Hacker News Reader

Comment by gowld | Hacker News Reader

Do AMD GPUs work well with ML drivers (CUDA and competitors)? Are higher-level drivers like PyTorch and TensorFlow compatible (and optimized, or else no point buying an expensive GPU to get only half-performance) with AMD backends?

View on HN

It's far better today than it used to be, but it's still a work in progress. Over the past year or two, they've been doing a lot of hiring to bring ROCm HIP up to par as a CUDA competitor.

I wasn't exactly thrilled to see yet another GPU interface, but they seem to be trying to match the CUDA way of doing things. That ought to make supporting AMD much easier than it once was.

In general, no. The highest performance software only runs on Nvidia GPUs (or internal hardware, like Google’s TPUs).

I think this depends very much on what exactly you mean by "highest performance software". Some frameworks have upstreamed support for ROCm. A few of my colleagues have been doing some testing and mentioned to me they were reasonably impressed with performance on newer hardware.

My current impression is that AMD is not the best option for ML but is getting better over time. So YMMV. That said, I do not use my local machine for ML (and don’t do ML in any serious capacity.)

TensorFlow works with AMD's CUDA alternative named ROCm. You'd need to check installation instructions though.

Yes, ROCm is upstreamed into Tensorflow. On our Hopsworks platform, you can run the same TensorFlow code on either ROCm or Cuda. No code changes needed:

https://www.logicalclocks.com/blog/welcoming-amd-rocm-to-hop...

Performance ain't so bad on AMD GPUs anymore - and none of the EULA crap. Just buy a Vega R7 for 500 bucks and get 16 GB of memory and the performance of close to a 2080 Ti (for convnets, at least) - but with no data center restrictions:

https://github.com/ROCmSoftwarePlatform/tensorflow-upstream/...

How about Pytorch?

You have to compile it yourself, which sucks. For some reason PyTorch aren't agreeing to upstream ROCm, yet.

You don't need to compile it, AMD provides docker images.

For example I was able to run fast.ai/pytorch oh my amd gpu https://github.com/briangorman/fastai_rocm_docker

Reply on news.ycombinator.com