Benchmarking of CNN part of a some project on different mainstreaming CPU platform with PyTorch. Besides the official PyTorch, we also provide an Intel version to maximize the performance on CPU platforms, in particular Intel® Xeon processors (HSW, BDW, Xeon Phi).
Tested CPU：
| Type | CPU Model | Sockets | Cores/Socket | Frequency |
|---|---|---|---|---|
| i7 | Intel(R) Core(TM) i7-5960X | 1 | 8 | 3.00GHz |
| BDW | Intel(R) Xeon(R) CPU E5-2699 v4 | 2 | 22 | 2.20GHz |
| KNL | Intel(R) Xeon Phi(TM) CPU 7250F | 1 | 68 | 1.40GHz |
| SKX | Intel(R) Xeon(R) Platinum 8180 CPU | 2 | 28 | 2.50GHz |
Refer to Intel version. It is important to warn you that the branch is icc.
Note: Part of optimization from Intel has been merged into official version. The reference data is outdated.
Unit:(ms)
| PyTorch version | compiler | AVX512 support | CPU platform | result | speedUp |
|---|---|---|---|---|---|
| Official | GNU | No | BDW | 1450 | (reference) |
| Intel | GNU | No | BDW | 220 | 6.59X |
| Intel | Intel | No | BDW | 172 | 8.43X |
| Intel | Intel | YES | KNL | 262 | -- |
| Intel | Intel | YES | SKX | 111 | -- |
| Intel | GNU | No | i7 | 460 | -- |