The amount of times I said that only to be quickly proven wrong by the fundamental forces of existence is the reason that’s going to be written on my tombstone.
I think. it’s just path stickiness at this point. CUDA works and then you can ignore it’s existence and do the thing you actually care about. ML in the pre CUDA days was painful. CUDA makes it not painful. Asking people to return to painfully…
Long term, I expect Vulkan to be the replacement to CUDA. ROCm isn’t going anywhere…
We just need fundamental Vulkan libraries to be developed that can replace the CUDA equivalents.
cuFFT->vkFFT(this definitely exists)cuBLAS->vkBLAS(is anyone working on this?)cuDNN->vkDNN(this definitely doesn’t exist)At that point, adding Vulkan support to XLA (Jax and TensorFlow) or ATen (PyTorch) wouldn’t be that difficult.
The amount of times I said that only to be quickly proven wrong by the fundamental forces of existence is the reason that’s going to be written on my tombstone.
I think. it’s just path stickiness at this point. CUDA works and then you can ignore it’s existence and do the thing you actually care about. ML in the pre CUDA days was painful. CUDA makes it not painful. Asking people to return to painfully…
Good luck…
Yeah, but I want both GPU compute and Wayland for my desktop.
Hybrid graphics. Works for me.