I started this part of the thread and mentioned Trainium but the person you replied to gave a link. Follow that and you can see Amazon's chips that they designed.
Amazon wants people to move away from Nvidia GPUs and to their own custom chips.
TBH I was just going off of that I've heard AWS is a terrible place to get h100 clusters at scale. And for the training I was looking at we didn't really want to consider going off CUDA.
Huh? That’s quite the assertion. They provide the infrastructure for Anthropic, so if that’s not large scale idk what is.
They have to use GCP as well, which is arguably a strong indictment of their experience with AWS. Coincidentally, this aligns with my experience trying to train on AWS.