5700xt 8gb vram 2 minutes for one picture, is that normal? (25 steps)

[chymace]

My commands are; --opt-split-attention --opt-channelslast --medvram --always-batch-cond-uncond --opt-sub-quad-attention --no-half --precision full --no-half-vae --disable-nan-check

is that normal or am i missing something?, i get 1.90s/it to 2.90/s it, sometimes it goes to 3s/it but lets say avarage is 2.5s. and best case scenario is 1.90.

windows 10, x64.
25 steps, 2m karras, upscale by 1.15, hires 3

[some9000]

To judge that you really need to mention resolution. Here a GeForce 2060 with 6Gb of RAM and all the possible improvements active does 5-6 iterations per second at the default 512x512 resolution. So that's a couple of seconds for one image at 20 iterations.

Upscale, by default as minimum doubles the processing time. But is it even worth it at 1.15?

[set-soft]

I have an RX5500XT with 8 GB. I usually use options to save RAM at a small performance cost.
A 512x512 image without optimizing VRAM usage takes around 14 s (not the first run after starting the server, this run takes 3 to 4 minutes while HIP compiles the computing kernels). A 512x768 image with VRAM optimizations takes between 24 an 27 seconds for 20 steps "Euler a", and I can batch 6 of them, which improves the speed.
But this is on Linux using HIP (ROCm). I saw many people complaining about bad AMD performance on Windows, where ROCm isnt avavilable.

[set-soft]

Oh, BTW! If you have poor performance avoid doing extra stuff during generation. Disable upscale and/or face restoration.
Apply them separately only to the best images.