Created by: xiaolil1
We enabled mkldnn INT8 inference with performance:
Paddle version | Topology | Performance measured on SKX8180 1S (HT On, Turbo On) (Throughput imgs/sec; Latency: ms; 1x1: batch size 1 x thread 1 | |
---|---|---|---|
FP32 Throughput | INT8 Throughput | FP32 Latency (1x1) | INT8 Latency (1x1) |
Our master: 77ca0bce | ResNet-50 (FB) | 338 | 537 |
MobileNet-V1 | 1206 | 1958 | 8.8 |
Baidu develop: b82a44ea | ResNet-50 (FB) | 335 | |
MobileNet-V1 | 1193 | 9.6 | |
Paddle version | Topology | Performance measured on SKX6148 1S (HT On, Turbo On) (Throughput imgs/sec; Latency: ms; 1x1: batch size 1 x thread 1 | |
FP32 Throughput | INT8 Throughput | FP32 Latency (1x1) | INT8 Latency (1x1) |
Our master: 77ca0bce | ResNet-50 (FB) | 240 | 384 |
MobileNet-V1 | 1032 | 1598 | 9.0 |
Baidu develop: b82a44ea | ResNet-50 (FB) | 239 | |
MobileNet-V1 | 1015 | 9.8 |