Hacker News Reader

Comment by semisight | Hacker News Reader

Exciting to see this! I haven't taken a look at TensorFlow code yet but it looks very similar (method names and all) to Caffe. I've wanted a CNN library for Rust for a while.

It sounds like performance is important to you. Are there any plans to add GPU paths soon? If it's CUDA, are you going to use cuDNN?

View on HN

Yes, GPU support has high priority but we are still investigating what library/tools we want to go with, since the current options in Rust are not that great. Maybe we'll use Arrayfire, maybe we'll use CUDA/cuDNN (would require writing Rust binding) or maybe something else.

If you're going for GPU support then please do not use CUDA. It's non-portable (NVIDIA only), proprietary, clunky and it requires a huge SDK which is very awkward to install. I highly recommend OpenCL which is portable, requires no SDK, is open and for properly optimized code is just as fast as CUDA.

While this sounds attractive, OpenCL performance just doesn't seem to be there at the moment.

In CNN benchmarks[1] the CL based implementations consistently finish at the bottom. I think this is mostly because of the weak nVidia OpenCL implementations, but that doesn't help, since most people use nVidia (eg, Amazon GPU instances are nVidia based)

[1] https://github.com/soumith/convnet-benchmarks

Another nice thing about OpenCL is that it is also supported by non-GPU hardware. Kronos maintains a pretty impressive list: https://www.khronos.org/conformance/adopters/conformant-prod...

Reply on news.ycombinator.com