Comment by felipetrz

As an AMD user, I have to say that it's nowhere near nvidia for ML. Basically ALL ML tooling is written with CUDA as a target, and while ROCm promises some level of compatibility, it's a buggy mess that crashes with most models out there.

If OpenCL took off as the standard, any GPU would be able to do everything just fine, but unfortunately industry always takes the "worse is better" approach to everything.