Zluda performance

[Omen090909]

Hi, i managed to use zluda but performance is basically exactly same as i had with directml
Am using rx 6700, and have the same performance also in sd next
What are your experiences with zluda guys?
The only upside for me is better memory management, but it still seems limited compared to what happenes on nvidia gpus

[Eminic]

Rx 6700 XT here, with zluda i get around 3.7 seconds/it with an SDXL model. Compared to rocm on linux, it's about 3X slower. Same performance in SD Next. Personally sticking with linux until it approaches anywhere near that performance.

[Omen090909]

thanks, that's good to hear

[Soulreaver90]

I bought a spare SSD for $50 and installed Linux on it (Ubuntu 22.04). Been getting the full power of my 6700xt and native support of web-ui. Honestly it isn’t that hard and it’s worth it since most of these developers focus on Linux first.

[Omen090909]

i had a spare ssd for it and tried it, and well, the lack of customisation for gpu kinda killed my interest, and i can't run it stock cause it's gonna burn. I might still try it, maybe find a workable sollution, but so far been quite disappointed with how linux works

[Eminic]

There's an program called core ctrl for linux which is very similar to msi afterburner in terms of functionality, controls frequencies, power delivery and you can make custom fan curves as well. I've been using it myself since my gpu also melts on stock settings.

[Omen090909]

thanks, i ll try it, altho i m sceptical about outside programs to change voltage on gpu, and i don't think i can find tons of footage on yt like i can for everyting on windows :x

[pinea00]

[Omen090909]

idk, i tried deep-cache in sd next, speed might be doubled, but quality is halfed
also yes, sd next is super unstable for me too, can't run sdxl at all (i mean i m getting pixels instead of image)

[Bercraft]

I do not know if its the good place but....... here https://github.com/brknsoul/ROCmLibs there are rocm libraries builds.
Ipersonally own a w5700 pro and it works wonders.

[pw405]

First test of Zluda on 7900XTX and uhh.. WOW, blown away!!

With SDXL model AeitherverseXL (https://civitai.com/models/308337/aetherverse-xl)
Generated a base image at 1600 X 768. Params: Steps: 30, Sampler: Restart, Schedule type: Exponential, CFG scale: 7

This was taking ~ 3-4 minutes on DirectML. Just finished TWO images for a total 54 seconds.

Next, tested Ultimate SD Upscale to increase to size 3X to 4800 X 2304. It only took 1 minute & 49 seconds for 18 tiles, 30 steps each! WOW! This could easily take ~8+ minutes or more on DirectML.

So far, ZLUDA looking to be a game changer. GPU runs cooler, it doesn't bog the entire system down and lock up before finishing every image. VERY excited to see where this goes!

[pinea00]

To use forge Zluda and use DPM karrs turbo sampler,you will wow again.

[pw405]

Is that sampler just not available in A1111? What sort of it/second do you get with it?

[pinea00]

RX6600
Generating 10241024 SDXL A1111-zluda (FP8) takes more than 2 minutes
Generating 10241024 SDXL forge-zluda takes 1 minute and 27 seconds
Generating 1024*1024 SDXL forge-zluda (DPM++ 2M SDE Turbo) takes about 43s

[lshqqytiger]

ROCm 5.7 does not support FP8 datatype and operations.

[gigend]

support Radeon RX 570 vram 4gb? how performance, better than directml?

[pw405]

I haven't heard of anybody running Zluda on an R570 - but its worth a shot to try, it can help VRAM use.

For reference - if I'm doing big image with SDXL - my VRAM allocation would always be maxed out with DirectML, but with ZLUDA I only see about ~12 GB of VRAM allocation.

Also be sure to use Scaled Dot Product in the optimizations settings within the Web UI.

[patientx]

ROCm 5.7 does not support FP8 datatype and operations.

Just came back to your app once again and trying zluda now, it automatically using doggettx , is that correct for zluda and do we need any specific options or commandline parameters other than zluda for an 8gb 6600 , maybe lowvram etc ? Thanks.

[lshqqytiger]

use scaled-dot product, and you don't need lowvram unless you are trying large images