I can tell you don't work with these cards. The MI300x is bug-ridden junk at the moment. As are all AMD cards. You cannot use them for serious compute.

Very weird to say something like that. Mine, and all the others who I know who have them, are being used for serious compute.

You use jax and pytorch to develop new models on an MI300x without encountering encountering random bugs? I'm shocked.

So what I hear you saying in this response is that you do not work with these cards, yet you have some strong opinions.

The only bugs we have run into so far are the lack of virtualization and the documented ones in the AMD rocm docs.

Before buying H100s we tested MI300s and they were borderline unusable.

If you're just running some popular pretrained models out of the box, sure.

But you're saying that you develop new model architectures on them? And it works reliably? That's not at all what we saw.

A lot has changed in only a few months. ROCm has gone through several revisions in a short amount of time. I'm curious where you did your testing.

By the way, this was just done...

https://www.reddit.com/r/AMD_MI300/comments/1dgimxt/benchmar...