Created by: xiaolil1
We enabled MKLDNN INT8 inference. We got these inference result:
- Accuracy
| Topology | FP32 Accuracy(Top-1/Top-5) | INT8 Accurary(Top-1/Top-5) |
|---|---|---|
| ResNet-50 (FB) | 76.63%/93.10% | 76.42%/93.07% |
| MobileNet-V1 | 70.78%/89.69% | 70.10%/89.30% |
- Performance
| Topology | FP32 Throughput | INT8 Throughput | FP32 latency(1x1) | INT8 latency(1x1) |
|---|---|---|---|---|
| ResNet-50 (FB) | 260 | 527 | 62.4 | 36.7 |
| MobileNet-V1 | 1207 | 1934 | 9.3 | 8.4 |
Note: The above performance measured on SKX8180 1S (HT On, Turbo On) (Throughput imgs/sec; Latency: ms; 1x1: batch size 1 x thread 1)