Very bad performance on AMD 6600 XT

[IcyFoxe]

Hello, not sure if anybody ran into this issue, but I'm having very slow on my AMD RX 6600 XT GPU.
It takes over 4 seconds to do 1 iteration on 512x512 image generation. (At first I thought this was normal, until I've seen a videos of people generating an image in 5 seconds).

I'm running on latest drivers, Windows 10, and followed the topmost tutorial on wiki for AMD GPUs. Only thing I had to add to the COMMANDLINE_ARGS was --lowvram, because otherwise it was throwing an error, that there's not enough memory (even though the GPU has 8GB of memory).

I'm not sure what to do anymore. I searched anywhere for this issue, but I can't find anything.

---

Edit:
When using --medvram instead of --lowvram, it results in ~1.5s for 1 iteration. Slightly better result, but still not what I would expect.

Another note / question: Even when using --medvram, I immediately receive an error, when trying to create images larger than 512x512px. Is 8GB not enough memory, or is it an issue related to my GPU?

[kainta]

It's memory hungry for sure. Pair that with the current lack of support for AMD it leaves a lot to be desired. Im on 6700 12GB and getting 0.4s/it. About 8 seconds to generate 512x512 20 steps.

Lack of VRAM seems to be the biggest hog.

By the way there is a big discussion thread on AMD here.

[GrgMdmn]

I don't understand... I am on a rx 6800 and I almost have 3s/it... :(

[ApoorvRChincholkar]

On a 6800 non-xt you should be getting around 7.3it/s with these conditions

1. You arerunning Automatic1111's or Vladmandic's SD WebUI (not any other branch, or fork).
2. Running this on Linux.
3. ROCm has been installed and configured correctly.

On Windows or on the DirectML fork, you won't get good performance.

[IcyFoxe]

Is this problem related only to AMD GPUs? Meaning if I had NVidia GPU with the same VRAM, would I be able to achieve better results?

[fractal-fumbler]

[IcyFoxe]

Ok, so based on that graph AMD GPUs seems to be pretty bad for this. (Best I managed to achieve was 1.5IT/s with --medvram --opt-split-attention, not sure what settings did they use).

I ordered GTX 3060 12GB model, so I'm curious what results will I get and how helpful will the extra 4GB of VRAM be.

[fractal-fumbler]

@IcyFoxe, also try --upcast-sampling and --opt-sub-quad-attention

[IcyFoxe]

Ok, I just did and adding to the commands above, that I currently use --upcast-sampling slowed the speed and --opt-sub-quad-attention is causing an error. So no luck there with those commands on my AMD card.

[fractal-fumbler]

it's either --opt-split-attention or --opt-sub-quad-attention
--upcast-sampling should slow speed, but allow to generate images using fp32 calculations via upcast

[jwvanderbeck]

Do you need to do something in SD after upgrading or changing your GPU? Like remove some caches or something that may be invalid now?

I'm asking because today I "upgraded" from a 3090 to a 3090ti and now SD is, no joke, about 5-10 times SLOWER than it was, and looking at perf monitor the GPU usage is very low.

[fractal-fumbler]

why do you ask in topic related to amd gpu problems? :)

ot: maybe need to recreate venv folder

[IcyFoxe]

The performance on the Nvidia GTX 3060 12GB is 6 times faster and it can handle producing much larger images than my AMD GPU could.
(Though sucks that the GPU I bought has such loud fans, even on the lowest RPM - MSI I trusted you, and you let me down like this?! Anyway, I'll replace it for a different brand, but basically if you're interested in AI, Nvidia is currently a way to go.)

[circle1-23]

That's cause windows does not support ROCM, it only support linux system.
You're using CPU for calculating, not GPU. that's why that slow.
If you don't want to use linux system, you cannot use automatic1111 for your GPU, try SHARK
tomshardware graph above shows under SHARK, which calculate under vulkan. Not native ROCM.
If you use linux, I'm using RX6600 right now, it takes about 4s for 512x512 20step eular a image

[circle1-23]

Yesterday following your guide I was able to use the GPU to create images, I put --share for the Gradio link but when trying to generate an image in the public link it stopped working and put no interface, but the local link worked so I don't know what It happens, but today when trying to use it, this message appears again in the terminal

https://pastebin.com/14JxJcW4

I don't know what to do, it started giving me errors after I changed demo.launch in the webui.py file (so that it gives me the public Gradio link instead of using --share I tried to use that and then no, that didn't work , I have already deleted and I have followed your guide about 3 times but it no longer works to generate images with the GPU, but the first time it did work, I await your response, thank you.

I don't know what actually happening on your PC... The message says no GPU found...
did you check the ROCM installation? the error occurs when proper ROCM environment is not set
try sudo rocminfo on terminal and do usermod commands again.
what version of mint are you using? the guide is made for ubuntu 22.04, which is same as mint 21
If you're using old branches, the command for ROCM installation is different.

also, the guide is saying pytorch 2.0 update, which may occur several errors, so don't do that. ROCM 5.4.2 + pytorch 1.13.1 works fine.

[circle1-23]

Yesterday following your guide I was able to use the GPU to create images, I put --share for the Gradio link but when trying to generate an image in the public link it stopped working and put no interface, but the local link worked so I don't know what It happens, but today when trying to use it, this message appears again in the terminal

https://pastebin.com/14JxJcW4

I don't know what to do, it started giving me errors after I changed demo.launch in the webui.py file (so that it gives me the public Gradio link instead of using --share I tried to use that and then no, that didn't work , I have already deleted and I have followed your guide about 3 times but it no longer works to generate images with the GPU, but the first time it did work, I await your response, thank you.

And if --share option is not working, don't change launch.py
https://bbs.ruliweb.com/userboard/board/700315/read/3194?search_type=member_srl&search_key=5308509
this is another guide for making share link. It'd better try --ngrok or --listen
I'm now using --listen so link is not changing

[Arion107]

I'm using Linux Mint but it's probably better that I switch to ubuntu and it's solved because I really don't know what's going on

[Arion107]

I deleted Linux Mint, installed Ubuntu, I followed step by step, when you put sudo rocminfo I also have RX6600 and I get the same as you, but after everything and already put the model, when I run webui.sh I still get the same error

https://pastebin.com/79TxwShS

I don't know what I can try anymore

[circle1-23]

I deleted Linux Mint, installed Ubuntu, I followed step by step, when you put sudo rocminfo I also have RX6600 and I get the same as you, but after everything and already put the model, when I run webui.sh I still get the same error

https://pastebin.com/79TxwShS

I don't know what I can try anymore

I reinstalled today and find out that original clone of stable diffusion is not working now
you need to retouch the webui.sh file.
I've renewed the guide, you need to change the pytrorch version now,
change
export TORCH_COMMAND="pip install torch torchvision --extra-index-url https://download.pytorch.org/whl/rocm5.2"
into
export TORCH_COMMAND="pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/rocm5.4.2"
and you also need to install libstdc++-12-dev manually

[ApoorvRChincholkar]

6600 XT runs pretty slow on windows (tiger's directML fork). I hardly get 1it/s speed there.
Whereas on Ubuntu, running on Rocm and the Automatic1111 or Vlad ui, 4.8it/s - 5it/s is very easy.

Check the results here by searching for 6600XT https://vladmandic.github.io/sd-extension-system-info/pages/benchmark.html

Installation of Rocm was difficult for me though. But after that it runs pretty fast. Tom's hardware benchmark is now pretty dated.

[Bartabok]

This card is so terrible on Stable Diffusion :(
I had a MSI GeForce GTX 1050 Ti 4G since years. I was running SD without any problems... it was slow ok, but it was working. I was doing 512x768 images.

A friend gave me a Radeon RX6600 8G yesterday. Double VRAM and it's supposed to be a much faster/more powerful card... ok. nice ! (i thought).
SD crash saying not enough VRam for same images i was doing before and all i can do now is very very slow 512x512.

what a hell :(

[ApoorvRChincholkar]

It's faster than 1050ti for sure, but only if you run Ubuntu and have proper Rocm installed. Outside of this setup it fairs really bad.

[Bartabok]

I'm on Windows 10.. not Linux... and yes. it's so bad on it.

[ApoorvRChincholkar]

True, don't even bother running it on Windows. It was an extremely frustrating experience.

[devlumba]

Hello!

Is it alright now? Almost a year passed, was it made better?

[ApoorvRChincholkar]

Not really :(