diff --git a/src/Makefile b/src/Makefile
index fd2618a4..e34fbf61 100644
--- a/src/Makefile
+++ b/src/Makefile
@@ -416,7 +416,7 @@ endif
 ifeq ($(avx512),yes)
 	CXXFLAGS += -DUSE_AVX512
 	ifeq ($(comp),$(filter $(comp),gcc clang mingw))
-		CXXFLAGS += -mavx512bw
+		CXXFLAGS += -mavx512f -mavx512bw
 	endif
 endif
 
diff --git a/src/nnue/layers/affine_transform.h b/src/nnue/layers/affine_transform.h
index 985ee71a..8d2acd18 100644
--- a/src/nnue/layers/affine_transform.h
+++ b/src/nnue/layers/affine_transform.h
@@ -126,8 +126,7 @@ namespace Eval::NNUE::Layers {
             const auto iv256  = reinterpret_cast<const __m256i*>(&input_vector[kNumChunks]);
             const auto row256 = reinterpret_cast<const __m256i*>(&row[kNumChunks]);
             __m256i product256 = _mm256_maddubs_epi16(_mm256_loadA_si256(&iv256[0]), _mm256_load_si256(&row256[0]));
-            product256 = _mm256_madd_epi16(product256, _mm256_set1_epi16(1));
-            sum = _mm512_add_epi32(sum, _mm512_zextsi256_si512(product256));
+            sum = _mm512_add_epi32(sum, _mm512_cvtepi16_epi32(product256));
         }
         output[i] = _mm512_reduce_add_epi32(sum) + biases_[i];