How is caffe timed? about nnpack HOT 2 CLOSED

It is benchmarked on cpu branch of convent-benchmarks, with some modifications:

1. We use Caffe+OpenBLAS as a reference
2. Batch size is 64 for VGG-A and 128 for AlexNet & OverFeat. It is important that you use protobufs from convnet-benchmarks: specifications from other sources (e.g. in Caffe itself) can use different image sizes and numbers of channels in hidden layers.
3. Use caffe-nnpack to integrate NNPACK into Caffe. As caffe-nnpack is currently incomplete, and doesn't implement backward pass, you'd need to remove this line, so that caffe-nnpack doesn't abort on backward pass.
4. Next, you need to modify prototxt with models to make some layers use NNPACK. Inside conv_param specify engine: NNPACK to make the layer compute through NNPACK.

from nnpack.

Thanks a lot! With your help I'm able to track down the problem: OpenBLAS I compiled myself leads to about 1.9x speedup compared to the one that comes with Ubuntu 14.04.

from nnpack.