performance dropped to the bottom when running on Atom device linked with MKL2017 about caffe HOT 5 CLOSED

Hi,
Intel® Math Kernel Library 2017 is optimized for the Intel® Xeon Phi™ processors. Please try using older Intel® MKL versions for Atom processor.

from caffe.

1. Did you tried on Atom or Xeon Phi processor?
2. Did you had AVX 512 enabled?
3. Did you used proper engine selection to use MKL 2017? (in command line add -engine=MKL2017)

from caffe.

No response - closing

from caffe.

back from vacation.

1. Did you tried on Atom or Xeon Phi processor?
   A: on Atom processor. Model name is Atom C2750 2.4GHz
2. Did you had AVX 512 enabled?
   A: This process does not support AVX512 or AVX2
3. proper engine selection is used, on latest Intel Caffe (with new MKL release 2017.0.2.20170110), performance is better but still lower than OOB Intel Caffe:
   batch size changed to 1
   ./build/tools/caffe time -engine=MKL2017 -model models/bvlc_alexnet/deploy.prototxt
   Average Forward pass: 147.002 ms.
   ./build/tools/caffe time -model models/bvlc_alexnet/deploy.prototxt
   Average Forward pass: 66.6948 ms.

from caffe.

This work is aimed at Intel(r) Xeon(tm) and Intel XeonPhi(tm) optimization.
Atom processors are not our target and we will not optimize for it.

from caffe.