + 2.95, 0.462) * optConstant,
-                            0.213 * limits.time[us] / double(timeLeft))
-                 * optExtra;
+                            0.213 * limits.time[us] / timeLeft)
+                 * originalTimeAdjust;
+
         maxScale = std::min(6.64, maxConstant + ply / 12.0);
     }
 
     // x moves in y seconds (+ z increment)
     else
     {
-        optScale = std::min((0.88 + ply / 116.4) / mtg, 0.88 * limits.time[us] / double(timeLeft));
+        optScale = std::min((0.88 + ply / 116.4) / mtg, 0.88 * limits.time[us] / timeLeft);
         maxScale = std::min(6.3, 1.5 + 0.11 * mtg);
     }
 
diff --git a/src/timeman.h b/src/timeman.h
index 35c3cfc0..10207a8a 100644
--- a/src/timeman.h
+++ b/src/timeman.h
@@ -36,14 +36,19 @@ struct LimitsType;
 // the maximum available time, the game move number, and other parameters.
 class TimeManagement {
    public:
-    void init(Search::LimitsType& limits, Color us, int ply, const OptionsMap& options);
+    void init(Search::LimitsType& limits,
+              Color               us,
+              int                 ply,
+              const OptionsMap&   options,
+              double&             originalTimeAdjust);
 
     TimePoint optimum() const;
     TimePoint maximum() const;
     template<typename FUNC>
     TimePoint elapsed(FUNC nodes) const {
-        return useNodesTime ? TimePoint(nodes()) : now() - startTime;
+        return useNodesTime ? TimePoint(nodes()) : elapsed_time();
     }
+    TimePoint elapsed_time() const { return now() - startTime; };
 
     void clear();
     void advance_nodes_time(std::int64_t nodes);
@@ -53,7 +58,7 @@ class TimeManagement {
     TimePoint optimumTime;
     TimePoint maximumTime;
 
-    std::int64_t availableNodes = 0;      // When in 'nodes as time' mode
+    std::int64_t availableNodes = -1;     // When in 'nodes as time' mode
     bool         useNodesTime   = false;  // True if we are in 'nodes as time' mode
 };
 
diff --git a/src/tt.cpp b/src/tt.cpp
index 4885a781..75689562 100644
--- a/src/tt.cpp
+++ b/src/tt.cpp
@@ -23,31 +23,89 @@
 #include <cstdlib>
 #include <cstring>
 #include <iostream>
-#include <thread>
-#include <vector>
 
+#include "memory.h"
 #include "misc.h"
+#include "syzygy/tbprobe.h"
+#include "thread.h"
 
 namespace Stockfish {
 
+
+// TTEntry struct is the 10 bytes transposition table entry, defined as below:
+//
+// key        16 bit
+// depth       8 bit
+// generation  5 bit
+// pv node     1 bit
+// bound type  2 bit
+// move       16 bit
+// value      16 bit
+// evaluation 16 bit
+//
+// These fields are in the same order as accessed by TT::probe(), since memory is fastest sequentially.
+// Equally, the store order in save() matches this order.
+
+struct TTEntry {
+
+    // Convert internal bitfields to external types
+    TTData read() const {
+        return TTData{Move(move16),           Value(value16),
+                      Value(eval16),          Depth(depth8 + DEPTH_ENTRY_OFFSET),
+                      Bound(genBound8 & 0x3), bool(genBound8 & 0x4)};
+    }
+
+    bool is_occupied() const;
+    void save(Key k, Value v, bool pv, Bound b, Depth d, Move m, Value ev, uint8_t generation8);
+    // The returned age is a multiple of TranspositionTable::GENERATION_DELTA
+    uint8_t relative_age(const uint8_t generation8) const;
+
+   private:
+    friend class TranspositionTable;
+
+    uint16_t key16;
+    uint8_t  depth8;
+    uint8_t  genBound8;
+    Move     move16;
+    int16_t  value16;
+    int16_t  eval16;
+};
+
+// `genBound8` is where most of the details are. We use the following constants to manipulate 5 leading generation bits
+// and 3 trailing miscellaneous bits.
+
+// These bits are reserved for other things.
+static constexpr unsigned GENERATION_BITS = 3;
+// increment for generation field
+static constexpr int GENERATION_DELTA = (1 << GENERATION_BITS);
+// cycle length
+static constexpr int GENERATION_CYCLE = 255 + GENERATION_DELTA;
+// mask to pull out generation number
+static constexpr int GENERATION_MASK = (0xFF << GENERATION_BITS) & 0xFF;
+
+// DEPTH_ENTRY_OFFSET exists because 1) we use `bool(depth8)` as the occupancy check, but
+// 2) we need to store negative depths for QS. (`depth8` is the only field with "spare bits":
+// we sacrifice the ability to store depths greater than 1<<8 less the offset, as asserted in `save`.)
+bool TTEntry::is_occupied() const { return bool(depth8); }
+
 // Populates the TTEntry with a new node's data, possibly
 // overwriting an old position. The update is not atomic and can be racy.
 void TTEntry::save(
   Key k, Value v, bool pv, Bound b, Depth d, Move m, Value ev, uint8_t generation8) {
 
-    // Preserve any existing move for the same position
+    // Preserve the old ttmove if we don't have a new one
     if (m || uint16_t(k) != key16)
         move16 = m;
 
     // Overwrite less valuable entries (cheapest checks first)
-    if (b == BOUND_EXACT || uint16_t(k) != key16 || d - DEPTH_OFFSET + 2 * pv > depth8 - 4
+    if (b == BOUND_EXACT || uint16_t(k) != key16 || d - DEPTH_ENTRY_OFFSET + 2 * pv > depth8 - 4
         || relative_age(generation8))
     {
-        assert(d > DEPTH_OFFSET);
-        assert(d < 256 + DEPTH_OFFSET);
+        assert(d > DEPTH_ENTRY_OFFSET);
+        assert(d < 256 + DEPTH_ENTRY_OFFSET);
 
         key16     = uint16_t(k);
-        depth8    = uint8_t(d - DEPTH_OFFSET);
+        depth8    = uint8_t(d - DEPTH_ENTRY_OFFSET);
         genBound8 = uint8_t(generation8 | uint8_t(pv) << 2 | b);
         value16   = int16_t(v);
         eval16    = int16_t(ev);
@@ -61,70 +119,117 @@ uint8_t TTEntry::relative_age(const uint8_t generation8) const {
     // is needed to keep the unrelated lowest n bits from affecting
     // the result) to calculate the entry age correctly even after
     // generation8 overflows into the next cycle.
-
-    return (TranspositionTable::GENERATION_CYCLE + generation8 - genBound8)
-         & TranspositionTable::GENERATION_MASK;
+    return (GENERATION_CYCLE + generation8 - genBound8) & GENERATION_MASK;
 }
 
 
+// TTWriter is but a very thin wrapper around the pointer
+TTWriter::TTWriter(TTEntry* tte) :
+    entry(tte) {}
+
+void TTWriter::write(
+  Key k, Value v, bool pv, Bound b, Depth d, Move m, Value ev, uint8_t generation8) {
+    entry->save(k, v, pv, b, d, m, ev, generation8);
+}
+
+
+// A TranspositionTable is an array of Cluster, of size clusterCount. Each cluster consists of ClusterSize number
+// of TTEntry. Each non-empty TTEntry contains information on exactly one position. The size of a Cluster should
+// divide the size of a cache line for best performance, as the cacheline is prefetched when possible.
+
+static constexpr int ClusterSize = 3;
+
+struct Cluster {
+    TTEntry entry[ClusterSize];
+    char    padding[2];  // Pad to 32 bytes
+};
+
+static_assert(sizeof(Cluster) == 32, "Suboptimal Cluster size");
+
+
 // Sets the size of the transposition table,
 // measured in megabytes. Transposition table consists
 // of clusters and each cluster consists of ClusterSize number of TTEntry.
-void TranspositionTable::resize(size_t mbSize, int threadCount) {
+void TranspositionTable::resize(size_t mbSize, ThreadPool& threads) {
     aligned_large_pages_free(table);
 
     clusterCount = mbSize * 1024 * 1024 / sizeof(Cluster);
 
     table = static_cast<Cluster*>(aligned_large_pages_alloc(clusterCount * sizeof(Cluster)));
+
     if (!table)
     {
         std::cerr << "Failed to allocate " << mbSize << "MB for transposition table." << std::endl;
         exit(EXIT_FAILURE);
     }
 
-    clear(threadCount);
+    clear(threads);
 }
 
 
 // Initializes the entire transposition table to zero,
 // in a multi-threaded way.
-void TranspositionTable::clear(size_t threadCount) {
-    std::vector<std::thread> threads;
+void TranspositionTable::clear(ThreadPool& threads) {
+    generation8              = 0;
+    const size_t threadCount = threads.num_threads();
 
-    for (size_t idx = 0; idx < size_t(threadCount); ++idx)
+    for (size_t i = 0; i < threadCount; ++i)
     {
-        threads.emplace_back([this, idx, threadCount]() {
-            // Thread binding gives faster search on systems with a first-touch policy
-            if (threadCount > 8)
-                WinProcGroup::bind_this_thread(idx);
-
+        threads.run_on_thread(i, [this, i, threadCount]() {
             // Each thread will zero its part of the hash table
-            const size_t stride = size_t(clusterCount / threadCount), start = size_t(stride * idx),
-                         len = idx != size_t(threadCount) - 1 ? stride : clusterCount - start;
+            const size_t stride = clusterCount / threadCount;
+            const size_t start  = stride * i;
+            const size_t len    = i + 1 != threadCount ? stride : clusterCount - start;
 
             std::memset(&table[start], 0, len * sizeof(Cluster));
         });
     }
 
-    for (std::thread& th : threads)
-        th.join();
+    for (size_t i = 0; i < threadCount; ++i)
+        threads.wait_on_thread(i);
 }
 
 
+// Returns an approximation of the hashtable
+// occupation during a search. The hash is x permill full, as per UCI protocol.
+// Only counts entries which match the current generation.
+int TranspositionTable::hashfull(int maxAge) const {
+    int maxAgeInternal = maxAge << GENERATION_BITS;
+    int cnt            = 0;
+    for (int i = 0; i < 1000; ++i)
+        for (int j = 0; j < ClusterSize; ++j)
+            cnt += table[i].entry[j].is_occupied()
+                && table[i].entry[j].relative_age(generation8) <= maxAgeInternal;
+
+    return cnt / ClusterSize;
+}
+
+
+void TranspositionTable::new_search() {
+    // increment by delta to keep lower bits as is
+    generation8 += GENERATION_DELTA;
+}
+
+
+uint8_t TranspositionTable::generation() const { return generation8; }
+
+
 // Looks up the current position in the transposition
-// table. It returns true and a pointer to the TTEntry if the position is found.
+// table. It returns true if the position is found.
 // Otherwise, it returns false and a pointer to an empty or least valuable TTEntry
 // to be replaced later. The replace value of an entry is calculated as its depth
 // minus 8 times its relative age. TTEntry t1 is considered more valuable than
 // TTEntry t2 if its replace value is greater than that of t2.
-TTEntry* TranspositionTable::probe(const Key key, bool& found) const {
+std::tuple<bool, TTData, TTWriter> TranspositionTable::probe(const Key key) const {
 
     TTEntry* const tte   = first_entry(key);
     const uint16_t key16 = uint16_t(key);  // Use the low 16 bits as key inside the cluster
 
     for (int i = 0; i < ClusterSize; ++i)
-        if (tte[i].key16 == key16 || !tte[i].depth8)
-            return found = bool(tte[i].depth8), &tte[i];
+        if (tte[i].key16 == key16)
+            // This gap is the main place for read races.
+            // After `read()` completes that copy is final, but may be self-inconsistent.
+            return {tte[i].is_occupied(), tte[i].read(), TTWriter(&tte[i])};
 
     // Find an entry to be replaced according to the replacement strategy
     TTEntry* replace = tte;
@@ -133,22 +238,12 @@ TTEntry* TranspositionTable::probe(const Key key, bool& found) const {
             > tte[i].depth8 - tte[i].relative_age(generation8) * 2)
             replace = &tte[i];
 
-    return found = false, replace;
+    return {false, TTData(), TTWriter(replace)};
 }
 
 
-// Returns an approximation of the hashtable
-// occupation during a search. The hash is x permill full, as per UCI protocol.
-// Only counts entries which match the current generation.
-int TranspositionTable::hashfull() const {
-
-    int cnt = 0;
-    for (int i = 0; i < 1000; ++i)
-        for (int j = 0; j < ClusterSize; ++j)
-            cnt += table[i].entry[j].depth8
-                && (table[i].entry[j].genBound8 & GENERATION_MASK) == generation8;
-
-    return cnt / ClusterSize;
+TTEntry* TranspositionTable::first_entry(const Key key) const {
+    return &table[mul_hi64(key, clusterCount)].entry[0];
 }
 
 }  // namespace Stockfish
diff --git a/src/tt.h b/src/tt.h
index 554a81a5..e7bb5c45 100644
--- a/src/tt.h
+++ b/src/tt.h
@@ -21,99 +21,76 @@
 
 #include <cstddef>
 #include <cstdint>
+#include <tuple>
 
-#include "misc.h"
+#include "memory.h"
 #include "types.h"
 
 namespace Stockfish {
 
-// TTEntry struct is the 10 bytes transposition table entry, defined as below:
+class ThreadPool;
+struct TTEntry;
+struct Cluster;
+
+// There is only one global hash table for the engine and all its threads. For chess in particular, we even allow racy
+// updates between threads to and from the TT, as taking the time to synchronize access would cost thinking time and
+// thus elo. As a hash table, collisions are possible and may cause chess playing issues (bizarre blunders, faulty mate
+// reports, etc). Fixing these also loses elo; however such risk decreases quickly with larger TT size.
 //
-// key        16 bit
-// depth       8 bit
-// generation  5 bit
-// pv node     1 bit
-// bound type  2 bit
-// move       16 bit
-// value      16 bit
-// eval value 16 bit
-struct TTEntry {
+// `probe` is the primary method: given a board position, we lookup its entry in the table, and return a tuple of:
+//   1) whether the entry already has this position
+//   2) a copy of the prior data (if any) (may be inconsistent due to read races)
+//   3) a writer object to this entry
+// The copied data and the writer are separated to maintain clear boundaries between local vs global objects.
 
-    Move  move() const { return Move(move16); }
-    Value value() const { return Value(value16); }
-    Value eval() const { return Value(eval16); }
-    Depth depth() const { return Depth(depth8 + DEPTH_OFFSET); }
-    bool  is_pv() const { return bool(genBound8 & 0x4); }
-    Bound bound() const { return Bound(genBound8 & 0x3); }
-    void  save(Key k, Value v, bool pv, Bound b, Depth d, Move m, Value ev, uint8_t generation8);
-    // The returned age is a multiple of TranspositionTable::GENERATION_DELTA
-    uint8_t relative_age(const uint8_t generation8) const;
 
-   private:
-    friend class TranspositionTable;
-
-    uint16_t key16;
-    uint8_t  depth8;
-    uint8_t  genBound8;
-    Move     move16;
-    int16_t  value16;
-    int16_t  eval16;
+// A copy of the data already in the entry (possibly collided). `probe` may be racy, resulting in inconsistent data.
+struct TTData {
+    Move  move;
+    Value value, eval;
+    Depth depth;
+    Bound bound;
+    bool  is_pv;
+};
+
+
+// This is used to make racy writes to the global TT.
+struct TTWriter {
+   public:
+    void write(Key k, Value v, bool pv, Bound b, Depth d, Move m, Value ev, uint8_t generation8);
+
+   private:
+    friend class TranspositionTable;
+    TTEntry* entry;
+    TTWriter(TTEntry* tte);
 };
 
 
-// A TranspositionTable is an array of Cluster, of size clusterCount. Each
-// cluster consists of ClusterSize number of TTEntry. Each non-empty TTEntry
-// contains information on exactly one position. The size of a Cluster should
-// divide the size of a cache line for best performance, as the cacheline is
-// prefetched when possible.
 class TranspositionTable {
 
-    static constexpr int ClusterSize = 3;
-
-    struct Cluster {
-        TTEntry entry[ClusterSize];
-        char    padding[2];  // Pad to 32 bytes
-    };
-
-    static_assert(sizeof(Cluster) == 32, "Unexpected Cluster size");
-
-    // Constants used to refresh the hash table periodically
-
-    // We have 8 bits available where the lowest 3 bits are
-    // reserved for other things.
-    static constexpr unsigned GENERATION_BITS = 3;
-    // increment for generation field
-    static constexpr int GENERATION_DELTA = (1 << GENERATION_BITS);
-    // cycle length
-    static constexpr int GENERATION_CYCLE = 255 + GENERATION_DELTA;
-    // mask to pull out generation number
-    static constexpr int GENERATION_MASK = (0xFF << GENERATION_BITS) & 0xFF;
-
    public:
     ~TranspositionTable() { aligned_large_pages_free(table); }
 
-    void new_search() {
-        // increment by delta to keep lower bits as is
-        generation8 += GENERATION_DELTA;
-    }
+    void resize(size_t mbSize, ThreadPool& threads);  // Set TT size
+    void clear(ThreadPool& threads);                  // Re-initialize memory, multithreaded
+    int  hashfull(int maxAge = 0)
+      const;  // Approximate what fraction of entries (permille) have been written to during this root search
 
-    TTEntry* probe(const Key key, bool& found) const;
-    int      hashfull() const;
-    void     resize(size_t mbSize, int threadCount);
-    void     clear(size_t threadCount);
-
-    TTEntry* first_entry(const Key key) const {
-        return &table[mul_hi64(key, clusterCount)].entry[0];
-    }
-
-    uint8_t generation() const { return generation8; }
+    void
+    new_search();  // This must be called at the beginning of each root search to track entry aging
+    uint8_t generation() const;  // The current age, used when writing new data to the TT
+    std::tuple<bool, TTData, TTWriter>
+    probe(const Key key) const;  // The main method, whose retvals separate local vs global objects
+    TTEntry* first_entry(const Key key)
+      const;  // This is the hash function; its only external use is memory prefetching.
 
    private:
     friend struct TTEntry;
 
     size_t   clusterCount;
-    Cluster* table       = nullptr;
-    uint8_t  generation8 = 0;  // Size must be not bigger than TTEntry::genBound8
+    Cluster* table = nullptr;
+
+    uint8_t generation8 = 0;  // Size must be not bigger than TTEntry::genBound8
 };
 
 }  // namespace Stockfish
diff --git a/src/tune.cpp b/src/tune.cpp
index 3e5ebe5e..dfcd3468 100644
--- a/src/tune.cpp
+++ b/src/tune.cpp
@@ -21,6 +21,7 @@
 #include <algorithm>
 #include <iostream>
 #include <map>
+#include <optional>
 #include <sstream>
 #include <string>
 
@@ -33,19 +34,19 @@ namespace Stockfish {
 bool          Tune::update_on_last;
 const Option* LastOption = nullptr;
 OptionsMap*   Tune::options;
-
-
 namespace {
 std::map<std::string, int> TuneResults;
 
-void on_tune(const Option& o) {
+std::optional<std::string> on_tune(const Option& o) {
 
     if (!Tune::update_on_last || LastOption == &o)
         Tune::read_options();
+
+    return std::nullopt;
+}
 }
 
-
-void make_option(OptionsMap* options, const string& n, int v, const SetRange& r) {
+void Tune::make_option(OptionsMap* opts, const string& n, int v, const SetRange& r) {
 
     // Do not generate option when there is nothing to tune (ie. min = max)
     if (r(v).first == r(v).second)
@@ -54,15 +55,17 @@ void make_option(OptionsMap* options, const string& n, int v, const SetRange& r)
     if (TuneResults.count(n))
         v = TuneResults[n];
 
-    (*options)[n] << Option(v, r(v).first, r(v).second, on_tune);
-    LastOption = &((*options)[n]);
+    (*opts)[n] << Option(v, r(v).first, r(v).second, on_tune);
+    LastOption = &((*opts)[n]);
 
     // Print formatted parameters, ready to be copy-pasted in Fishtest
-    std::cout << n << "," << v << "," << r(v).first << "," << r(v).second << ","
-              << (r(v).second - r(v).first) / 20.0 << ","
+    std::cout << n << ","                                  //
+              << v << ","                                  //
+              << r(v).first << ","                         //
+              << r(v).second << ","                        //
+              << (r(v).second - r(v).first) / 20.0 << ","  //
               << "0.0020" << std::endl;
 }
-}
 
 string Tune::next(string& names, bool pop) {
 
@@ -118,7 +121,6 @@ void Tune::Entry<Tune::PostUpdate>::read_option() {
 
 namespace Stockfish {
 
-void Tune::read_results() { /* ...insert your values here... */
-}
+void Tune::read_results() { /* ...insert your values here... */ }
 
 }  // namespace Stockfish
diff --git a/src/tune.h b/src/tune.h
index 079614db..ed4738cd 100644
--- a/src/tune.h
+++ b/src/tune.h
@@ -145,6 +145,8 @@ class Tune {
         return add(value, (next(names), std::move(names)), args...);
     }
 
+    static void make_option(OptionsMap* options, const std::string& n, int v, const SetRange& r);
+
     std::vector<std::unique_ptr<EntryBase>> list;
 
    public:
diff --git a/src/types.h b/src/types.h
index 7e523e25..b0bb1c07 100644
--- a/src/types.h
+++ b/src/types.h
@@ -137,9 +137,9 @@ enum Bound {
     BOUND_EXACT = BOUND_UPPER | BOUND_LOWER
 };
 
-// Value is used as an alias for int16_t, this is done to differentiate between
-// a search value and any other integer value. The values used in search are always
-// supposed to be in the range (-VALUE_NONE, VALUE_NONE] and should not exceed this range.
+// Value is used as an alias for int, this is done to differentiate between a search
+// value and any other integer value. The values used in search are always supposed
+// to be in the range (-VALUE_NONE, VALUE_NONE] and should not exceed this range.
 using Value = int;
 
 constexpr Value VALUE_ZERO     = 0;
@@ -155,6 +155,21 @@ constexpr Value VALUE_TB                 = VALUE_MATE_IN_MAX_PLY - 1;
 constexpr Value VALUE_TB_WIN_IN_MAX_PLY  = VALUE_TB - MAX_PLY;
 constexpr Value VALUE_TB_LOSS_IN_MAX_PLY = -VALUE_TB_WIN_IN_MAX_PLY;
 
+
+constexpr bool is_valid(Value value) { return value != VALUE_NONE; }
+
+constexpr bool is_win(Value value) {
+    assert(is_valid(value));
+    return value >= VALUE_TB_WIN_IN_MAX_PLY;
+}
+
+constexpr bool is_loss(Value value) {
+    assert(is_valid(value));
+    return value <= VALUE_TB_LOSS_IN_MAX_PLY;
+}
+
+constexpr bool is_decisive(Value value) { return is_win(value) || is_loss(value); }
+
 // In the code, we make the assumption that these values
 // are such that non_pawn_material() can be used to uniquely
 // identify the material on the board.
@@ -187,12 +202,21 @@ constexpr Value PieceValue[PIECE_NB] = {
 using Depth = int;
 
 enum : int {
-    DEPTH_QS_CHECKS    = 0,
-    DEPTH_QS_NO_CHECKS = -1,
-
-    DEPTH_NONE = -6,
-
-    DEPTH_OFFSET = -7  // value used only for TT entry occupancy check
+    // The following DEPTH_ constants are used for transposition table entries
+    // and quiescence search move generation stages. In regular search, the
+    // depth stored in the transposition table is literal: the search depth
+    // (effort) used to make the corresponding transposition table value. In
+    // quiescence search, however, the transposition table entries only store
+    // the current quiescence move generation stage (which should thus compare
+    // lower than any regular search depth).
+    DEPTH_QS = 0,
+    // For transposition table entries where no searching at all was done
+    // (whether regular or qsearch) we use DEPTH_UNSEARCHED, which should thus
+    // compare lower than any quiescence or regular depth. DEPTH_ENTRY_OFFSET
+    // is used only for the transposition table entry occupancy check (see tt.cpp),
+    // and should thus be lower than DEPTH_UNSEARCHED.
+    DEPTH_UNSEARCHED   = -2,
+    DEPTH_ENTRY_OFFSET = -3
 };
 
 // clang-format off
@@ -357,9 +381,10 @@ enum MoveType {
 // bit 14-15: special move flag: promotion (1), en passant (2), castling (3)
 // NOTE: en passant bit is set only when a pawn can be captured
 //
-// Special cases are Move::none() and Move::null(). We can sneak these in because in
-// any normal move destination square is always different from origin square
-// while Move::none() and Move::null() have the same origin and destination square.
+// Special cases are Move::none() and Move::null(). We can sneak these in because
+// in any normal move the destination square and origin square are always different,
+// but Move::none() and Move::null() have the same origin and destination square.
+
 class Move {
    public:
     Move() = default;
diff --git a/src/uci.cpp b/src/uci.cpp
index cb686a02..8388cad8 100644
--- a/src/uci.cpp
+++ b/src/uci.cpp
@@ -22,6 +22,7 @@
 #include <cctype>
 #include <cmath>
 #include <cstdint>
+#include <iterator>
 #include <optional>
 #include <sstream>
 #include <string_view>
@@ -30,20 +31,19 @@
 
 #include "benchmark.h"
 #include "engine.h"
-#include "evaluate.h"
+#include "memory.h"
 #include "movegen.h"
 #include "position.h"
 #include "score.h"
 #include "search.h"
-#include "syzygy/tbprobe.h"
 #include "types.h"
 #include "ucioption.h"
 
 namespace Stockfish {
 
-constexpr auto StartFEN  = "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1";
-constexpr int  MaxHashMB = Is64Bit ? 33554432 : 2048;
+constexpr auto BenchmarkCommand = "speedtest";
 
+constexpr auto StartFEN = "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1";
 template<typename... Ts>
 struct overload: Ts... {
     using Ts::operator()...;
@@ -52,50 +52,40 @@ struct overload: Ts... {
 template<typename... Ts>
 overload(Ts...) -> overload<Ts...>;
 
+void UCIEngine::print_info_string(std::string_view str) {
+    sync_cout_start();
+    for (auto& line : split(str, "\n"))
+    {
+        if (!is_whitespace(line))
+        {
+            std::cout << "info string " << line << '\n';
+        }
+    }
+    sync_cout_end();
+}
+
 UCIEngine::UCIEngine(int argc, char** argv) :
     engine(argv[0]),
     cli(argc, argv) {
 
-    auto& options = engine.get_options();
-
-    options["Debug Log File"] << Option("", [](const Option& o) { start_logger(o); });
-
-    options["Threads"] << Option(1, 1, 1024, [this](const Option&) { engine.resize_threads(); });
-
-    options["Hash"] << Option(16, 1, MaxHashMB, [this](const Option& o) { engine.set_tt_size(o); });
-
-    options["Clear Hash"] << Option([this](const Option&) { engine.search_clear(); });
-    options["Ponder"] << Option(false);
-    options["MultiPV"] << Option(1, 1, MAX_MOVES);
-    options["Skill Level"] << Option(20, 0, 20);
-    options["Move Overhead"] << Option(10, 0, 5000);
-    options["nodestime"] << Option(0, 0, 10000);
-    options["UCI_Chess960"] << Option(false);
-    options["UCI_LimitStrength"] << Option(false);
-    options["UCI_Elo"] << Option(1320, 1320, 3190);
-    options["UCI_ShowWDL"] << Option(false);
-    options["SyzygyPath"] << Option("<empty>", [](const Option& o) { Tablebases::init(o); });
-    options["SyzygyProbeDepth"] << Option(1, 1, 100);
-    options["Syzygy50MoveRule"] << Option(true);
-    options["SyzygyProbeLimit"] << Option(7, 0, 7);
-    options["EvalFile"] << Option(EvalFileDefaultNameBig,
-                                  [this](const Option& o) { engine.load_big_network(o); });
-    options["EvalFileSmall"] << Option(EvalFileDefaultNameSmall,
-                                       [this](const Option& o) { engine.load_small_network(o); });
+    engine.get_options().add_info_listener([](const std::optional<std::string>& str) {
+        if (str.has_value())
+            print_info_string(*str);
+    });
 
+    init_search_update_listeners();
+}
 
+void UCIEngine::init_search_update_listeners() {
     engine.set_on_iter([](const auto& i) { on_iter(i); });
     engine.set_on_update_no_moves([](const auto& i) { on_update_no_moves(i); });
-    engine.set_on_update_full([&](const auto& i) { on_update_full(i, options["UCI_ShowWDL"]); });
+    engine.set_on_update_full(
+      [this](const auto& i) { on_update_full(i, engine.get_options()["UCI_ShowWDL"]); });
     engine.set_on_bestmove([](const auto& bm, const auto& p) { on_bestmove(bm, p); });
-
-    engine.load_networks();
-    engine.resize_threads();
-    engine.search_clear();  // After threads are up
+    engine.set_on_verify_networks([](const auto& s) { print_info_string(s); });
 }
 
 void UCIEngine::loop() {
-
     std::string token, cmd;
 
     for (int i = 1; i < cli.argc; ++i)
@@ -123,13 +113,22 @@ void UCIEngine::loop() {
             engine.set_ponderhit(false);
 
         else if (token == "uci")
+        {
             sync_cout << "id name " << engine_info(true) << "\n"
-                      << engine.get_options() << "\nuciok" << sync_endl;
+                      << engine.get_options() << sync_endl;
+
+            sync_cout << "uciok" << sync_endl;
+        }
 
         else if (token == "setoption")
             setoption(is);
         else if (token == "go")
+        {
+            // send info strings after the go command is sent for old GUIs and python-chess
+            print_info_string(engine.numa_config_information_as_string());
+            print_info_string(engine.thread_allocation_information_as_string());
             go(is);
+        }
         else if (token == "position")
             position(is);
         else if (token == "ucinewgame")
@@ -143,6 +142,8 @@ void UCIEngine::loop() {
             engine.flip();
         else if (token == "bench")
             bench(is);
+        else if (token == BenchmarkCommand)
+            benchmark(is);
         else if (token == "d")
             sync_cout << engine.visualize() << sync_endl;
         else if (token == "eval")
@@ -258,7 +259,7 @@ void UCIEngine::bench(std::istream& args) {
                 Search::LimitsType limits = parse_limits(is);
 
                 if (limits.perft)
-                    nodes = perft(limits);
+                    nodesSearched = perft(limits);
                 else
                 {
                     engine.go(limits);
@@ -286,14 +287,178 @@ void UCIEngine::bench(std::istream& args) {
 
     dbg_print();
 
-    std::cerr << "\n==========================="
-              << "\nTotal time (ms) : " << elapsed << "\nNodes searched  : " << nodes
+    std::cerr << "\n==========================="    //
+              << "\nTotal time (ms) : " << elapsed  //
+              << "\nNodes searched  : " << nodes    //
               << "\nNodes/second    : " << 1000 * nodes / elapsed << std::endl;
 
     // reset callback, to not capture a dangling reference to nodesSearched
     engine.set_on_update_full([&](const auto& i) { on_update_full(i, options["UCI_ShowWDL"]); });
 }
 
+void UCIEngine::benchmark(std::istream& args) {
+    // Probably not very important for a test this long, but include for completeness and sanity.
+    static constexpr int NUM_WARMUP_POSITIONS = 3;
+
+    std::string token;
+    uint64_t    nodes = 0, cnt = 1;
+    uint64_t    nodesSearched = 0;
+
+    engine.set_on_update_full([&](const Engine::InfoFull& i) { nodesSearched = i.nodes; });
+
+    engine.set_on_iter([](const auto&) {});
+    engine.set_on_update_no_moves([](const auto&) {});
+    engine.set_on_bestmove([](const auto&, const auto&) {});
+    engine.set_on_verify_networks([](const auto&) {});
+
+    Benchmark::BenchmarkSetup setup = Benchmark::setup_benchmark(args);
+
+    const int numGoCommands = count_if(setup.commands.begin(), setup.commands.end(),
+                                       [](const std::string& s) { return s.find("go ") == 0; });
+
+    TimePoint totalTime = 0;
+
+    // Set options once at the start.
+    auto ss = std::istringstream("name Threads value " + std::to_string(setup.threads));
+    setoption(ss);
+    ss = std::istringstream("name Hash value " + std::to_string(setup.ttSize));
+    setoption(ss);
+    ss = std::istringstream("name UCI_Chess960 value false");
+    setoption(ss);
+
+    // Warmup
+    for (const auto& cmd : setup.commands)
+    {
+        std::istringstream is(cmd);
+        is >> std::skipws >> token;
+
+        if (token == "go")
+        {
+            // One new line is produced by the search, so omit it here
+            std::cerr << "\rWarmup position " << cnt++ << '/' << NUM_WARMUP_POSITIONS;
+
+            Search::LimitsType limits = parse_limits(is);
+
+            TimePoint elapsed = now();
+
+            // Run with silenced network verification
+            engine.go(limits);
+            engine.wait_for_search_finished();
+
+            totalTime += now() - elapsed;
+
+            nodes += nodesSearched;
+            nodesSearched = 0;
+        }
+        else if (token == "position")
+            position(is);
+        else if (token == "ucinewgame")
+        {
+            engine.search_clear();  // search_clear may take a while
+        }
+
+        if (cnt > NUM_WARMUP_POSITIONS)
+            break;
+    }
+
+    std::cerr << "\n";
+
+    cnt   = 1;
+    nodes = 0;
+
+    int           numHashfullReadings = 0;
+    constexpr int hashfullAges[]      = {0, 999};  // Only normal hashfull and touched hash.
+    int           totalHashfull[std::size(hashfullAges)] = {0};
+    int           maxHashfull[std::size(hashfullAges)]   = {0};
+
+    auto updateHashfullReadings = [&]() {
+        numHashfullReadings += 1;
+
+        for (int i = 0; i < static_cast<int>(std::size(hashfullAges)); ++i)
+        {
+            const int hashfull = engine.get_hashfull(hashfullAges[i]);
+            maxHashfull[i]     = std::max(maxHashfull[i], hashfull);
+            totalHashfull[i] += hashfull;
+        }
+    };
+
+    engine.search_clear();  // search_clear may take a while
+
+    for (const auto& cmd : setup.commands)
+    {
+        std::istringstream is(cmd);
+        is >> std::skipws >> token;
+
+        if (token == "go")
+        {
+            // One new line is produced by the search, so omit it here
+            std::cerr << "\rPosition " << cnt++ << '/' << numGoCommands;
+
+            Search::LimitsType limits = parse_limits(is);
+
+            TimePoint elapsed = now();
+
+            // Run with silenced network verification
+            engine.go(limits);
+            engine.wait_for_search_finished();
+
+            totalTime += now() - elapsed;
+
+            updateHashfullReadings();
+
+            nodes += nodesSearched;
+            nodesSearched = 0;
+        }
+        else if (token == "position")
+            position(is);
+        else if (token == "ucinewgame")
+        {
+            engine.search_clear();  // search_clear may take a while
+        }
+    }
+
+    totalTime = std::max<TimePoint>(totalTime, 1);  // Ensure positivity to avoid a 'divide by zero'
+
+    dbg_print();
+
+    std::cerr << "\n";
+
+    static_assert(
+      std::size(hashfullAges) == 2 && hashfullAges[0] == 0 && hashfullAges[1] == 999,
+      "Hardcoded for display. Would complicate the code needlessly in the current state.");
+
+    std::string threadBinding = engine.thread_binding_information_as_string();
+    if (threadBinding.empty())
+        threadBinding = "none";
+
+    // clang-format off
+
+    std::cerr << "==========================="
+              << "\nVersion                    : "
+              << engine_version_info()
+              // "\nCompiled by                : "
+              << compiler_info()
+              << "Large pages                : " << (has_large_pages() ? "yes" : "no")
+              << "\nUser invocation            : " << BenchmarkCommand << " "
+              << setup.originalInvocation << "\nFilled invocation          : " << BenchmarkCommand
+              << " " << setup.filledInvocation
+              << "\nAvailable processors       : " << engine.get_numa_config_as_string()
+              << "\nThread count               : " << setup.threads
+              << "\nThread binding             : " << threadBinding
+              << "\nTT size [MiB]              : " << setup.ttSize
+              << "\nHash max, avg [per mille]  : "
+              << "\n    single search          : " << maxHashfull[0] << ", "
+              << totalHashfull[0] / numHashfullReadings
+              << "\n    single game            : " << maxHashfull[1] << ", "
+              << totalHashfull[1] / numHashfullReadings
+              << "\nTotal nodes searched       : " << nodes
+              << "\nTotal search time [s]      : " << totalTime / 1000.0
+              << "\nNodes/second               : " << 1000 * nodes / totalTime << std::endl;
+
+    // clang-format on
+
+    init_search_update_listeners();
+}
 
 void UCIEngine::setoption(std::istringstream& is) {
     engine.wait_for_search_finished();
@@ -344,12 +509,12 @@ WinRateParams win_rate_params(const Position& pos) {
     int material = pos.count<PAWN>() + 3 * pos.count<KNIGHT>() + 3 * pos.count<BISHOP>()
                  + 5 * pos.count<ROOK>() + 9 * pos.count<QUEEN>();
 
-    // The fitted model only uses data for material counts in [10, 78], and is anchored at count 58.
-    double m = std::clamp(material, 10, 78) / 58.0;
+    // The fitted model only uses data for material counts in [17, 78], and is anchored at count 58.
+    double m = std::clamp(material, 17, 78) / 58.0;
 
     // Return a = p_a(material) and b = p_b(material), see github.com/official-stockfish/WDL_model
-    constexpr double as[] = {-150.77043883, 394.96159472, -321.73403766, 406.15850091};
-    constexpr double bs[] = {62.33245393, -91.02264855, 45.88486850, 51.63461272};
+    constexpr double as[] = {-37.45051876, 121.19101539, -132.78783573, 420.70576692};
+    constexpr double bs[] = {90.26261072, -137.26549898, 71.10130540, 51.35259597};
 
     double a = (((as[0] * m + as[1]) * m + as[2]) * m) + as[3];
     double b = (((bs[0] * m + bs[1]) * m + bs[2]) * m) + bs[3];
@@ -390,8 +555,8 @@ std::string UCIEngine::format_score(const Score& s) {
 // without treatment of mate and similar special scores.
 int UCIEngine::to_cp(Value v, const Position& pos) {
 
-    // In general, the score can be defined via the the WDL as
-    // (log(1/L - 1) - log(1/W - 1)) / ((log(1/L - 1) + log(1/W - 1))
+    // In general, the score can be defined via the WDL as
+    // (log(1/L - 1) - log(1/W - 1)) / (log(1/L - 1) + log(1/W - 1)).
     // Based on our win_rate_model, this simply yields v / a.
 
     auto [a, b] = win_rate_params(pos);
diff --git a/src/uci.h b/src/uci.h
index 55d580f9..6adf74cb 100644
--- a/src/uci.h
+++ b/src/uci.h
@@ -19,10 +19,10 @@
 #ifndef UCI_H_INCLUDED
 #define UCI_H_INCLUDED
 
+#include <cstdint>
 #include <iostream>
 #include <string>
 #include <string_view>
-#include <cstdint>
 
 #include "engine.h"
 #include "misc.h"
@@ -58,8 +58,11 @@ class UCIEngine {
     Engine      engine;
     CommandLine cli;
 
+    static void print_info_string(std::string_view str);
+
     void          go(std::istringstream& is);
     void          bench(std::istream& args);
+    void          benchmark(std::istream& args);
     void          position(std::istringstream& is);
     void          setoption(std::istringstream& is);
     std::uint64_t perft(const Search::LimitsType&);
@@ -68,6 +71,8 @@ class UCIEngine {
     static void on_update_full(const Engine::InfoFull& info, bool showWDL);
     static void on_iter(const Engine::InfoIter& info);
     static void on_bestmove(std::string_view bestmove, std::string_view ponder);
+
+    void init_search_update_listeners();
 };
 
 }  // namespace Stockfish
diff --git a/src/ucioption.cpp b/src/ucioption.cpp
index e1ffe546..455803cf 100644
--- a/src/ucioption.cpp
+++ b/src/ucioption.cpp
@@ -36,6 +36,8 @@ bool CaseInsensitiveLess::operator()(const std::string& s1, const std::string& s
       [](char c1, char c2) { return std::tolower(c1) < std::tolower(c2); });
 }
 
+void OptionsMap::add_info_listener(InfoListener&& message_func) { info = std::move(message_func); }
+
 void OptionsMap::setoption(std::istringstream& is) {
     std::string token, name, value;
 
@@ -57,13 +59,20 @@ void OptionsMap::setoption(std::istringstream& is) {
 
 Option OptionsMap::operator[](const std::string& name) const {
     auto it = options_map.find(name);
-    return it != options_map.end() ? it->second : Option();
+    return it != options_map.end() ? it->second : Option(this);
 }
 
-Option& OptionsMap::operator[](const std::string& name) { return options_map[name]; }
+Option& OptionsMap::operator[](const std::string& name) {
+    if (!options_map.count(name))
+        options_map[name] = Option(this);
+    return options_map[name];
+}
 
 std::size_t OptionsMap::count(const std::string& name) const { return options_map.count(name); }
 
+Option::Option(const OptionsMap* map) :
+    parent(map) {}
+
 Option::Option(const char* v, OnChange f) :
     type("string"),
     min(0),
@@ -118,6 +127,8 @@ bool Option::operator==(const char* s) const {
     return !CaseInsensitiveLess()(currentValue, s) && !CaseInsensitiveLess()(s, currentValue);
 }
 
+bool Option::operator!=(const char* s) const { return !(*this == s); }
+
 
 // Inits options and assigns idx in the correct printing order
 
@@ -125,10 +136,12 @@ void Option::operator<<(const Option& o) {
 
     static size_t insert_order = 0;
 
-    *this = o;
-    idx   = insert_order++;
-}
+    auto p = this->parent;
+    *this  = o;
 
+    this->parent = p;
+    idx          = insert_order++;
+}
 
 // Updates currentValue and triggers on_change() action. It's up to
 // the GUI to check for option's limits, but we could receive the new value
@@ -153,11 +166,18 @@ Option& Option::operator=(const std::string& v) {
             return *this;
     }
 
-    if (type != "button")
+    if (type == "string")
+        currentValue = v == "<empty>" ? "" : v;
+    else if (type != "button")
         currentValue = v;
 
     if (on_change)
-        on_change(*this);
+    {
+        const auto ret = on_change(*this);
+
+        if (ret && parent != nullptr && parent->info != nullptr)
+            parent->info(ret);
+    }
 
     return *this;
 }
@@ -170,10 +190,16 @@ std::ostream& operator<<(std::ostream& os, const OptionsMap& om) {
                 const Option& o = it.second;
                 os << "\noption name " << it.first << " type " << o.type;
 
-                if (o.type == "string" || o.type == "check" || o.type == "combo")
+                if (o.type == "check" || o.type == "combo")
                     os << " default " << o.defaultValue;
 
-                if (o.type == "spin")
+                else if (o.type == "string")
+                {
+                    std::string defaultValue = o.defaultValue.empty() ? "<empty>" : o.defaultValue;
+                    os << " default " << defaultValue;
+                }
+
+                else if (o.type == "spin")
                     os << " default " << int(stof(o.defaultValue)) << " min " << o.min << " max "
                        << o.max;
 
diff --git a/src/ucioption.h b/src/ucioption.h
index b575d164..a47cc98d 100644
--- a/src/ucioption.h
+++ b/src/ucioption.h
@@ -23,6 +23,7 @@
 #include <functional>
 #include <iosfwd>
 #include <map>
+#include <optional>
 #include <string>
 
 namespace Stockfish {
@@ -31,31 +32,14 @@ struct CaseInsensitiveLess {
     bool operator()(const std::string&, const std::string&) const;
 };
 
-class Option;
-
-class OptionsMap {
-   public:
-    void setoption(std::istringstream&);
-
-    friend std::ostream& operator<<(std::ostream&, const OptionsMap&);
-
-    Option  operator[](const std::string&) const;
-    Option& operator[](const std::string&);
-
-    std::size_t count(const std::string&) const;
-
-   private:
-    // The options container is defined as a std::map
-    using OptionsStore = std::map<std::string, Option, CaseInsensitiveLess>;
-
-    OptionsStore options_map;
-};
+class OptionsMap;
 
 // The Option class implements each option as specified by the UCI protocol
 class Option {
    public:
-    using OnChange = std::function<void(const Option&)>;
+    using OnChange = std::function<std::optional<std::string>(const Option&)>;
 
+    Option(const OptionsMap*);
     Option(OnChange = nullptr);
     Option(bool v, OnChange = nullptr);
     Option(const char* v, OnChange = nullptr);
@@ -63,18 +47,57 @@ class Option {
     Option(const char* v, const char* cur, OnChange = nullptr);
 
     Option& operator=(const std::string&);
-    void    operator<<(const Option&);
     operator int() const;
     operator std::string() const;
     bool operator==(const char*) const;
+    bool operator!=(const char*) const;
 
     friend std::ostream& operator<<(std::ostream&, const OptionsMap&);
 
    private:
-    std::string defaultValue, currentValue, type;
-    int         min, max;
-    size_t      idx;
-    OnChange    on_change;
+    friend class OptionsMap;
+    friend class Engine;
+    friend class Tune;
+
+    void operator<<(const Option&);
+
+    std::string       defaultValue, currentValue, type;
+    int               min, max;
+    size_t            idx;
+    OnChange          on_change;
+    const OptionsMap* parent = nullptr;
+};
+
+class OptionsMap {
+   public:
+    using InfoListener = std::function<void(std::optional<std::string>)>;
+
+    OptionsMap()                             = default;
+    OptionsMap(const OptionsMap&)            = delete;
+    OptionsMap(OptionsMap&&)                 = delete;
+    OptionsMap& operator=(const OptionsMap&) = delete;
+    OptionsMap& operator=(OptionsMap&&)      = delete;
+
+    void add_info_listener(InfoListener&&);
+
+    void setoption(std::istringstream&);
+
+    Option  operator[](const std::string&) const;
+    Option& operator[](const std::string&);
+
+    std::size_t count(const std::string&) const;
+
+   private:
+    friend class Engine;
+    friend class Option;
+
+    friend std::ostream& operator<<(std::ostream&, const OptionsMap&);
+
+    // The options container is defined as a std::map
+    using OptionsStore = std::map<std::string, Option, CaseInsensitiveLess>;
+
+    OptionsStore options_map;
+    InfoListener info;
 };
 
 }
diff --git a/tests/instrumented.py b/tests/instrumented.py
new file mode 100644
index 00000000..db5ec8e0
--- /dev/null
+++ b/tests/instrumented.py
@@ -0,0 +1,520 @@
+import argparse
+import re
+import sys
+import subprocess
+import pathlib
+import os
+
+from testing import (
+    EPD,
+    TSAN,
+    Stockfish as Engine,
+    MiniTestFramework,
+    OrderedClassMembers,
+    Valgrind,
+    Syzygy,
+)
+
+PATH = pathlib.Path(__file__).parent.resolve()
+CWD = os.getcwd()
+
+
+def get_prefix():
+    if args.valgrind:
+        return Valgrind.get_valgrind_command()
+    if args.valgrind_thread:
+        return Valgrind.get_valgrind_thread_command()
+
+    return []
+
+
+def get_threads():
+    if args.valgrind_thread or args.sanitizer_thread:
+        return 2
+    return 1
+
+
+def get_path():
+    return os.path.abspath(os.path.join(CWD, args.stockfish_path))
+
+
+def postfix_check(output):
+    if args.sanitizer_undefined:
+        for idx, line in enumerate(output):
+            if "runtime error:" in line:
+                # print next possible 50 lines
+                for i in range(50):
+                    debug_idx = idx + i
+                    if debug_idx < len(output):
+                        print(output[debug_idx])
+                return False
+
+    if args.sanitizer_thread:
+        for idx, line in enumerate(output):
+            if "WARNING: ThreadSanitizer:" in line:
+                # print next possible 50 lines
+                for i in range(50):
+                    debug_idx = idx + i
+                    if debug_idx < len(output):
+                        print(output[debug_idx])
+                return False
+
+    return True
+
+
+def Stockfish(*args, **kwargs):
+    return Engine(get_prefix(), get_path(), *args, **kwargs)
+
+
+class TestCLI(metaclass=OrderedClassMembers):
+
+    def beforeAll(self):
+        pass
+
+    def afterAll(self):
+        pass
+
+    def beforeEach(self):
+        self.stockfish = None
+
+    def afterEach(self):
+        assert postfix_check(self.stockfish.get_output()) == True
+        self.stockfish.clear_output()
+
+    def test_eval(self):
+        self.stockfish = Stockfish("eval".split(" "), True)
+        assert self.stockfish.process.returncode == 0
+
+    def test_go_nodes_1000(self):
+        self.stockfish = Stockfish("go nodes 1000".split(" "), True)
+        assert self.stockfish.process.returncode == 0
+
+    def test_go_depth_10(self):
+        self.stockfish = Stockfish("go depth 10".split(" "), True)
+        assert self.stockfish.process.returncode == 0
+
+    def test_go_perft_4(self):
+        self.stockfish = Stockfish("go perft 4".split(" "), True)
+        assert self.stockfish.process.returncode == 0
+
+    def test_go_movetime_1000(self):
+        self.stockfish = Stockfish("go movetime 1000".split(" "), True)
+        assert self.stockfish.process.returncode == 0
+
+    def test_go_wtime_8000_btime_8000_winc_500_binc_500(self):
+        self.stockfish = Stockfish(
+            "go wtime 8000 btime 8000 winc 500 binc 500".split(" "),
+            True,
+        )
+        assert self.stockfish.process.returncode == 0
+
+    def test_go_wtime_1000_btime_1000_winc_0_binc_0(self):
+        self.stockfish = Stockfish(
+            "go wtime 1000 btime 1000 winc 0 binc 0".split(" "),
+            True,
+        )
+        assert self.stockfish.process.returncode == 0
+
+    def test_go_wtime_1000_btime_1000_winc_0_binc_0_movestogo_5(self):
+        self.stockfish = Stockfish(
+            "go wtime 1000 btime 1000 winc 0 binc 0 movestogo 5".split(" "),
+            True,
+        )
+        assert self.stockfish.process.returncode == 0
+
+    def test_go_movetime_200(self):
+        self.stockfish = Stockfish("go movetime 200".split(" "), True)
+        assert self.stockfish.process.returncode == 0
+
+    def test_go_nodes_20000_searchmoves_e2e4_d2d4(self):
+        self.stockfish = Stockfish(
+            "go nodes 20000 searchmoves e2e4 d2d4".split(" "), True
+        )
+        assert self.stockfish.process.returncode == 0
+
+    def test_bench_128_threads_8_default_depth(self):
+        self.stockfish = Stockfish(
+            f"bench 128 {get_threads()} 8 default depth".split(" "),
+            True,
+        )
+        assert self.stockfish.process.returncode == 0
+
+    def test_bench_128_threads_3_bench_tmp_epd_depth(self):
+        self.stockfish = Stockfish(
+            f"bench 128 {get_threads()} 3 {os.path.join(PATH,'bench_tmp.epd')} depth".split(
+                " "
+            ),
+            True,
+        )
+        assert self.stockfish.process.returncode == 0
+
+    def test_d(self):
+        self.stockfish = Stockfish("d".split(" "), True)
+        assert self.stockfish.process.returncode == 0
+
+    def test_compiler(self):
+        self.stockfish = Stockfish("compiler".split(" "), True)
+        assert self.stockfish.process.returncode == 0
+
+    def test_license(self):
+        self.stockfish = Stockfish("license".split(" "), True)
+        assert self.stockfish.process.returncode == 0
+
+    def test_uci(self):
+        self.stockfish = Stockfish("uci".split(" "), True)
+        assert self.stockfish.process.returncode == 0
+
+    def test_export_net_verify_nnue(self):
+        current_path = os.path.abspath(os.getcwd())
+        self.stockfish = Stockfish(
+            f"export_net {os.path.join(current_path , 'verify.nnue')}".split(" "), True
+        )
+        assert self.stockfish.process.returncode == 0
+
+    # verify the generated net equals the base net
+
+    def test_network_equals_base(self):
+        self.stockfish = Stockfish(
+            ["uci"],
+            True,
+        )
+
+        output = self.stockfish.process.stdout
+
+        # find line
+        for line in output.split("\n"):
+            if "option name EvalFile type string default" in line:
+                network = line.split(" ")[-1]
+                break
+
+        # find network file in src dir
+        network = os.path.join(PATH.parent.resolve(), "src", network)
+
+        if not os.path.exists(network):
+            print(
+                f"Network file {network} not found, please download the network file over the make command."
+            )
+            assert False
+
+        diff = subprocess.run(["diff", network, f"verify.nnue"])
+
+        assert diff.returncode == 0
+
+
+class TestInteractive(metaclass=OrderedClassMembers):
+    def beforeAll(self):
+        self.stockfish = Stockfish()
+
+    def afterAll(self):
+        self.stockfish.quit()
+        assert self.stockfish.close() == 0
+
+    def afterEach(self):
+        assert postfix_check(self.stockfish.get_output()) == True
+        self.stockfish.clear_output()
+
+    def test_startup_output(self):
+        self.stockfish.starts_with("Stockfish")
+
+    def test_uci_command(self):
+        self.stockfish.send_command("uci")
+        self.stockfish.equals("uciok")
+
+    def test_set_threads_option(self):
+        self.stockfish.send_command(f"setoption name Threads value {get_threads()}")
+
+    def test_ucinewgame_and_startpos_nodes_1000(self):
+        self.stockfish.send_command("ucinewgame")
+        self.stockfish.send_command("position startpos")
+        self.stockfish.send_command("go nodes 1000")
+        self.stockfish.starts_with("bestmove")
+
+    def test_ucinewgame_and_startpos_moves(self):
+        self.stockfish.send_command("ucinewgame")
+        self.stockfish.send_command("position startpos moves e2e4 e7e6")
+        self.stockfish.send_command("go nodes 1000")
+        self.stockfish.starts_with("bestmove")
+
+    def test_fen_position_1(self):
+        self.stockfish.send_command("ucinewgame")
+        self.stockfish.send_command("position fen 5rk1/1K4p1/8/8/3B4/8/8/8 b - - 0 1")
+        self.stockfish.send_command("go nodes 1000")
+        self.stockfish.starts_with("bestmove")
+
+    def test_fen_position_2_flip(self):
+        self.stockfish.send_command("ucinewgame")
+        self.stockfish.send_command("position fen 5rk1/1K4p1/8/8/3B4/8/8/8 b - - 0 1")
+        self.stockfish.send_command("flip")
+        self.stockfish.send_command("go nodes 1000")
+        self.stockfish.starts_with("bestmove")
+
+    def test_depth_5_with_callback(self):
+        self.stockfish.send_command("ucinewgame")
+        self.stockfish.send_command("position startpos")
+        self.stockfish.send_command("go depth 5")
+
+        def callback(output):
+            regex = r"info depth \d+ seldepth \d+ multipv \d+ score cp \d+ nodes \d+ nps \d+ hashfull \d+ tbhits \d+ time \d+ pv"
+            if output.startswith("info depth") and not re.match(regex, output):
+                assert False
+            if output.startswith("bestmove"):
+                return True
+            return False
+
+        self.stockfish.check_output(callback)
+
+    def test_ucinewgame_and_go_depth_9(self):
+        self.stockfish.send_command("ucinewgame")
+        self.stockfish.send_command("setoption name UCI_ShowWDL value true")
+        self.stockfish.send_command("position startpos")
+        self.stockfish.send_command("go depth 9")
+
+        depth = 1
+
+        def callback(output):
+            nonlocal depth
+
+            regex = rf"info depth {depth} seldepth \d+ multipv \d+ score cp \d+ wdl \d+ \d+ \d+ nodes \d+ nps \d+ hashfull \d+ tbhits \d+ time \d+ pv"
+
+            if output.startswith("info depth"):
+                if not re.match(regex, output):
+                    assert False
+                depth += 1
+
+            if output.startswith("bestmove"):
+                assert depth == 10
+                return True
+
+            return False
+
+        self.stockfish.check_output(callback)
+
+    def test_clear_hash(self):
+        self.stockfish.send_command("setoption name Clear Hash")
+
+    def test_fen_position_mate_1(self):
+        self.stockfish.send_command("ucinewgame")
+        self.stockfish.send_command(
+            "position fen 5K2/8/2qk4/2nPp3/3r4/6B1/B7/3R4 w - e6"
+        )
+        self.stockfish.send_command("go depth 18")
+
+        self.stockfish.expect("* score mate 1 * pv d5e6")
+        self.stockfish.equals("bestmove d5e6")
+
+    def test_fen_position_mate_minus_1(self):
+        self.stockfish.send_command("ucinewgame")
+        self.stockfish.send_command(
+            "position fen 2brrb2/8/p7/Q7/1p1kpPp1/1P1pN1K1/3P4/8 b - -"
+        )
+        self.stockfish.send_command("go depth 18")
+        self.stockfish.expect("* score mate -1 *")
+        self.stockfish.starts_with("bestmove")
+
+    def test_fen_position_fixed_node(self):
+        self.stockfish.send_command("ucinewgame")
+        self.stockfish.send_command(
+            "position fen 5K2/8/2P1P1Pk/6pP/3p2P1/1P6/3P4/8 w - - 0 1"
+        )
+        self.stockfish.send_command("go nodes 500000")
+        self.stockfish.starts_with("bestmove")
+
+    def test_fen_position_with_mate_go_depth(self):
+        self.stockfish.send_command("ucinewgame")
+        self.stockfish.send_command(
+            "position fen 8/5R2/2K1P3/4k3/8/b1PPpp1B/5p2/8 w - -"
+        )
+        self.stockfish.send_command("go depth 18 searchmoves c6d7")
+        self.stockfish.expect("* score mate 2 * pv c6d7 * f7f5")
+
+        self.stockfish.starts_with("bestmove")
+
+    def test_fen_position_with_mate_go_mate(self):
+        self.stockfish.send_command("ucinewgame")
+        self.stockfish.send_command(
+            "position fen 8/5R2/2K1P3/4k3/8/b1PPpp1B/5p2/8 w - -"
+        )
+        self.stockfish.send_command("go mate 2 searchmoves c6d7")
+        self.stockfish.expect("* score mate 2 * pv c6d7 *")
+
+        self.stockfish.starts_with("bestmove")
+
+    def test_fen_position_with_mate_go_nodes(self):
+        self.stockfish.send_command("ucinewgame")
+        self.stockfish.send_command(
+            "position fen 8/5R2/2K1P3/4k3/8/b1PPpp1B/5p2/8 w - -"
+        )
+        self.stockfish.send_command("go nodes 500000 searchmoves c6d7")
+        self.stockfish.expect("* score mate 2 * pv c6d7 * f7f5")
+
+        self.stockfish.starts_with("bestmove")
+
+    def test_fen_position_depth_27(self):
+        self.stockfish.send_command("ucinewgame")
+        self.stockfish.send_command(
+            "position fen r1b2r1k/pp1p2pp/2p5/2B1q3/8/8/P1PN2PP/R4RK1 w - - 0 18"
+        )
+        self.stockfish.send_command("go")
+        self.stockfish.contains("score mate 1")
+
+        self.stockfish.starts_with("bestmove")
+
+    def test_fen_position_with_mate_go_depth_and_promotion(self):
+        self.stockfish.send_command("ucinewgame")
+        self.stockfish.send_command(
+            "position fen 8/5R2/2K1P3/4k3/8/b1PPpp1B/5p2/8 w - - moves c6d7 f2f1q"
+        )
+        self.stockfish.send_command("go depth 18")
+        self.stockfish.expect("* score mate 1 * pv f7f5")
+        self.stockfish.starts_with("bestmove f7f5")
+
+    def test_fen_position_with_mate_go_depth_and_searchmoves(self):
+        self.stockfish.send_command("ucinewgame")
+        self.stockfish.send_command(
+            "position fen 8/5R2/2K1P3/4k3/8/b1PPpp1B/5p2/8 w - -"
+        )
+        self.stockfish.send_command("go depth 18 searchmoves c6d7")
+        self.stockfish.expect("* score mate 2 * pv c6d7 * f7f5")
+
+        self.stockfish.starts_with("bestmove c6d7")
+
+    def test_fen_position_with_moves_with_mate_go_depth_and_searchmoves(self):
+        self.stockfish.send_command("ucinewgame")
+        self.stockfish.send_command(
+            "position fen 8/5R2/2K1P3/4k3/8/b1PPpp1B/5p2/8 w - - moves c6d7"
+        )
+        self.stockfish.send_command("go depth 18 searchmoves e3e2")
+        self.stockfish.expect("* score mate -1 * pv e3e2 f7f5")
+        self.stockfish.starts_with("bestmove e3e2")
+
+    def test_verify_nnue_network(self):
+        current_path = os.path.abspath(os.getcwd())
+        Stockfish(
+            f"export_net {os.path.join(current_path , 'verify.nnue')}".split(" "), True
+        )
+
+        self.stockfish.send_command("setoption name EvalFile value verify.nnue")
+        self.stockfish.send_command("position startpos")
+        self.stockfish.send_command("go depth 5")
+        self.stockfish.starts_with("bestmove")
+
+    def test_multipv_setting(self):
+        self.stockfish.send_command("setoption name MultiPV value 4")
+        self.stockfish.send_command("position startpos")
+        self.stockfish.send_command("go depth 5")
+        self.stockfish.starts_with("bestmove")
+
+    def test_fen_position_with_skill_level(self):
+        self.stockfish.send_command("setoption name Skill Level value 10")
+        self.stockfish.send_command("position startpos")
+        self.stockfish.send_command("go depth 5")
+        self.stockfish.starts_with("bestmove")
+
+        self.stockfish.send_command("setoption name Skill Level value 20")
+
+
+class TestSyzygy(metaclass=OrderedClassMembers):
+    def beforeAll(self):
+        self.stockfish = Stockfish()
+
+    def afterAll(self):
+        self.stockfish.quit()
+        assert self.stockfish.close() == 0
+
+    def afterEach(self):
+        assert postfix_check(self.stockfish.get_output()) == True
+        self.stockfish.clear_output()
+
+    def test_syzygy_setup(self):
+        self.stockfish.starts_with("Stockfish")
+        self.stockfish.send_command("uci")
+        self.stockfish.send_command(
+            f"setoption name SyzygyPath value {os.path.join(PATH, 'syzygy')}"
+        )
+        self.stockfish.expect(
+            "info string Found 35 WDL and 35 DTZ tablebase files (up to 4-man)."
+        )
+
+    def test_syzygy_bench(self):
+        self.stockfish.send_command("bench 128 1 8 default depth")
+        self.stockfish.expect("Nodes searched  :*")
+
+    def test_syzygy_position(self):
+        self.stockfish.send_command("ucinewgame")
+        self.stockfish.send_command("position fen 4k3/PP6/8/8/8/8/8/4K3 w - - 0 1")
+        self.stockfish.send_command("go depth 5")
+
+        def check_output(output):
+            if "score cp 20000" in output or "score mate" in output:
+                return True
+
+        self.stockfish.check_output(check_output)
+        self.stockfish.expect("bestmove *")
+
+    def test_syzygy_position_2(self):
+        self.stockfish.send_command("ucinewgame")
+        self.stockfish.send_command("position fen 8/1P6/2B5/8/4K3/8/6k1/8 w - - 0 1")
+        self.stockfish.send_command("go depth 5")
+
+        def check_output(output):
+            if "score cp 20000" in output or "score mate" in output:
+                return True
+
+        self.stockfish.check_output(check_output)
+        self.stockfish.expect("bestmove *")
+
+    def test_syzygy_position_3(self):
+        self.stockfish.send_command("ucinewgame")
+        self.stockfish.send_command("position fen 8/1P6/2B5/8/4K3/8/6k1/8 b - - 0 1")
+        self.stockfish.send_command("go depth 5")
+
+        def check_output(output):
+            if "score cp -20000" in output or "score mate" in output:
+                return True
+
+        self.stockfish.check_output(check_output)
+        self.stockfish.expect("bestmove *")
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description="Run Stockfish with testing options")
+    parser.add_argument("--valgrind", action="store_true", help="Run valgrind testing")
+    parser.add_argument(
+        "--valgrind-thread", action="store_true", help="Run valgrind-thread testing"
+    )
+    parser.add_argument(
+        "--sanitizer-undefined",
+        action="store_true",
+        help="Run sanitizer-undefined testing",
+    )
+    parser.add_argument(
+        "--sanitizer-thread", action="store_true", help="Run sanitizer-thread testing"
+    )
+
+    parser.add_argument(
+        "--none", action="store_true", help="Run without any testing options"
+    )
+    parser.add_argument("stockfish_path", type=str, help="Path to Stockfish binary")
+
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    args = parse_args()
+
+    EPD.create_bench_epd()
+    TSAN.set_tsan_option()
+    Syzygy.download_syzygy()
+
+    framework = MiniTestFramework()
+
+    # Each test suite will be ran inside a temporary directory
+    framework.run([TestCLI, TestInteractive, TestSyzygy])
+
+    EPD.delete_bench_epd()
+    TSAN.unset_tsan_option()
+
+    if framework.has_failed():
+        sys.exit(1)
+
+    sys.exit(0)
diff --git a/tests/instrumented.sh b/tests/instrumented.sh
deleted file mode 100755
index 4c63fc57..00000000
--- a/tests/instrumented.sh
+++ /dev/null
@@ -1,301 +0,0 @@
-#!/bin/bash
-# check for errors under Valgrind or sanitizers.
-
-error()
-{
-  echo "instrumented testing failed on line $1"
-  exit 1
-}
-trap 'error ${LINENO}' ERR
-
-# define suitable post and prefixes for testing options
-case $1 in
-  --valgrind)
-    echo "valgrind testing started"
-    prefix=''
-    exeprefix='valgrind --error-exitcode=42 --errors-for-leak-kinds=all --leak-check=full'
-    postfix=''
-    threads="1"
-  ;;
-  --valgrind-thread)
-    echo "valgrind-thread testing started"
-    prefix=''
-    exeprefix='valgrind --fair-sched=try --error-exitcode=42'
-    postfix=''
-    threads="2"
-  ;;
-  --sanitizer-undefined)
-    echo "sanitizer-undefined testing started"
-    prefix='!'
-    exeprefix=''
-    postfix='2>&1 | grep -A50 "runtime error:"'
-    threads="1"
-  ;;
-  --sanitizer-thread)
-    echo "sanitizer-thread testing started"
-    prefix='!'
-    exeprefix=''
-    postfix='2>&1 | grep -A50 "WARNING: ThreadSanitizer:"'
-    threads="2"
-
-cat << EOF > tsan.supp
-race:Stockfish::TTEntry::move
-race:Stockfish::TTEntry::depth
-race:Stockfish::TTEntry::bound
-race:Stockfish::TTEntry::save
-race:Stockfish::TTEntry::value
-race:Stockfish::TTEntry::eval
-race:Stockfish::TTEntry::is_pv
-
-race:Stockfish::TranspositionTable::probe
-race:Stockfish::TranspositionTable::hashfull
-
-EOF
-
-    export TSAN_OPTIONS="suppressions=./tsan.supp"
-
-  ;;
-  *)
-    echo "unknown testing started"
-    prefix=''
-    exeprefix=''
-    postfix=''
-    threads="1"
-  ;;
-esac
-
-cat << EOF > bench_tmp.epd
-Rn6/1rbq1bk1/2p2n1p/2Bp1p2/3Pp1pP/1N2P1P1/2Q1NPB1/6K1 w - - 2 26
-rnbqkb1r/ppp1pp2/5n1p/3p2p1/P2PP3/5P2/1PP3PP/RNBQKBNR w KQkq - 0 3
-3qnrk1/4bp1p/1p2p1pP/p2bN3/1P1P1B2/P2BQ3/5PP1/4R1K1 w - - 9 28
-r4rk1/1b2ppbp/pq4pn/2pp1PB1/1p2P3/1P1P1NN1/1PP3PP/R2Q1RK1 w - - 0 13
-EOF
-
-# simple command line testing
-for args in "eval" \
-            "go nodes 1000" \
-            "go depth 10" \
-            "go perft 4" \
-            "go movetime 1000" \
-            "go wtime 8000 btime 8000 winc 500 binc 500" \
-            "go wtime 1000 btime 1000 winc 0 binc 0" \
-            "go wtime 1000 btime 1000 winc 0 binc 0" \
-            "go wtime 1000 btime 1000 winc 0 binc 0 movestogo 5" \
-            "go movetime 200" \
-            "go nodes 20000 searchmoves e2e4 d2d4" \
-            "bench 128 $threads 8 default depth" \
-            "bench 128 $threads 3 bench_tmp.epd depth" \
-            "export_net verify.nnue" \
-            "d" \
-            "compiler" \
-            "license" \
-            "uci"
-do
-
-   echo "$prefix $exeprefix ./stockfish $args $postfix"
-   eval "$prefix $exeprefix ./stockfish $args $postfix"
-
-done
-
-# verify the generated net equals the base net
-network=`./stockfish uci | grep 'option name EvalFile type string default' | awk '{print $NF}'`
-echo "Comparing $network to the written verify.nnue"
-diff $network verify.nnue
-
-# more general testing, following an uci protocol exchange
-cat << EOF > game.exp
- set timeout 240
- # to correctly catch eof we need the following line
- # expect_before timeout { exit 2 } eof { exit 3 }
- expect_before timeout { exit 2 }
-
- spawn $exeprefix ./stockfish
- expect "Stockfish"
-
- send "uci\n"
- expect "uciok"
-
- # send "setoption name Debug Log File value debug.log\n"
- send "setoption name Threads value $threads\n"
-
- send "ucinewgame\n"
- send "position startpos\n"
- send "go nodes 1000\n"
- expect "bestmove"
-
- send "ucinewgame\n"
- send "position startpos moves e2e4 e7e6\n"
- send "go nodes 1000\n"
- expect "bestmove"
-
- send "ucinewgame\n"
- send "position fen 5rk1/1K4p1/8/8/3B4/8/8/8 b - - 0 1\n"
- send "go depth 10\n"
- expect "bestmove"
-
- send "ucinewgame\n"
- send "position fen 5rk1/1K4p1/8/8/3B4/8/8/8 b - - 0 1\n"
- send "flip\n"
- send "go depth 10\n"
- expect "bestmove"
-
- send "ucinewgame\n"
- send "position startpos\n"
- send "go depth 5\n"
- expect -re {info depth \d+ seldepth \d+ multipv \d+ score cp \d+ nodes \d+ nps \d+ hashfull \d+ tbhits \d+ time \d+ pv}
- expect "bestmove"
-
- send "ucinewgame\n"
- send "setoption name UCI_ShowWDL value true\n"
- send "position startpos\n"
- send "go depth 9\n"
- expect -re {info depth 1 seldepth \d+ multipv \d+ score cp \d+ wdl \d+ \d+ \d+ nodes \d+ nps \d+ hashfull \d+ tbhits \d+ time \d+ pv}
- expect -re {info depth 2 seldepth \d+ multipv \d+ score cp \d+ wdl \d+ \d+ \d+ nodes \d+ nps \d+ hashfull \d+ tbhits \d+ time \d+ pv}
- expect -re {info depth 3 seldepth \d+ multipv \d+ score cp \d+ wdl \d+ \d+ \d+ nodes \d+ nps \d+ hashfull \d+ tbhits \d+ time \d+ pv}
- expect -re {info depth 4 seldepth \d+ multipv \d+ score cp \d+ wdl \d+ \d+ \d+ nodes \d+ nps \d+ hashfull \d+ tbhits \d+ time \d+ pv}
- expect -re {info depth 5 seldepth \d+ multipv \d+ score cp \d+ wdl \d+ \d+ \d+ nodes \d+ nps \d+ hashfull \d+ tbhits \d+ time \d+ pv}
- expect -re {info depth 6 seldepth \d+ multipv \d+ score cp \d+ wdl \d+ \d+ \d+ nodes \d+ nps \d+ hashfull \d+ tbhits \d+ time \d+ pv}
- expect -re {info depth 7 seldepth \d+ multipv \d+ score cp \d+ wdl \d+ \d+ \d+ nodes \d+ nps \d+ hashfull \d+ tbhits \d+ time \d+ pv}
- expect -re {info depth 8 seldepth \d+ multipv \d+ score cp \d+ wdl \d+ \d+ \d+ nodes \d+ nps \d+ hashfull \d+ tbhits \d+ time \d+ pv}
- expect -re {info depth 9 seldepth \d+ multipv \d+ score cp \d+ wdl \d+ \d+ \d+ nodes \d+ nps \d+ hashfull \d+ tbhits \d+ time \d+ pv}
- expect "bestmove"
-
- send "setoption name Clear Hash\n"
-
- send "ucinewgame\n"
- send "position fen 5K2/8/2qk4/2nPp3/3r4/6B1/B7/3R4 w - e6\n"
- send "go depth 18\n"
- expect "score mate 1"
- expect "pv d5e6"
- expect "bestmove d5e6"
-
- send "ucinewgame\n"
- send "position fen 2brrb2/8/p7/Q7/1p1kpPp1/1P1pN1K1/3P4/8 b - -\n"
- send "go depth 18\n"
- expect "score mate -1"
- expect "bestmove"
-
- send "ucinewgame\n"
- send "position fen 8/5R2/2K1P3/4k3/8/b1PPpp1B/5p2/8 w - -\n"
- send "go depth 18\n"
- expect "score mate 2 * pv c6d7 * f7f5"
- expect "bestmove c6d7"
-
- send "ucinewgame\n"
- send "position fen 8/5R2/2K1P3/4k3/8/b1PPpp1B/5p2/8 w - -\n"
- send "go mate 2\n"
- expect "score mate 2 * pv c6d7"
- expect "bestmove c6d7"
-
- send "ucinewgame\n"
- send "position fen 8/5R2/2K1P3/4k3/8/b1PPpp1B/5p2/8 w - -\n"
- send "go nodes 10000\n"
- expect "score mate 2 * pv c6d7 * f7f5"
- expect "bestmove c6d7"
-
- send "ucinewgame\n"
- send "position fen 1NR2B2/5p2/5p2/1p1kpp2/1P2rp2/2P1pB2/2P1P1K1/8 b - - \n"
- send "go depth 18\n"
- expect "score mate -2"
- expect "pv d5e6 c8d8"
- expect "bestmove d5e6"
-
- send "ucinewgame\n"
- send "position fen 8/5R2/2K1P3/4k3/8/b1PPpp1B/5p2/8 w - - moves c6d7 f2f1q\n"
- send "go depth 18\n"
- expect "score mate 1 * pv f7f5"
- expect "bestmove f7f5"
- 
- send "ucinewgame\n"
- send "position fen 8/5R2/2K1P3/4k3/8/b1PPpp1B/5p2/8 w - -\n"
- send "go depth 18 searchmoves c6d7\n"
- expect "score mate 2 * pv c6d7 * f7f5"
- expect "bestmove c6d7"
- 
- send "ucinewgame\n"
- send "position fen 8/5R2/2K1P3/4k3/8/b1PPpp1B/5p2/8 w - - moves c6d7\n"
- send "go depth 18 searchmoves e3e2\n"
- expect "score mate -1 * pv e3e2 f7f5"
- expect "bestmove e3e2"
-
- send "setoption name EvalFile value verify.nnue\n"
- send "position startpos\n"
- send "go depth 5\n"
- expect "bestmove"
-
- send "setoption name MultiPV value 4\n"
- send "position startpos\n"
- send "go depth 5\n"
- expect "bestmove"
-
- send "setoption name Skill Level value 10\n"
- send "position startpos\n"
- send "go depth 5\n"
- expect "bestmove"
- send "setoption name Skill Level value 20\n"
-
- send "quit\n"
- expect eof
-
- # return error code of the spawned program, useful for Valgrind
- lassign [wait] pid spawnid os_error_flag value
- exit \$value
-EOF
-
-#download TB as needed
-if [ ! -d ../tests/syzygy ]; then
-   curl -sL https://api.github.com/repos/niklasf/python-chess/tarball/9b9aa13f9f36d08aadfabff872882f4ab1494e95 | tar -xzf -
-   mv niklasf-python-chess-9b9aa13 ../tests/syzygy
-fi
-
-cat << EOF > syzygy.exp
- set timeout 240
- # to correctly catch eof we need the following line
- # expect_before timeout { exit 2 } eof { exit 3 }
- expect_before timeout { exit 2 }
- spawn $exeprefix ./stockfish
- expect "Stockfish"
- send "uci\n"
- send "setoption name SyzygyPath value ../tests/syzygy/\n"
- expect "info string Found 35 tablebases"
- send "bench 128 1 8 default depth\n"
- expect "Nodes searched  :"
- send "ucinewgame\n"
- send "position fen 4k3/PP6/8/8/8/8/8/4K3 w - - 0 1\n"
- send "go depth 5\n"
- expect -re {score cp 20000|score mate}
- expect "bestmove"
- send "ucinewgame\n"
- send "position fen 8/1P6/2B5/8/4K3/8/6k1/8 w - - 0 1\n"
- send "go depth 5\n"
- expect -re {score cp 20000|score mate}
- expect "bestmove"
- send "ucinewgame\n"
- send "position fen 8/1P6/2B5/8/4K3/8/6k1/8 b - - 0 1\n"
- send "go depth 5\n"
- expect -re {score cp -20000|score mate}
- expect "bestmove"
- send "quit\n"
- expect eof
-
- # return error code of the spawned program, useful for Valgrind
- lassign [wait] pid spawnid os_error_flag value
- exit \$value
-EOF
-
-for exp in game.exp syzygy.exp
-do
-
-  echo "======== $exp =============="
-  cat $exp
-  echo "============================"
-  echo "$prefix expect $exp $postfix"
-  eval "$prefix expect $exp $postfix"
-
-  rm $exp
-
-done
-
-rm -f tsan.supp bench_tmp.epd
-
-echo "instrumented testing OK"
diff --git a/tests/perft.sh b/tests/perft.sh
index 545e750f..c1532c20 100755
--- a/tests/perft.sh
+++ b/tests/perft.sh
@@ -1,5 +1,5 @@
 #!/bin/bash
-# verify perft numbers (positions from www.chessprogramming.org/Perft_Results)
+# verify perft numbers (positions from https://www.chessprogramming.org/Perft_Results)
 
 error()
 {
diff --git a/tests/testing.py b/tests/testing.py
new file mode 100644
index 00000000..d51ca89a
--- /dev/null
+++ b/tests/testing.py
@@ -0,0 +1,378 @@
+import subprocess
+from typing import List
+import os
+import collections
+import time
+import sys
+import traceback
+import fnmatch
+from functools import wraps
+from contextlib import redirect_stdout
+import io
+import tarfile
+import pathlib
+import concurrent.futures
+import tempfile
+import shutil
+import requests
+
+CYAN_COLOR = "\033[36m"
+GRAY_COLOR = "\033[2m"
+RED_COLOR = "\033[31m"
+GREEN_COLOR = "\033[32m"
+RESET_COLOR = "\033[0m"
+WHITE_BOLD = "\033[1m"
+
+MAX_TIMEOUT = 60 * 5
+
+PATH = pathlib.Path(__file__).parent.resolve()
+
+
+class Valgrind:
+    @staticmethod
+    def get_valgrind_command():
+        return [
+            "valgrind",
+            "--error-exitcode=42",
+            "--errors-for-leak-kinds=all",
+            "--leak-check=full",
+        ]
+
+    @staticmethod
+    def get_valgrind_thread_command():
+        return ["valgrind", "--error-exitcode=42", "--fair-sched=try"]
+
+
+class TSAN:
+    @staticmethod
+    def set_tsan_option():
+        with open(f"tsan.supp", "w") as f:
+            f.write(
+                """
+race:Stockfish::TTEntry::read
+race:Stockfish::TTEntry::save
+race:Stockfish::TranspositionTable::probe
+race:Stockfish::TranspositionTable::hashfull
+"""
+            )
+
+        os.environ["TSAN_OPTIONS"] = "suppressions=./tsan.supp"
+
+    @staticmethod
+    def unset_tsan_option():
+        os.environ.pop("TSAN_OPTIONS", None)
+        os.remove(f"tsan.supp")
+
+
+class EPD:
+    @staticmethod
+    def create_bench_epd():
+        with open(f"{os.path.join(PATH,'bench_tmp.epd')}", "w") as f:
+            f.write(
+                """
+Rn6/1rbq1bk1/2p2n1p/2Bp1p2/3Pp1pP/1N2P1P1/2Q1NPB1/6K1 w - - 2 26
+rnbqkb1r/ppp1pp2/5n1p/3p2p1/P2PP3/5P2/1PP3PP/RNBQKBNR w KQkq - 0 3
+3qnrk1/4bp1p/1p2p1pP/p2bN3/1P1P1B2/P2BQ3/5PP1/4R1K1 w - - 9 28
+r4rk1/1b2ppbp/pq4pn/2pp1PB1/1p2P3/1P1P1NN1/1PP3PP/R2Q1RK1 w - - 0 13
+"""
+            )
+
+    @staticmethod
+    def delete_bench_epd():
+        os.remove(f"{os.path.join(PATH,'bench_tmp.epd')}")
+
+
+class Syzygy:
+    @staticmethod
+    def get_syzygy_path():
+        return os.path.abspath("syzygy")
+
+    @staticmethod
+    def download_syzygy():
+        if not os.path.isdir(os.path.join(PATH, "syzygy")):
+            url = "https://api.github.com/repos/niklasf/python-chess/tarball/9b9aa13f9f36d08aadfabff872882f4ab1494e95"
+            file = "niklasf-python-chess-9b9aa13"
+
+            with tempfile.TemporaryDirectory() as tmpdirname:
+                tarball_path = os.path.join(tmpdirname, f"{file}.tar.gz")
+
+                response = requests.get(url, stream=True)
+                with open(tarball_path, 'wb') as f:
+                    for chunk in response.iter_content(chunk_size=8192):
+                        f.write(chunk)
+
+                with tarfile.open(tarball_path, "r:gz") as tar:
+                    tar.extractall(tmpdirname)
+
+                shutil.move(os.path.join(tmpdirname, file), os.path.join(PATH, "syzygy"))
+
+class OrderedClassMembers(type):
+    @classmethod
+    def __prepare__(self, name, bases):
+        return collections.OrderedDict()
+
+    def __new__(self, name, bases, classdict):
+        classdict["__ordered__"] = [
+            key for key in classdict.keys() if key not in ("__module__", "__qualname__")
+        ]
+        return type.__new__(self, name, bases, classdict)
+
+
+class TimeoutException(Exception):
+    def __init__(self, message: str, timeout: int):
+        self.message = message
+        self.timeout = timeout
+
+
+def timeout_decorator(timeout: float):
+    def decorator(func):
+        @wraps(func)
+        def wrapper(*args, **kwargs):
+            with concurrent.futures.ThreadPoolExecutor() as executor:
+                future = executor.submit(func, *args, **kwargs)
+                try:
+                    result = future.result(timeout=timeout)
+                except concurrent.futures.TimeoutError:
+                    raise TimeoutException(
+                        f"Function {func.__name__} timed out after {timeout} seconds",
+                        timeout,
+                    )
+            return result
+
+        return wrapper
+
+    return decorator
+
+
+class MiniTestFramework:
+    def __init__(self):
+        self.passed_test_suites = 0
+        self.failed_test_suites = 0
+        self.passed_tests = 0
+        self.failed_tests = 0
+
+    def has_failed(self) -> bool:
+        return self.failed_test_suites > 0
+
+    def run(self, classes: List[type]) -> bool:
+        self.start_time = time.time()
+
+        for test_class in classes:
+            with tempfile.TemporaryDirectory() as tmpdirname:
+                original_cwd = os.getcwd()
+                os.chdir(tmpdirname)
+
+                try:
+                    if self.__run(test_class):
+                        self.failed_test_suites += 1
+                    else:
+                        self.passed_test_suites += 1
+                finally:
+                    os.chdir(original_cwd)
+
+        self.__print_summary(round(time.time() - self.start_time, 2))
+        return self.has_failed()
+
+    def __run(self, test_class) -> bool:
+        test_instance = test_class()
+        test_name = test_instance.__class__.__name__
+        test_methods = [m for m in test_instance.__ordered__ if m.startswith("test_")]
+
+        print(f"\nTest Suite: {test_name}")
+
+        if hasattr(test_instance, "beforeAll"):
+            test_instance.beforeAll()
+
+        fails = 0
+
+        for method in test_methods:
+            fails += self.__run_test_method(test_instance, method)
+
+        if hasattr(test_instance, "afterAll"):
+            test_instance.afterAll()
+
+        self.failed_tests += fails
+
+        return fails > 0
+
+    def __run_test_method(self, test_instance, method: str) -> int:
+        print(f"    Running {method}... \r", end="", flush=True)
+
+        buffer = io.StringIO()
+        fails = 0
+
+        try:
+            t0 = time.time()
+
+            with redirect_stdout(buffer):
+                if hasattr(test_instance, "beforeEach"):
+                    test_instance.beforeEach()
+
+                getattr(test_instance, method)()
+
+                if hasattr(test_instance, "afterEach"):
+                    test_instance.afterEach()
+
+            duration = time.time() - t0
+
+            self.print_success(f" {method} ({duration * 1000:.2f}ms)")
+            self.passed_tests += 1
+        except Exception as e:
+            if isinstance(e, TimeoutException):
+                self.print_failure(
+                    f" {method} (hit execution limit of {e.timeout} seconds)"
+                )
+
+            if isinstance(e, AssertionError):
+                self.__handle_assertion_error(t0, method)
+
+            fails += 1
+        finally:
+            self.__print_buffer_output(buffer)
+
+        return fails
+
+    def __handle_assertion_error(self, start_time, method: str):
+        duration = time.time() - start_time
+        self.print_failure(f" {method} ({duration * 1000:.2f}ms)")
+        traceback_output = "".join(traceback.format_tb(sys.exc_info()[2]))
+
+        colored_traceback = "\n".join(
+            f"  {CYAN_COLOR}{line}{RESET_COLOR}"
+            for line in traceback_output.splitlines()
+        )
+
+        print(colored_traceback)
+
+    def __print_buffer_output(self, buffer: io.StringIO):
+        output = buffer.getvalue()
+        if output:
+            indented_output = "\n".join(f"    {line}" for line in output.splitlines())
+            print(f"    {RED_COLOR}⎯⎯⎯⎯⎯OUTPUT⎯⎯⎯⎯⎯{RESET_COLOR}")
+            print(f"{GRAY_COLOR}{indented_output}{RESET_COLOR}")
+            print(f"    {RED_COLOR}⎯⎯⎯⎯⎯OUTPUT⎯⎯⎯⎯⎯{RESET_COLOR}")
+
+    def __print_summary(self, duration: float):
+        print(f"\n{WHITE_BOLD}Test Summary{RESET_COLOR}\n")
+        print(
+            f"    Test Suites: {GREEN_COLOR}{self.passed_test_suites} passed{RESET_COLOR}, {RED_COLOR}{self.failed_test_suites} failed{RESET_COLOR}, {self.passed_test_suites + self.failed_test_suites} total"
+        )
+        print(
+            f"    Tests:       {GREEN_COLOR}{self.passed_tests} passed{RESET_COLOR}, {RED_COLOR}{self.failed_tests} failed{RESET_COLOR}, {self.passed_tests + self.failed_tests} total"
+        )
+        print(f"    Time:        {duration}s\n")
+
+    def print_failure(self, add: str):
+        print(f"    {RED_COLOR}✗{RESET_COLOR}{add}", flush=True)
+
+    def print_success(self, add: str):
+        print(f"    {GREEN_COLOR}✓{RESET_COLOR}{add}", flush=True)
+
+
+class Stockfish:
+    def __init__(
+        self,
+        prefix: List[str],
+        path: str,
+        args: List[str] = [],
+        cli: bool = False,
+    ):
+        self.path = path
+        self.process = None
+        self.args = args
+        self.cli = cli
+        self.prefix = prefix
+        self.output = []
+
+        self.start()
+
+    def start(self):
+        if self.cli:
+            self.process = subprocess.run(
+                self.prefix + [self.path] + self.args,
+                capture_output=True,
+                text=True,
+            )
+
+            self.process.stdout
+
+            return
+
+        self.process = subprocess.Popen(
+            self.prefix + [self.path] + self.args,
+            stdin=subprocess.PIPE,
+            stdout=subprocess.PIPE,
+            stderr=subprocess.STDOUT,
+            universal_newlines=True,
+            bufsize=1,
+        )
+
+    def setoption(self, name: str, value: str):
+        self.send_command(f"setoption name {name} value {value}")
+
+    def send_command(self, command: str):
+        if not self.process:
+            raise RuntimeError("Stockfish process is not started")
+
+        self.process.stdin.write(command + "\n")
+        self.process.stdin.flush()
+
+    @timeout_decorator(MAX_TIMEOUT)
+    def equals(self, expected_output: str):
+        for line in self.readline():
+            if line == expected_output:
+                return
+
+    @timeout_decorator(MAX_TIMEOUT)
+    def expect(self, expected_output: str):
+        for line in self.readline():
+            if fnmatch.fnmatch(line, expected_output):
+                return
+
+    @timeout_decorator(MAX_TIMEOUT)
+    def contains(self, expected_output: str):
+        for line in self.readline():
+            if expected_output in line:
+                return
+
+    @timeout_decorator(MAX_TIMEOUT)
+    def starts_with(self, expected_output: str):
+        for line in self.readline():
+            if line.startswith(expected_output):
+                return
+
+    @timeout_decorator(MAX_TIMEOUT)
+    def check_output(self, callback):
+        if not callback:
+            raise ValueError("Callback function is required")
+
+        for line in self.readline():
+            if callback(line) == True:
+                return
+
+    def readline(self):
+        if not self.process:
+            raise RuntimeError("Stockfish process is not started")
+
+        while True:
+            line = self.process.stdout.readline().strip()
+            self.output.append(line)
+
+            yield line
+
+    def clear_output(self):
+        self.output = []
+
+    def get_output(self) -> List[str]:
+        return self.output
+
+    def quit(self):
+        self.send_command("quit")
+
+    def close(self):
+        if self.process:
+            self.process.stdin.close()
+            self.process.stdout.close()
+            return self.process.wait()
+
+        return 0