For those of us doing any ML that requires CUDA/CuDNN. It feels that AMD (specifically for GPU) still isn't a viable option at this moment in time. I sincerely hope that this changes.
While I agree with the current status quo, I don't think that it is too hard to break the vendor lock-in here. Most ML people code against pytorch or tensorflow/keras APIs, so as long as AMD is a viable backend for these frameworks, people won't really care. But whether and when this will happen is another story.
AMD apparently has a patent to include specific ML-related chips on their GPU cards. Maybe they intend to strengthen their position there.
Unfortunately, there's far too much inertia to ever change back, unless they can absolutely crush CUDA-based performance somehow.
Hopefully OneAPI just makes this whole conversation irrelevant.