[LeanderK]

lol, this is so stupid. Don't they realise that people usually develop locally and train on a server? You don't need a super beefy GPUs to do it, so you buy nvidia. So people are used to nvidia, debug and fix bugs etc. It's not a very smart decision, looks like the decision makers have no idea what's going on.

[paulmd]

AMD’s opencl runtime also historically has an incredible number of bugs and paper features that make any sort of portability between nvidia and AMD cards quite difficult - you are running a special build and a different codepath for AMD anyway, there was no gain from using the ostensibly portable approach.

[latchkey]

The gain is that you can't find time on NVIDIA cards right now. Decentralizing away from a single provider lowers the risk on your business significantly.

Case in point, OpenAI closed new signups because they couldn't keep up with demand and they literally have all the resources in the world to make things happen.