News for hackers. Stuff that matters.

A comment on Run CUDA, unmodified, on AMD GPUs. Parent comment

Re:Run CUDA, unmodified, on AMD GPUs by rjurney on Monday July 15, 2024 @09:21PM (#40972924)
Not having a layer like this has left AMD completely out of the AI game that has made NVDA the world's most valuable company.
[ Parent | on HN ]
Re:Run CUDA, unmodified, on AMD GPUs by ChoGGi on Tuesday July 16, 2024 @02:53AM (#40974162)
Self-inflicted wounds hurt the most.
[ Parent | on HN ]
Re:Run CUDA, unmodified, on AMD GPUs by HarHarVeryFunny on Tuesday July 16, 2024 @03:10PM (#40979340)
Well, they kinda have it with their hipify tool, although this is for porting CUDA code to AMD's HIP which supports both AMD and NVIDIA. This supports CUDA C code and libraries with AMD equivalents like cuDNN, cuBLAS, cuRAND, but doesn't support porting of CUDA C inline PTX assembler. AMD have their own inline GCN assembler, but seem to discourage it's use.

There are also versions of PyTorch, TensorFlow and JAX with AMD support.

PyTorch's torch.compile can generate Triton (OpenAI's GPU compiler) kernels, with Triton also supporting AMD.

[ Parent | on HN ]