Increasing speed on AMD RX 6600XT. Also, please share the ur speed information!

[hvdxxx]

Hello everyone! My English is very bad, so I will use a translator!
Computer specifications:

• CPU: Ryzen 7 5700x3d
• GPU: AMD RX 6600XT (undervolted to 1100 mV, 2449 MHz because during generation the hotspot is >91 degrees Celsius)
• RAM: 3200mhz
• OC: Windows 10

WARNING: The speed was increased from an average of 2-3 s/it to 1.7 s/it in 832x1280, depending on the used LoRA your speed will drop, also changing the generation settings affects the first generation!

The installation of Forge was carried out according to https://github.com/lshqqytiger/stable-diffusion-webui-amdgpu-forge

HIP SDK5.7 was installed and ROCmLibs 5.7 was used. From the tests I found out that the optimized library in the instructions is inferior to the made by likelovewant. You can also try to build your own assembly and try it out.

The following arguments were used to launch on webui-user.bat:
set COMMANDLINE_ARGS= --zluda --opt-split-attention --no-half-vae --disable-nan-check
set SAFETENSORS_FAST_GPU=1
set CUDA_LAUNCH_BLOCKING=1
I highly recommend experimenting with arguments!

In Settings:

• Negative Guidance minimum sigma = 0.7
• Token merging ratio = 0.5
• Show live previews of the created image = none
  I highly recommend experimenting with settings!

These settings gave the following result in:
512x512, Euler A, Clip Skip=2, Sampling Steps=20:
avg it/s = 2.7

768x1024, Euler A, Clip Skip=2, Sampling Steps=20:
avg it/s = 1.2

832x1280, Euler A, Clip Skip=2, Sampling Steps=20:
avg s/it = 1.7

To further increase the resolution, I recommend using Img2Img with upscale

In conclusion, I want to say that I will switch to NVIDIA... ZLUDA better than directML, i also want to know what is your generation speed guys? No matter what GPU or OS, tell us so we can compare!