This page details benchmark results comparing MXNet 1.3.0 with MKLDNN vs without MKLDNN. The results clearly shows that MKL-DNN boosts inference throughput between 6x to 37x, latency reduced between 2x to 41x, while accuracy is equivalent up to an epsilon of 1e-8.
Inference Performance
This group of the performance test is gathered on AWS EC2 instance C5.18xLarge with 1 socket and 1 processor.
For the throughput, 2 sockets can provide about 2X speedup while latency will keep the constant.
Performance boost with Intel MKL-DNN backend in release 1.3
The c5.18xlarge instance offers a 2-socket Intel Xeon Platinum processor with 72 vCPUs.
...
Category | Model | Latency batchsize=1 (ms, small is better) | Throughput batchsize=128 (fps, higher is better) | ||||
---|---|---|---|---|---|---|---|
no mkldnn | release 1.3 + mkldnn | speedup | no mkldnn | release 1.3 + mkldnn | speedup | ||
CNN/classification | ResNet-50 v1 | 97.19 | 18.94 | 5.13 | 10.29 | 132.05 | 12.84 |
ResNet-50 v2 | 98.69 | 18.93 | 5.21 | 9.94 | 127.17 | 12.79 | |
Inception v3 | 175.17 | 26.34 | 6.65 | 5.74 | 110.00 | 19.16 | |
Inception v4 | 330.93 | 66.96 | 4.94 | 3.04 | 59.28 | 19.47 | |
DenseNet | 111.66 | 53.31 | 2.09 | 8.52 | 121.79 | 14.30 | |
MobileNet | 38.56 | 7.32 | 5.27 | 24.87 | 380.54 | 15.30 | |
VGG16 | 406.50 | 40.08 | 10.14 | 2.91 | 69.84 | 23.96 | |
AlexNet | 64.60 | 4.33 | 14.90 | 26.58 | 689.86 | 25.96 | |
inception-resnet v2 | 181.10 | 111.28 | 1.63 | 5.48 | 69.39 | 12.66 | |
CNN/object detection | Faster R-CNN | 1175.74 | 95.15 | 12.36 | 0.85 | 10.51 | 12.36 |
SSD-VGG16 | 721.03 | 127.48 | 5.66 | 1.43(batchsize=224) | 27.35(batchsize=224) | 19.13 | |
SSD-MobileNet | 239.40 | 100.75 | 2.39 | 4.07 (batchsize=256) | 57.73(batchsize=256) | 14.18 | |
RNN | GNMT | 683.43 | 100.30 | 6.81 | 1.46(batchsize=64) | 9.97(batchsize=64) | 6.83 |
GAN | DCGAN | 8.94 | 0.22 | 41.36 | 109.13 | 4059.74 | 37.20 |
Inference Accuracy
The model is from gluon model zoo by pre-trained parameters. The top1 and top5 accuracy are verified by MKL-DNN backend.
...