Performance difference between GPU version with/without MKL

Open Moulick opened this issue 7 years ago • 2 comments

Not a issue, but a general question. Will having MKL matter at all if the calculations are being done on the GPU?

Feb 09 '18 03:02 Moulick

Theoretically no, but depending on your model and parameters, MKL builds may be slower. There are some discussions in #12.

Feb 27 '18 19:02 danqing

I am confused because TensorFlow's performance guide writes:

Note: MKL was added as of TensorFlow 1.2 and currently only works on Linux. It also does not work when also using --config=cuda.

Are we sure that the GPU+MKL wheels are actually built with MKL support?

Mar 05 '18 21:03 kaczmarj