Super slow generation with RTX 3060

[Mashkiki]

Hey, if people are getting ~14it/s with RTX 3070, why is my 3060 juicing only 0.85it/s? It's over 15x slower and I still don't know why. What are some checks I could do and maybe possible fixes?

RTX 3060 12GB | Ryzen 5 5600X | 16gb 2666hz
pyTorch 2.0+cu118 | xFormers 0.0.18
512x512 px | 20 Steps | CFG 7 | DPM++ SDE Karras

[Sakura-Luna]

The information you provided is not enough. At present, I can only tell you that the SDE sampler is slower, and Euler a is twice as fast as him.
Post your log and the screenshot at the bottom of the WebUI, and I can reply you more.

[valden80]

I have the same type of card, 3060 with 12GB vram, 32Gb memory, i5-8400
On similar image generation parameters, speed of samplers on my system:
DPM++ SDE Karras - ~3.5it/s (7 sec to full image generation)
Euler a - ~7it/s. (4 sec to full image)
DPM++ 2M Karras - 7it/s (3.8 sec to full image)

Default setup from this repo on Manjaro Linux, no upgrading to Torch 2.0 or anything, just clean setup about 2 week ago.

[CeciliaXCIX]

This is what I'm getting on my system. I have a 3070 though, but maybe it can give some insight?

RTX 3070 | Ryzen 7 3700x | ram 16Gb
python: 3.10.9 | torch: 1.13.1+cu117 | xformers: 0.0.17.dev464

8.5-9it/s - 512x512 px | 20 Steps | CFG 7 | Euler a
8.5-9it/s - 512x512 px | 20 Steps | CFG 7 | DPM++ 2M Karras
4.2it/s - 512x512 px | 20 Steps | CFG 7 | DPM++ SDE Karras

[JPyle]

Hey I'm running into your exact problem and am lost, did you ever find a fix ?

[undrh2o]

is everyone using a batch size of 1?

I have the same problem here 3060 12g, 32g RAM, i9-9900k, windows 11, I've turned off GPU acceleration in windows and Chrome.

A fresh clone of master which includes pytorch 2.0.1, and nothing else added. I get 1.2it/s running 512 x 512, 20 steps, euler a, , no negative prompt and only "cat" as a prompt. I only get 5.8it/s with a batch size of 1, with a batch size of 8 i only get 1.2 it/s

I even tried command line args below as suggested elsewhere with no change
--opt-sdp-attention --medvram --always-batch-cond-uncond --opt-channelslast

[Sakura-Luna]

Not sure what you're trying to express, it's a given that increasing the batch size will slow things down.

[undrh2o]

I'm just trying to get an idea what batch size people are using for the sake of comparison. if it's a batch size of 1 then my 6-7 it/s seems like about right based on the Tom's hardware article I found.

[Sakura-Luna]

This conclusion is obviously wrong. According to the speed shared by others above (size1 9it/s), your size8 1it/s is very normal.

[wzgrx]

The recent updates have been experiencing this issue, with a random decrease in speed of 5-10 times，Previously, a 960 resolution image took 30 seconds, but now it takes 5-10 minutes

[BundoBytes]

This thread has brought to my attention that I've been getting low performance as well on my 3060 12GB.
I found that if I remove the "--medvram" commandline argument and hide my webui tab (by opening a new tab or minimizing the browser completely) my generations went from ~1.47it/s to ~3.6it/s when generating 512x512, 20steps DPM++ SDE, and from ~3.3it/s to ~6.8it/s when generating 512x512, 20steps Euler a.

I kept these commandline arguments: --opt-sdp-attention --opt-channelslast
Using Automatic1111's May 14 commit, torch 2.0.1+cu118, python 3.10.9 in a Docker environment. My web browser has HW acceleration disabled (so I can get more VRAM :P).

[locke12456]

Back in April, I encountered the same issue as you did, but no matter how much I tried to make changes, nothing worked. So, I gave up and decided to use an older version instead. These past few days, I decided to give it another try and found that the same issues persisted.

The problems I faced back then were twofold: first, I couldn't achieve a resolution of 1920x1280, and second, the performance was extremely slow. To tackle these problems, I used the following parameters:

set COMMANDLINE_ARGS= --xformers --lowvram --medvram --precision full --no-half --no-half-vae --disable-nan-check --opt-split-attention-v1 --upcast-sampling --opt-channelslast --theme dark --autolaunch
set PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.6,max_split_size_mb:128

After testing, I confirmed that these settings were effective, but it resulted in extremely slow processing times, taking about 20 minutes. Therefore, I decided to revert to the parameters I have been using all along, and I ran the test again using the following parameters:

set COMMANDLINE_ARGS= --medvram --xformers --force-enable-xformers --always-batch-cond-uncond --opt-channelslast --no-hashing --disable-nan-check --api --xformers-flash-attention --opt-split-attention --no-half-vae
set PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.6, max_split_size_mb:32

This time, the results were similar in terms of speed to what I experienced with the older version.
result:
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 3346099050, Size: 1920x1280, Model hash: 61e23e57ea, Denoising strength: 0.35, Clip skip: 2, ENSD: 31337, Version: v1.4.1

Time taken: 1m 4.92sTorch active/reserved: 11238/11582 MiB, Sys VRAM: 12288/12288 MiB (100.0%)

[GloriHumer]

Hello, guys,
I got the Stable Diffusion, and a bunch of models, tried to make some AI pics, and noticed, that my GPU (RTX 3060 laptop) doesn't get activated at all, and that the sampling takes too long, and the final result looks worse, than the same prompt on my friends PC.
Can anybody help?

[StudioDUzes]

set CUDA_VISIBLE_DEVICES 1 or =1 in webui-user.bat ?