ROCm on RX570 (gfx803)

McLink · Posted: Thu Oct 12, 2023 5:53 pm Post subject: ROCm on RX570 (gfx803)

I've been trying to get ROCm to work on my RX570 (I want to use it for machine learning in PyTorch), but while the card is detected, any real-life tests seem to fail, usually with a segfault. For example, the vadd_hip example mentioned by wgo:HIP gives this output: http://0x0.st/H4fz.txt followed by Segmentation fault (core dumped).

Installed roc* packages:

grknight · Retired Dev Joined: 20 Feb 2015 Posts: 1662

I have very poor knowledge on OpenCL, but for my RX 550 to work properly, I had to install dev-libs/amdgpu-pro-opencl
Don't know if it helps here.

McLink · Posted: Thu Oct 12, 2023 6:14 pm Post subject:

Ahh, the proprietary one. The whole reason I buy AMD cards is to avoid proprietary drivers, so I'd very much like to avoid that. If I can't get this card to work with the FLOSS driver, I think I'd rather get a new one (besides, this one is kind of low on memory for ML anyway).
_________________

depontius · Advocate Joined: 05 May 2004 Posts: 3509

It might be even worse than that. I've had similar problems with my RC 5500 based card. (Navi14 - gfx1012) I've managed to build the ROCm stuff, but I've never been able to actually DO anything with other than some opencv tests. More recently I tried to start building some other things like rocBLAS or pytorch, and none of it builds. I haven't had a lot of time to spend fiddling with it, however.

Then just the other day I saw on Phoronix about ZLCUDA - that will allow you to run Cuda binaries unchanged on an AMD card. Then I started looking into it harder. As of ROCm 5.7 the only supported cards are in the 7900 series. Not just only the newest generation, but only the top-end of the newest generation. I did manage to find some patches to get Navi14 running with a newer ROCm, but I'm not sure how new and they warn that not everything will work.

I saw others saying that ROCm was a waste of time about a year ago, but at the time hadn't looked much into it. Honestly, I still don't need it. I bought the hardware anticipating heading that direction and built some software necessary to get there. But as I said, I had mixed results and didn't have time to pursue.

At this point my next efforts into scientific computing won't need the GPU anyway, so I'm going to bide my time. When I get to the point of actually needing GPU computing I'll take another look at the market. At this time I'm expecting a GPU purchase will be nVidia and it's Cuda time. But we'll see.

STUPID STUPID STUPID moves on AMDs part. At the very least, today when I was looking at Cuda the API is versioned, so that it's documented what will work on what generation of cards.
_________________
.sigs waste space and bandwidth

McLink · Posted: Wed Feb 14, 2024 12:17 pm Post subject:

FWIW, I ended up buying an RX 7800 XT for Christmas, and ROCm works fine on that. The downside is that I get occasional crashes with this card, even when I'm not using ROCm, but I guess that's just a matter of waiting for the graphics stack to stabilise. The biggest issue on the machine learning end is that if I inadvertently fill up the GPU memory, the system instantly hard-locks, so I have to be very careful with that.
_________________

depontius · Advocate Joined: 05 May 2004 Posts: 3509

When I built my current computer I called it "Retirement Computer #2". I planned to dabble in scientific computing after retirement and wanted a capable machine. I made the best hardware decisions I could back in the Fall of 2020. I may not have been careful enough on my choice of GPU and may not have looked carefully enough at AMDs support statements. However the currently supported GPU didn't even exist back then, so it seems pretty clear that "supported" is a moving target.

Almost a year into retirement I've still been too busy to take up doing my own scientific computing. That which I've already done has been CPU based anyway and my system has been fully adequate to the tasks. I was thinking of "Retirement Computer #3", but it makes more sense to wait until my current system hits the wall, then see what's available. I've also been more closely formulating some things I want to try in the realm of scientific computing, and so far none of it really offloads into the GPU - yet. For me this can all wait.
_________________
.sigs waste space and bandwidth