AMD CEO Teases ROCm Support Coming to Radeon Consumer GPUs Soon | Hacker News Reader

AMD CEO Teases ROCm Support Coming to Radeon Consumer GPUs Soon

wccftech.com

15 pointsmindcrime8 comments

Great to hear about it again. I thought ROCm was in half abandoned limbo. Wonder what happened to the architecture issue of shipping a binary for every GPU and code size balooning...

That's still a thing. There's a tradeoff between ahead of time compiling for the specific architecture and shipping some more generic thing that you maybe try to specialise at the last moment. Some of the tooling now bundles the N different binaries in a single archive file which makes it less annoying in practice.

I thought ROCm was in half abandoned limbo.

Hmm... for all of the (very valid) criticisms of ROCm, I don't think I've heard anybody suggest that it's "half abandoned". From what I've seen, AMD have been working on it regularly and it seems to be pretty strategic to them. They just published a new documentation site, and the various Github repos get commits more or less constantly.

Of course one can quibble about the value of the output of all of this activity, but to their credit, they appear to be giving a serious effort here.

Guess I have gotten the wrong impression when I tried to deploy a project requiring some GPU horse power on a bit of FreeBSD infrastructure. There was no support for it in BSD land and all the issues were years of maybes. Now that I had a second look you are right.

> I thought ROCm was in half abandoned limbo.

They have to join the AI race so it's back from the dead.

I don't think it was ever abandoned. Had a reasonable size team on it for at least the last five years. Just very focused on HPC for most of that. I believe it's running nicely on Frontier.

It looks like there have been some dead ends on the path, e.g. HSAIL doesn't seem to be a thing any more. Also looks like making it run on older generations of graphics cards was hard work which might be why the supported card list was mostly the instinct line.

There's already support for like the 6950 xt & one other consumer card.

This is good news though. It needs to expand a lot. But there's so many other big needs ROCm has, so much quality/stability improvement necessary. And CUDA has such a huge huge range of libraries that AMD has so much yet to reimplement.

Developing on windows is zero fun, not totally happy to see that part of the announcement.

Compiling the libraries for the various consumer GPUs sounds good though. It definitely annoys people that they have to build from source (and tread the "officially unsupported" waters) when using consumer cards.

Reply on news.ycombinator.com