#ROCm on Windows + RDNA2 (RX 6700) — anyone tried this?

1 messages · Page 1 of 1 (latest)

astral plover
#

Hey everyone — quick question for anyone who’s messed with ROCm on Windows.

I’m running an RX 6700 (RDNA2, 10GB) on Windows 11 and I’m trying to see how far ROCm can realistically be pushed on Windows, even if it’s unofficial / hacky.

I know ROCm is Linux-first, but Windows is a hard requirement for me right now. I’m mainly looking at local LLM inference, and I’m totally fine with:

WSL2

Forcing env vars (like HSA_OVERRIDE_GFX_VERSION)

Compiling from source

“Mostly works” setups if performance is decent

Has anyone here actually gotten ROCm + RDNA2 working under WSL?

Which ROCm versions behaved best?

Is this worth the effort compared to Vulkan or DirectML on Windows?

Any gotchas specific to RX 6000 cards?

Not expecting anything officially supported — just trying to learn what’s actually possible before I give up or switch paths.

Appreciate any firsthand experience 🙏

wary flame
# 
pip install --index-url https://rocm.nightlies.amd.com/v2-staging/gfx103X-dgpu/ rocm[devel,libraries] torch torchaudio torchvision
rocm-sdk init
rocm-sdk test
#

see what happens

#

it might be useable, can't vouch for performance but when windows gets ck I think it should be on par with linux, probably a little slower.

#

I would try on native windows first and see what happens

uncut igloo
uncut igloo
# astral plover Hey everyone — quick question for anyone who’s messed with ROCm on Windows. I’m...

@astral plover hey! just want to add some context here:

**official support status: **RX 6700 (RDNA2 / gfx1031) is not on the official Windows ROCm support list right now. native PyTorch via ROCm currently only supports RX 7000/9000 series + some Ryzen AI APUs. source: https://rocm.docs.amd.com/projects/radeon-ryzen/en/latest/docs/install/installrad/windows/install-pytorch.html

that said, @wary flame's approach using the gfx103X-dgpu nightly builds is the community workaround path — might work, no guarantees on stability/perf

**on ROCm versions: **if you do try this, 7.10.0 preview might be more stable than 7.1.0. we've seen reports of HIP error: illegal memory access on 7.1.0 under high memory load / batch processing (mainly RDNA4, but worth noting). see: https://discord.com/channels/1239631572886491286/1435355465067266130

wary flame
#

Do not be discouraged if it takes an absurd amount of time to run every time you encounter a new shape. That is MIOpen doing it's job, just badly. You will encounter this problem with every non Instinct gpu, regardless of os. There has been a lot of screaming about it for months so surely it has already been fixed and said fix just hasn't landed yet. I put an example patchy fix over the weekend in an issue I opened in rocm-libraries, which should take no time at all to fix properly. So reasonably one would expect to see it fixed today...

(I try my best not to be passive aggressive, but sometimes I can't help it even if I realize I'm doing it.)

uncut igloo
#

@wary flame good point on the MIOpen warmup — that's definitely something OP should expect on first runs with new shapes. do you have a link to that rocm-libraries issue? might be helpful for others hitting the same thing

wary flame
#

More helpful for devs really