Intel-caffe + MKL-DNN inference performance droped from 230 to 5 FPS about caffe HOT 7 CLOSED

Thanks very much for reporting this, could you please send us model of yours (could be narrowed down) that You used for this comparison?

from caffe.

It's a general problem, you can use bvlc_googlenet.

from caffe.

Ok, We just pushed some changes, that may affect/fix perofrmance of MKLDNN. Could you please retest and let us now if problem does still occur?

from caffe.

Seems it's been fixed, and I see there is a huge improvement on bvlc_alexnet.

from caffe.

Are you looking at average forward time of caffe time output, or just measure time it takes to do actual classification ?

from caffe.

Both, Is it normal that alexnet' inference perfromance has more than 30x improvement, but googlenet not ?

from caffe.

MKL DNN is still a demo library AlexNet performacne is first priority. Googlenet performance will be improved soon.

from caffe.