GPU Library for AMD GPUs


#1

I have followed the setup scripts given in http://wiki.fast.ai/index.php/Installation but all of them uses NVIDIA GPUs and installs the cuda library. Is there any equivalent to cuda for AMD GPUs for local installation?


#2

Hey, in the first place, you should not be following that setup guide. It was meant for fast.ai part 1 version 1. For AI Saturdays, we will be learning part 1 version 2 (v2) of the online course.

CUDA is a parallel computing platform and application programming interface (API) model created by NVIDIA. So, CUDA is limited to NVIDIA hardware. For AMD GPU, OpenCL would be the best alternative.

https://developer.amd.com/tools-and-sdks/opencl-zone/

AMD released ROCm, their open-source platform for GPU computing.

https://rocm.github.io/dl.html

Deep Learning Framework support for ROCm:

  • TensorFlow, PyTorch [status: under development]
  • Caffe [status: public]
  • and more

AMD GPU doesn’t work well with most machine learning/deep learning libraries yet.


#3

Thanks for sharing the information. Yes I have setup the cloud environment as per the docs. But I was just curious if there is a better alternative in local environment. As cloud instances costs so I was exploring some local alternatives for that.


#4

Alternative in local environment: Build your own box.

Here’s a lengthy thread from fast.ai forums where people ask questions, share what components they are using, and post other useful links and tips. The cheapest new NVIDIA GPUs are around $300 and others paying more for more powerful GPUs. A few of fast.ai’s students wrote blog posts documenting how they built their machines:


#5

The cloud instances are a good place to start when you’re exploring the field. Once you’re sure, then you might want to shift to an economic alternative.

AMD doesn’t support the libraries that we use for DL as of now.
The AI Saturday communities will be using GCP, which offers 300$ as a signing up credit (for every gmail account used). I think that might be good place for you start at.

Otherwise if you’re a student, you can try the AWS Educate program for 150$ credit.