easydiffusion
Multi GPU performance bug
I love the multi-gpu support with Easy Diffusion but I've noticed something. While it does work across all the cards, the performance doesn't actually really increase. The AI rig that I have uses four 4090s.
When I use a single card this is the iterations per second that I get:
When I use all four cards, I only get about 1/4th of that on each card:
So effectively the same performance but spread across all the cards.
This test was done with the exact same model, prompt, and image settings, etc. I get the same result no matter what models, prompts, or image settings I use
Im using a fresh install of EasyDiffusion. Linux Kernel: 5.14.0-362.24.1.el9_3.0.1.x86_64 Distro: RHEL 9.3 Nvidia Driver: 550.54.15 NVML: 12.550.54.15
I would expect that when I am using multiple GPUs that I would see an increase in performance with each card added. I did a test of 100 images using SDXL with 1 card and 4 cards, and the difference in time length was only about 5%.
Is there anything I need to manually adjust in the source to get this to work properly? Are there any log files or other debug output that I can get you to help resolve this?
