Hacker News Reader

Comment by wronglebowski | Hacker News Reader

No ROCm at launch? After they delayed it for months? What a joke. That's like not having CUDA available at launch.

https://www.phoronix.com/news/AMD-ROCm-RX-9070-Launch-Day

This card is marketed to consumers, not AI developers. Most people who will buy these cards don't even know what ROCm is.

Edit: it just doesn't matter for launch day. It'll likely be supported eventually.

Every Nvidia card has CUDA support from day one, regardless of who the market is for it. I wouldn't much as much if all their slides weren't covered in AI, AI, AI, and it didn't ship with some stupid LLM chatbot to help you tune your GPU.

I think having CUDA available in consumer cards probably played a big part in it being the de facto "AI framework." AMD would be wise to do the same thing.

Why can't consumers dabble in AI?

When I explained my non-technical friend who is addicted to ChatGPT that she could run models locally her eyes got lighten up and she wants to buy a graphics card, but she is not doing any gaming.

Developers introduction to a technical ecosystem is more than likely using the card they already own. And "It'll likely be supported eventually" just signals that they're not serious about investing in their software. AMD is selling enough CPUs to finance a few developers.

> not AI developers

Yet their slides show INT8 and INT8 with Sparsity performance improvements. As well as "Supercharged AI Performance".

Those features are probably used by FSR4. The 9070 isn't very attractive for LLMs though.

The slides mention Stable Diffusion and Flux as a benchmark.

I mean that's one way to guarantee availability for gamers!

The ROCm libraries were built for RDNA 4 starting in ROCm 6.3 (the current ROCm release). I'm not sure whether that means it will be considered officially supported in ROCm 6.3 or if it's just experimental in that release.

Im running ROCm ok on my 9070XT. You can build it from source today if you have a card.

rocminfo:

**** Agent 2 **** Name: gfx1201 Uuid: GPU-cea119534ea1127a Marketing Name: AMD Radeon Graphics Vendor Name: AMD Feature: KERNEL_DISPATCH Profile: BASE_PROFILE Float Round Mode: NEAR Max Queue Number: 128(0x80) Queue Min Size: 64(0x40) Queue Max Size: 131072(0x20000)

[32.624s](rocm-venv) a@Shark:~/github/TheRock$ ./build/dist/rocm/bin/rocm-smi

======================================== ROCm System Management Interface ======================================== ================================================== Concise Info ================================================== Device Node IDs Temp Power Partitions SCLK MCLK Fan Perf PwrCap VRAM% GPU% (DID, GUID) (Edge) (Avg) (Mem, Compute, ID) ================================================================================================================== 0 2 0x73a5, 59113 N/A N/A N/A, N/A, 0 N/A N/A 0% unknown N/A 0% 0% 1 1 0x7550, 24524 36.0°C 2.0W N/A, N/A, 0 0Mhz 96Mhz 0% auto 245.0W 4% 0% ================================================================================================================== ============================================== End of ROCm SMI Log ===============================================

Can't Vulkan backends do the job? Not to defend AMD, but so long that perf/dollar stays above NVIDIA just any how, isn't that more than bare minimum effort for them?

I am not surprised, ROCm focuses on CDNA. Was RDNA in ROCm ever officially supported?

My 6900 XT has had pretty good support in everything I've tried.

But it didn't launch with AMD support, it is support now.

RX 7900 (RDNA 3 flagships) are officially supported. I think the 6900 (RDNA 2) was supported at one point as well, and in both cases other cards with the same architecture could usually be made to work as well.

Do you assume that it working means official support? I had been under the impression that RDNA had unofficial ROCm support.

I've been using my 6750XT with rocm ever since I picked it up ~2 years ago. Idk if it's "officially" supported but it's works with SD and LLMs just by setting an environment variable.

On Linux, there is official support for a few Navi 21 (gfx1030), Navi 31 (gfx1100) and Navi 32 (gfx1101) cards [1]. As you mention, in practice, there's also quite a few cards that work but are not officially supported.

[1]: https://rocm.docs.amd.com/projects/install-on-linux/en/docs-...

Thanks for the link. Those are the professional cards. For consumer cards, the only supported models on Linux are the 7900 series and Radeon VII according to that. Interestingly, all of the supported cards aside from Radeon VII have 24GB of RAM or more as if high VRAM were a prerequisite for support.

Seeing this, I am not surprised I misremembered the state of RDNA support in ROCm. The support matrix it pitiful compared to Nvidia’s CUDA support matrix.

The official support list is anemic, but in practice, any RDNA 2 or RDNA 3 GPUs will either work out of the box or can be made to work without much difficulty.

Debian has a particularly good compatibility story. Its ROCm packages carry patches to work on all discrete Vega, RDNA 1, RDNA 2 and RDNA 3 GPUs. That's not official support, but it is tested.

https://rocm.docs.amd.com/projects/install-on-linux/en/lates...

Looks like on Linux the only GPUs which are currently officially supported are the 7900 series.

https://rocm.docs.amd.com/projects/install-on-windows/en/lat...

On Windows the official support covers most (all?) of the discrete desktop 7000 and 6000 series.

Thanks for the clarification.

Reply on news.ycombinator.com