This repository was archived by the owner on Mar 4, 2026. It is now read-only.
Repository navigation
This repository was archived by the owner on Mar 4, 2026. It is now read-only.
ResNet50 inference model seemed not support FP16 #173
Description
I want to evaluate inference performance using AMX-FP16 on new CPU while default docker or AI tool package not supported AMX-FP16 because of old version of onednn(3.2.0), so building onednn-3.3.0 and intel extensions for tensorflow-2.15.0 from source and using following model:
FP32, FP16 and BFloat32 Pretrained model:
wget https://zenodo.org/record/2535873/files/resnet50_v1.pb
howerver, log from onednn showed precision of this model was FP32 not FP16.
Activity
sramakintel commentedon Mar 25, 2024
@chenming22 we will work on adding
FP16to resnet50 model. Thanks for reaching out.sramakintel commentedon Mar 25, 2024
@chenming22 we currently have AMP FP16 TF support based on the PR upstream here: tensorflow/tensorflow#62817. Since the script already supports FP16 input, installing
tf-nightyfrom pip can provide FP16 support. Can you run the workload again to see if FP16 works as expected?chenming22 commentedon Mar 26, 2024
thanks @sramakintel , i will have a try.
chenming22 commentedon Jun 6, 2024
hi, @sramakintel , noticed above pr been merged to tf2.16.1, so i install tf2.16.1 from pip and then ran inference.sh in ResNet50 V1.5 inference with precision=fp16, ONEDNN_MAX_CPU_ISA=AVX512_CORE_AMX_FP16 and model: resnet50_v1.pb, but log and metric on emon showed no AMX instructions used, any other environment variables or configurations needed?