Well the problem is all of the packages and libraries you can find related to ML and AI have been programmed keeping in mind an NVIDIA GPU. And this was not overnight as well.
They introduced CUDA in 2006 and encouraged development on it then finally AlexNet happened in 2012 and that’s where everything took off.
They introduced CUDA in 2006 and encouraged development on it then finally AlexNet happened in 2012 and that’s where everything took off.
NVIDIA doesn’t have good support on Linux and it’s due to their drivers being closed source so all the Linux users prefer AMD GPUs but still they haven’t found a replacement for the kind of performance NVIDIA is able to provide for training models.
If any company were to be motivated enough to research on a replacement for CUDA, it would take years before it would be feasible (that’s not even a guarantee) and the R&D cost would be insane. Intel OneAPI and AMD ROCm are two examples of companies trying.
One can hope for a miracle and it could happen that there’ll be a competitor in the future so let’s see. But US is hell bent on keeping their lead on AI hence why they are putting such sanctions on the export of NVIDIA GPUs.
