Comment by mattnewton

people are trying, especially for inference. For training, it’s just too high risk to tank your training I think.

TPUs are at least dogfooded by Google deepmind, no team AFAIK has gotten the AMD stack to train well.