Re: [eigen] Intel Caffe2 benchmark

[Gael Guennebaud <gael.guennebaud@xxxxxxxxx>]

Hi,

it's hard to draw any conclusion without more information on which kind of BLAS operation have been used. Do they perform large GEMM in which case it is very odd to see Eigen performs so badly on the sequential version (did they forgot to enable AVX2/FMA?), or do they rely on many small gemm batched in a single MKL call?

gael

On Fri, Apr 21, 2017 at 7:13 PM, Francois fayard <fayard@xxxxxxxxxxxxx> wrote:

Hi,

Intel has just released a new blog post on Caffe2, the Facebook fork of Caffe. To promote the MKL, they usually benchmark the MKL BLAS against OpenBLAS. But it seems that the target has moved to Eigen.

https://software.intel.com/en-us/blogs/2017/04/18/intel-and-facebook-collaborate-to-boost-caffe2-performance-on-intel-cpu-s

My experience with Intel benchmark is that they are usually right. They tend to use hardware where the MKL really outperforms their contender. Here they use a dual socket system with a lot of cores. I don't know how well Eigen does here, but I know that OpenBLAS was really bad with so many cores last summer.

François