diff --git a/tensorflow/lite/delegates/gpu/metal/kernels/conv.cc b/tensorflow/lite/delegates/gpu/metal/kernels/conv.cc
index 8f63fab7cf5..0a851a00a57 100644
--- a/tensorflow/lite/delegates/gpu/metal/kernels/conv.cc
+++ b/tensorflow/lite/delegates/gpu/metal/kernels/conv.cc
@@ -920,7 +920,7 @@ ConvParams GetConvParamsForIntel(const Convolution2DAttributes& attr,
   const int dst_slices = IntegralDivideRoundUp(dst_shape.c, 4);
   const int src_slices = IntegralDivideRoundUp(attr.weights.shape.i, 4);
   ConvParams params;
-  params.weights_upload_type = WeightsUploadType::PRIVATE_MEM_SIMD16_BROADCAST;
+  params.weights_upload_type = WeightsUploadType::PRIVATE_MEM_SIMD8_BROADCAST;
   params.x_kernel_is_1 = IsKernelXIs1(attr);
   params.y_kernel_is_1 = IsKernelYIs1(attr);
   params.src_depth_loop_size = 1;
@@ -1132,8 +1132,7 @@ std::vector<ComputeTaskDescriptorPtr> ConvolutionWino4x4To6x6(
     }
   } else if (device_info.IsIntelGPU()) {
     params.weight_layout = WeightsInnerBlockLayout::I4O4;
-    params.weights_upload_type =
-        WeightsUploadType::PRIVATE_MEM_SIMD16_BROADCAST;
+    params.weights_upload_type = WeightsUploadType::PRIVATE_MEM_SIMD8_BROADCAST;
     params.work_group_size = int3(16, 1, 1);
     params.block_size = int3(1, 1, 4);
   } else if (device_info.IsAMDGPU()) {