Hacker News Reader

Comment by superkuh | Hacker News Reader

My wishlist for ROCm support is actually supporting the cards they already released. But that's not going to happen.

By the time an (consumer) AMD device is supported by ROCm it'll only have a few years of ROCm support left before support is removed. Lifespan of support for AMD cards with ROCm is very short. You end up having to use Vulkan which is not optimized, of course, and a bit slower. I once bought an AMD GPU 2 years after release and 1 year after I bought it ROCm support was dropped.

View on HN

FWIW, every ROCm library currently in the Debian 13 'main' and Ubuntu 24.04 'universe' repository has been built for and tested on every discrete consumer GPU architecture since Vega. Not every package is available that way, but the ones that are have been tested on and work on Vega 10, Vega 20, RDNA 1, 2 and 3.

Note that these are not the packages distributed by AMD. They are the packages in the OS repositories. Not all the ROCm packages are there, but most of them are. The biggest downside is that some of them are a little old and don't have all the latest performance optimizations for RDNA 3.

Those operating systems will be around for the next decade, so that should at least provide one option for users of older hardware.

I can confirm this, Debian's ROCm distribution worked great for me on some "unsupported" cards.

Packages existing and the software actually working are very different things. You can run rocm on unsupported GPUs like a 780m, but as soon as you hit an issue you are out of luck. And you’ll hit an issue.

For example, my 780m gets 1-2 inferences from llama.cpp before dropping off the bus due to a segfault in the driver. It’s a bad enough lockup that linux can’t cleanly shutdown and will hang under hard rebooted.

The 780m is an integrated GPU. I specified discrete GPUs because that's what I have tested and can confirm will work.

I have dozens of different AMD GPUs and I personally host most of the Debian ROCm Team's continuous integration servers. Over the past year, I have worked together with other members of the Debian project to ensure that every potentially affected ROCm library is tested on every discrete consumer AMD GPU architecture since Vega whenever a new version of a package is uploaded to Debian.

FWIW, Framework Computers donated a few laptops to Debian last year, which I plan to use to enable the 780m too. I just haven't had the time yet. Fedora has some patches that add support for that architecture.

As the underdog AMD can't afford to have their efforts perceived as half-assed or a hobby or whatever. They should be moving heaven and earth to maximize their value proposition, promising and delivering on longer support horizons to demonstrate the long term value of their ecosystem.

Honestly at this point half-assed support would be a significant step up from their historical position. The one thing they have pioneered is new tiers of fractional assedness asymptotically approaching zero.

I mean at this point my next card is going to be an nvidia. It has been a total waste of time trying to use rocm for anything machine-learning based. No one uses it. No one can use it. The card I have is somehow always not quite supported.

We go from:

Support is coming in three months!

This card is ancient and will be no longer developed for. Buy our brand new card released in three months!

Every damned time.

I have a mi50 with 16gb of hbm thats collecting dust (its Vega bases, so it can play games, I guess) because I don’t want to bother setting up a system with Ubuntu 20.04, the last version of Ubuntu the last version of ROCM that supported the MI50 works on.

With situations like this, its not hard to see why Nvidia totally dominates in the compute/ai market.

The MI50 may be considered deprecated in newer releases, but it seems to work fine in my experience. I have a Radeon VII in my workstation (which shares the same architecture) and I host the MI60 test machine for Debian AI Team. I haven't had any trouble with them.

I don’t think the mi60 has reached deprecated status yet (the last time I look at prices for the mi50 and mi60, the mi60 was something like 3x expensive, and I think thats because its still officially supported), but I’ll check this all out. Thanks.

The MI60 is basically just a faster MI50 with more memory. They were deprecated together. It's plausible there could be small firmware or driver differences that cause issues in one but not the other, but I think that's unlikely.

I had the impression Debian applied patches that widen arch support from what upstream officially supports, including for the MI50/MI60.

https://salsa.debian.org/rocm-team/rocm-hipamd/-/raw/d6d2014... (one patch of many)

I wrote that patch. It's not actually used for MI50/MI60 in any of the Debian system packages, since Debian builds for gfx906 rather than using the gfx900 fallback path that patch provides. Debian is not relying on any special patches to enhance gfx906 support. That architecture is the same as upstream.

Now, for some other GPU architectures, you're absolutely right. There are indeed important patches in Debian that enable its extra-wide hardware compatibility.

Thanks for all your work on this.

AMD did over $5 billion in GPU compute (Instinct line) last year. Not nVidia numbers but also not bad. Customers love that they can actually get Instinct system rather than trying to compete with the hyperscalers for limited supplies of nVidia systems. Meta and Microsoft are the two biggest buyers of AMD Instincts, though...

AMD Instinct is also more power efficient and has comparable (if not better) performance for the same (or less) price.

Meta and Microsoft buys hundreds of thousands of Nvidia accelerators a year, and are a big reason why everyone else has to compete for nvidia units.

Seeing Radeon VII on the deprecation list is a little saddening, unless they start putting out more 16gb+ GPUs that aren't overly expensive...

They should have at a minimum 5 year support release cycle.

It kinda seems like they do - 5 years would only include the RX 6xxx and 7xxx.

5 years is not very long tbh.

True but business hardware (and home for that matter) often goes on 3-5 year cycles though. At 5 years it's kinda expected hardware will get replaced.

It doesn’t work for the first three years, so it’s two years in practice.

And that's if you're buying bleeding edge new hardware. If you can't afford to buy it the year of release the window is even smaller.

RX 7800 XT was supported for 15 months before being dropped. Significantly less than 5 years.

AMD has separate architectures for GPU compute (Instinct https://www.amd.com/en/products/accelerators/instinct/mi300....) and consumer video (Radeon).

AMD are merging the architectures (UDNA) like nVidia but it's not going to be before 2026. (https://wccftech.com/amd-ryzen-zen-6-cpus-radeon-udna-gpus-u...)

You can use ROCM on consumer radeon as long as you pay more than 400 dollars for one of their gpus. Meanwhile, you can run stable diffusion with the -lowvram flag on a 3050 6gb that goes for 180 dollars

Reply on news.ycombinator.com