Disclaimer: I used to work for AWS ML Buried in the press release is the announcement that there will be a push to bring up models on the AWS Trainium accelera

Disclaimer: I used to work for AWS ML Buried in the press release is the announcement that there will be a push to bring up models on the AWS Trainium accelerator. In the name of "ML democratization" this agreement could help shift tides away from the Nvidia HW + SW monopoly. Amazon (trn1) and Google (TPU) are little guys compared to Nvidia. The book may already be written. Pytorch is heavily biased towards libraries that are customized for cuda/Nvidia (see FSDP). This leaves the other HW folks playing SW catch-up, with everyone rallying behind the open source XLA compiler. It's too bad the press release didn't touch on these points and used cliche rhetoric instead.