#Extremely high perf/latency on AMD card

1 messages · Page 1 of 1 (latest)

tawny mountain
#

AMD Sapphire Pulse 9070 XT
Windows 11 home

Earlier in the year I upgraded my entire machine from Intel/Nvidia(I7-10700/RTX3070) to an AMD(7600x/9070xt) machine. It was a fresh install and everything, so no leftover drivers there. I reinstalled deiteris' W-okada with the directml version and things were absolutely fantastic for a few months. I went from 250-300ms latency to 80ms with 2.7s extra processing time in both instances. I quit using it for awhile as I had to focus on work, and after trying to use okada again the performance... well... had become absolutely abysmal. It's worse than my RTX 3070 at this point, jumping all the way up to 300-500ms of latency. I had done absolutely nothing to my settings so I couldn't figure out why.

After a bit of experimenting and trying a few other things I only seem to have made the problem worse...

Installation method: One thing I noticed in deiteris' and TG's github pages are their instructions... Specifically the known issues section where it states "AMD Radeon RX 7000 series may be unable to achieve low latency (below 256ms)."

Things I have tried:
-Reinstalling Deiteris' W-Okada directml fork
-Installing TG's W-Okada directml fork
-Forcing FP32 mode
-Converting models to ONNX
-Using different models
-Lowering processing time to 0.3s (This was the only thing that seemed to remotely help. Yet, it's still very bad.)
-Switching the device to CPU and back to my GPU

Things that have changed recently:
-Installed a bunch of games
-Updated AMD's drivers
-Installed comfy UI-Zluda and it's dependencies. Quite possibly down/upgrading python.
-Ran a few virus scans

Figured I'd ask before I even consider fucking around aimlessly some more. Am I blind? Am I missing something so blatantly obvious that I'll feel like a troglodyte after it's been pointed out? Is AMD cursed? Any suggestions or ideas are welcome.

At the very least, thanks for taking the time to stop by and read this!

tawny mountain
#

After changing what device W-okada uses I have found the following.
-0:Amd Radeon(TM) graphics (directml) has a low perf rate of 50ms. However the voice that is output sounds terrible, like it's in a can.
-1: AMD Radeon RX 9070 XT (directml) has an extremely high perf rate, and for whatever reason it seems to be using my CPU's IGPU.

I'm going to make the bold assumption that W-Okada isn't using my GPU here. I'll try to have it use my 9070 xt by disabling my igpu and leave an update here.