Re:Hugging Face and AWS partner to make AI more accessible
by locknessness on Tuesday February 21, 2023 @05:09PM
(#34888301)
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.
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.