I’m curious if anyone has had much luck leveraging older AMD hardware to use ROCm, I have an 6700 xt that I’ve just begun inquiring about, and it seems it falls outside of official support.
Right now I intend to pass it through to my Debian Docker VM to support transcoding in some containers in addition to machine learning applications.
You must log in or # to comment.
I run it on a 6650xt just fine. I have to explicitly set what version I want, but no issues.
You should be in a better spot with a 6700xt.
I had it running on my Vega 64. But it had to be exactly one specific version of ROCm. Been a while since I’ve played around with that so I don’t remember the specifics.
Just use Vulkan backends.
Kind of a tangent : this depends a lot on your use case but I’ve found that transcoding with GPU is not necessarily a good thing. You generally get larger files, and it’s not always faster than CPU. That is because ffmpeg can distribute the load among all your CPU cores. If you’ve got enough of those you’ll get better multipliers than on an old GPU.
You need to set an override in your environment variables to force it to use the gfx1030 kernel modules, but otherwise you shouldn’t have too many issues.
It’s unofficial, but the 6700xt uses the exact same core as one of the supported enterprise cards, so just using the drivers for it generally works just fine. I use a 6800M personally.
If you are struggling to get rocm installed at all then stop using the amd guides and just install the pre built binaries directly. Fedora packages them in their repository and in my experience rocm just works once you run dnf install rocm*.
Mostly bad, but running AI models on Vulkan surprisingly isn’t any worse than ROCm, so there is little point in using that it seems.
My RX 6600 can use HIP just fine in blender
It’s a lot easier to just use the vulkan support for models and seems to work well enough
My 6700xt works pretty well with nixos, arch, Debian and Ubuntu. Can’t seem to get comfyui to recognize it, but ollama and lama.ccp use it just fine. Just cause it’s not supported doesn’t mean it won’t work. I have an instinct mi25 that I flashed to a wx9100, per this. About 6 months ago, was working great on Debian 12, Ubuntu 20 and 22, nixos and arch. Now, can’t even get it to work on Ubuntu 20 using rocm 5.4. Super sad I can’t leverage the extra vram.
ROCm on my 7900xt is solid. ROCm on my MI50s (Vega) is a NIGHTMARE
and here I thought the instinct vega line was within supported scope, no wonder nvidia is eating amd’s lunch
I have the rx6650xt and managed to make it work in NixOS, although there is some environment variable you have to set on ollama-rocm iirc. GFX something
I had it working on a 5700xt a couple years ago
Yeah, it technically works, but requires telling ROCm you have a 6800 XT instead
God after buying an amd machine last year I’m never doing it again.
What are you trying to use rocm for? Their own guides don’t work.
A couple of years ago, I bought a Radeon laptop, just to see how things were going with ROCm. It worked out ok, but convinced me not to buy a discrete Radeon GPU for my desktop.
I decided to buy another Nvidia card, and finally start investing in NVDA.
Their guides specifically call for an exact kernel version, distribution, and hardware. If you are trying to operate outside of the official requirements then it shouldn’t come as a surprise when the official documentation doesn’t work for you.
do you know how insane it is their official guides don’t work with kernel point updates?
https://github.com/ROCm/ROCm/issues/5824
This has been an issue for a long time.
I have to maintain a file of which specific kernel+os+firmware versions I’m on and have downgraded to just to get the most popular ML library in the world to du a matrix multiply.
I don’t get how this bug gets into production branch, let alone shipped requiring firmware downgrades, on their new line of GPUs/chips. How do they not test their latest hardware with their own firmware?
mostly it is the hardware I have on hand; first project in mind is ROCm machine learning for immich. after that it’s pretty much trying to understand the technology, I’m sure I’ll come up with something fun.
I don’t know how the immich ml works, but if you’re going LLMs stick to llama.cpp.
going beyond that, I’ve had serious kernel bugs with PyTorch and onnx that are still unresolved. The most popular ML/AI frameworks basically don’t work due to drivers for me.
Vulkan flows are fine and generally comparable in speed so far, so if there’s a vulkan option try rock first then revert to vulkan.
thanks for the heads up, in truth I’d probably be headed to vulkan now if it were compatible with immich. I’ll put llama.cpp on my radar.
@roundup5381 I have at ROCM 6.4.2 working on my RX6600XT with 8GB well. Still having had the time to upgrade to ROCM 7.x.
Works well with the #ubuntu packages on my #linuxmint desktop.
it has me considering moving my docker containers over to something ubuntu flavored, that ubuntu is specifically supported
I’m using Debian trixie on two systems with (newer) AMD hardware:
ROCm 7.0.1.70001-42~24.04 on an RX 7900 XTX
ROCm 7.0.2.70002-56~24.04 on an AMD AI Max 395+.
