PyTorch-benchmark

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

Installation

Refer to Intel version. It is important to warn you that the branch is icc.

Benchmarking data  

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 The reference data is outdated.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 --