Anomaly Low Performance (AMD)

[AhmadCoffee]

I updated my graphic card to 6700xt and when i tried to generate something generating time was totally not normal (between 50min and 2h) and around 60s/it. On my old 1660ti pics generated around 10 mins (same specs).

I used various models to test and different optimization methods (best result was ~35s/it).
Also i tried different COMMANDLINE_ARGS. I tried --opt-sdp-attention --opt-sdp-no-mem-attention --opt-split-attention --opt-sub-quad-attention and some others. Now it is = --no-half --precision full --skip-torch-cuda-test --medvram. Nothing of it really helped.

Gen info of last completed pic (~50 mins):
city ruins, 1 girl, rooftop, black long hair, blue coat, black pants, iron boots, floating hair, looking back, polar lights, hdr, best quality, masterpiece,
Negative prompt: lowres, (((bad anatomy))), (((bad hands)):1.4), bad faces, text, error, extra limbs, bad proportions, extra faces, missing fingers, missing hand, three hand, extra digit, fewer digits, cropped, (worst quality, low quality:1.4), normal quality, jpeg artifacts, signature, watermark, username, blurry,
Steps: 42, Sampler: DPM++ 2M Karras, CFG scale: 6, Seed: 4221526795, Size: 752x752, Model hash: 2727c13b25, Model: yottaMix_v30, Clip skip: 2

My specs is:
6700XT Sapphire Pulse 12GB (Earlier 1660 ti MSI OC Armor 6GB)
Ryzen 7 2700
Adata RAM 8x4 2666

Sorry if i didn't write something important, i'm almost new in all this theme

[yamfun]

Here's what to do:
Step 1. go search about stuff like AMD stable diffusion Windows DirectML vs Linux ROCm, and try the dual boot option
Step 2. regret about AMD
Step 3. return the card and get a NV card

I got a Rx6600 too but too late to return it. If I can travel back in time for world peace, I will get a 4060Ti 16gb instead

[AhmadCoffee]

I'll stay with AMD anyway, cause old card worked strange and AI isn't main thing i bought it for.
I guess exectly for SD it won't work as fast as NVidia and i can can only try DirectML and hope for AMD optimization?

[yamfun]

Well yeah, and my Rx6600 on DirectML has like 1.5it/s, so yours should be better with the it

[AhmadCoffee]

So SD DML version helped a little bit. Now it takes around 15-20 mins for 1 pic in 700x700 (20-10s/it), but it still to low.
I just don't understand why your 6600 output is 1.5it/s. My performance is just anomaly. And additional CommandArgs except already wroten just make worse.
Or is something with my CPU when it work with 70-95%?

[midnight-god-01]

Ensure all your drivers are up to date. If you've confirmed this, attempt to resolve the issue by performing a disk cleanup. This action will remove outdated driver files. If the problem persists, consider deleting the virtual environment (venv) folder.

Afterwards, restart the web UI. If the issue continues, try reinstalling the Stable Diffusion. After reinstallation, opt running ui with the default settings without any extensions or command line additions.

This will help determine if the problem originates from these additional components. Also, I have noticed you are using (.CKPT) format I would recommend using (.Safetensors) as it loads faster and is a lot safer to use.

Also, you can use different UI and check If they complement your setup better.

[yamfun]

Waaaaat, ppl usually compare speed of 512x512 so I didn't look at your details. What is your simple 512 512 speed?

[peterpernhofer]

I'm not sure if I understand correctly what you wrote. Image-Generation for your samples should be extremely simple and fast. I tried it out with the model "juggernaut_final" instead of your model (I don't have it) and my generation time was:
13.5 seconds (3.28 it/s)

Not minutes! So there might be something wrong with your installation or config after all - even with your old card. My GPU isn't the fastest and most expensive card, but thank goodness for this use case, an NVIDIA:
RTX3060 with 12GB vram. And after enabling the Nvidia CUDA extensions (--xformers) it speeds up about 2-3 times.

[AhmadCoffee]

Thanks everyone for answers. Now i stabilized my performance (512x512 ~2it/s).
Default SD don't work for me without CommandLineARGS (--no-half --precision full --skip-torch-cuda-test).

For optimization i used:
SD with DirectML (https://github.com/lshqqytiger/stable-diffusion-webui-directml),
COMMANDLINE_ARGS= --opt-sdp-no-mem-attention,
Settings > Optimization > Token Merging ratio = 0.3 (or set between 0.2 and 0.5).
Also i tried add MS Oliva to SD DML (nothing changed or i'm just installed it wrong).

And when i tried generate pics like 1200x900 (~2s/it) in my case --medvram crashed almost every time, without this command it worked sometimes.

[Soulreaver90]

I have a 6700xt Nitro+. I am running linux Ubuntu with 512x512 generations at 6.6-6.7 it/s.
If you are on Windows, you'll have to use the DML version of A1111. If you are patient, you can potentially wait for the Windows ROCM drivers to be officially supported in python. If you want to take advantage of the full power of your card, dual boot to Ubuntu 22.04 and run with the ROCM drivers. There are instructions on how to get it running but feel free to hmu if you need assistance.

[AlmostHuman34]

So what did ROCM for windows changed? My performance with 7800 XT is terrible at 1.02-1.05 it/s. What guide should I follow to fix it? I have A111 + ROCM driver, new amd drivers, python and git. What else do I need?

[Soulreaver90]

Nothing, Rocm is one piece of the puzzle. We are still waiting for other components to be updated to utilize Rocm on Windows. There is a tracker for SD.Next, but the same prereqs are needed for A1111.

(vladmandic/automatic#1880)