Very low it/s on a RTX 3090

[AkeiZelionne]

Hey!
I recently bought a RTX 3090 to create faster images on Stable Diffusion
However I kind have bad it/s. It's better than my older GPU (RTX 2080 super) but I got 3.5 it/s using SDXL model, 30 sampling steps, 720*1280 7 CFG scale and no hires fix or anything

All my specs should be good, I have a AMD RYZEN 9 3900X 12-core, 32GB ram, Seasonic Focus GX - 750W and a Asus TUF X570 PLUS GAMING

I run SD with : api --no-half-vae --skip-torch-cuda-test --xformers --opt-split-attention --theme=dark
I also tried without but seems there is little, if not any differences
I made a clean installation as well to make sure everything is up to date

Is this normal or should I have much higher it/s?

[w-e-w]

take my numbers with a grain of salt

3.5 it/s seems to be about the same for XL 720x1280
I also use a 3090

theres a database of performence
https://vladmandic.github.io/sd-extension-system-info/pages/benchmark.html
for refrence SD1 512x512, for 3090 it should be about 20 it/s
irrc the test are all done at 512x512

---

sometimes you could squeeze out some performance if you increase your batch size

---

generally you should not be using --no-half-vae
if you use it for xl is because of NaN then use the fp16 nan fix vae form
https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl.vae.safetensors

[AkeiZelionne]

Thank for your answer!
You seems right, I tried with a SD 1.5 models on 512x512 and got like 17 it/s, it's not 20 but oh well haha
I'm mostly using SDXL now since I got better results, I thought I had low it/s but based on your informations I should be good then

Increasing batch size does not seems to increase my it/s, it decrease it instead
about no half vae I used it because I found in online haha... but I don't remember having "NaN" issues with my renders, so I think I'll get ride of it!

[w-e-w]

Increasing batch size does not seems to increase my it/s, it decrease it instead

of course it doesn't increase the raw it/s, the it/s is not important
the important thing is the actual work / time
batch size 2 process 2 images at once
so a 5 it/s of batch size 2 is essentially should be compare to 10 it/s of batch size 1

and often when you increase the batch size you can squeeze out a bit of performance

example: if at batch size 1 you get 10 it/sm at batch size 2 you can may get 6/its, this is a +20% performance

note there are other things that are not counted into the it/s

example not using actual numbers