Running Caffe AlexNet/GoogleNet On Some CPUs Compared To NVIDIA CUDA

[phoronix]

Phoronix: Running Caffe AlexNet/GoogleNet On Some CPUs Compared To NVIDIA CUDA

With working on some Broadwell-EP Linux comparison benchmarks this weekend, as part of that onslaught of benchmarks I decided to run the CPU-only Caffe build on a few different Intel CPUs. For fun, afterwards I checked to see how the performance compares to Caffe with CUDA+cuDNN on a few Maxwell/Pascal GPUs...

http://www.phoronix.com/scan.php?page=news_item&px=Caffe-CPU-GPU-Fun

[nslay]

Nice work!

I'm not sure you can get a hold of Intel MKL, but I think MKL vs GPU would be a very interesting benchmark. About 2 years ago, some of my colleagues found MKL was faster for certain kinds of networks/layers. Not sure if this is still true anymore.

[Michael]

Nice work!

I'm not sure you can get a hold of Intel MKL, but I think MKL vs GPU would be a very interesting benchmark. About 2 years ago, some of my colleagues found MKL was faster for certain kinds of networks/layers. Not sure if this is still true anymore.

Unfortunately Intel has never sent over any of their workstation/server.HPC products... Even these days for consumer CPUs it's tough to get them directly from Intel for Linux reviews.

[edwaleni]

I have a presentation on GPU use in the enterprise coming up, this will be a good reference point. These are definitely hitting the mainstream now in application development. I just had my first analytics app come across my desk using CUDA and .NET. The rest mostly use Linux. It's just a matter of time before MSFT uses something like DirectX to scale up SSAS with GPU's.

[fuzz]

Isn't there a utility from AMD to convert CUDA to OpenCL? Would that be worth trying?

[intelmatt]

Unfortunately Intel has never sent over any of their workstation/server.HPC products... Even these days for consumer CPUs it's tough to get them directly from Intel for Linux reviews.

Michael- Intel has optimized Caffe specifically for their CPUs and made it available on github. You should give it a try!

https://software.intel.com/en-us/articles/training-and-deploying-deep-learning-networks-with-caffe-optimized-for-intel-architecture

-Matt
(disclaimer: I work at intel but am speaking only for myself)

[coder]

Unfortunately Intel has never sent over any of their workstation/server.HPC products... Even these days for consumer CPUs it's tough to get them directly from Intel for Linux reviews.

MKL is a software library. You can download Intel's optimized version, freely from their site.

https://software.intel.com/en-us/intel-mkl

There's also an OpenCL fork of Caffe, for use with AMD GPUs and the iGPUs inside Intel's desktop/laptop/mobile processors.

https://github.com/BVLC/caffe/tree/opencl