Best optimization for AMD Cards?

[RandomLegend]

Hey Guys,

so i just switched from my RTX3070 to an RX7900 XT and got it all running just fine.

However i coulnd't really improve my speed compared to the Nvidia card. It got up by approx. 10% which is fine but i believe there could be more to have?

I stumbled across these posts for automatic1111 LINK1 and LINK2 and tried all of the args but i couldn't really get more performance out of them. Those topics aren't quite up to date though and don't consider stuff like ONNX and ZLUDA.

How do you guys optimize your AMD cards? Is there some newer workflow i could go for?

Thanks!

[likelovewant]

For amd， I guess zluda is the speed favorite way. AMD have already implemented Rocm on windows, with the help of ZLUDA, the speed quite boosted. Even many GPUs not officially supported ,doesn't means they are never been worked. AMD had those code available on GitHub. Eg, Roclabs and tensile, follow the official guide and some tweaks, anyone can easily recompile the rocblas .ie ,the library file to fit rocm on windows or Linux. Then to achieve the unofficial support Rocm way of speeding.

[RandomLegend]

Yeah i have no idea on how to actually set up all this stuff... For ZLUDA i heard it's a drop in replacement but my auto1111doesn't work with it.. so idk

[likelovewant]

sure, There are several AUTOMATIC VERSIONs ,NO all support zlude , you can use the this one "git clone https://githubfast.com/lshqqytiger/stable-diffusion-webui-directml.git" or https://github.com/vladmandic/automatic/ (SD.NEXT) .Which have the zluda support otherwise , you had to tweek the code to make it through .There is a tutorial on youtube , google it .7900xt is official support rocm ,so it more easier .Give it a try

[RandomLegend]

The directml version of lshqqtiger is ~25% faster for me but i can't get sd next running with ZLUDA.

Only tutorial i found was this https://www.youtube.com/watch?v=n8RhNoAenvM but i am on linux and on the lshqqtiger zluda github is only installation tutorial for windows but not linux...

[likelovewant]

yes, that's the one ,The sd.Next guide is here https://github.com/vladmandic/automatic/wiki/ZLUDA, except the tutorial delete the part which replace the two dlll in vevn fold, The sd code can ran both windows or linux , I guess the step is same . Have not try it .

[JonathanDiez]

Ey, how u can use SD with AMD? I have this problem

[likelovewant]

This is forge version ,not support zluda so far .