Nvidia RTX4090 and AMD 7900XTX benchmark

[Ryuukeisyou]

I made a simple benchmark test for the two gpus on my computers. Hope this can help anyone who is interested.

7900XTX is about 57% of RTX4090 at low resolution and batchsize (512x512x1),
yet only achieves 33% at higher resolution and batchsize (768x768x4).
would get even lower to 29% if RTX4090 uses xformers (tested but not listed below).

And so I wonder what would be the cause? Would anyone have any ideas?

Here are some details.

Nvidia RTX4090

cpu: Ryzen9 7950x, 16core-32thread, 4.5GHz to 5.7GHz
mem: 64GB (32GB x 2) 6000MHz
os: Windows 11 23H2
pytorch: 2.1.2+cu121

AMD 7900XTX

cpu: EPYC 7F52, 16core-32thread, 3.5GHz to 3.9GHz
mem: 64GB (32GB x 2) 2666MHz
os: Ubuntu 22.04
pytorch: 2.3.0.dev20240118+rocm5.7

Common

checkpoint: v1.5-pruned-emaonly.safetensors
positive prompt: 1girl
steps: 100
sampler: euler
args: --dont-upcast-attention --use-pytorch-cross-attention
seed: run random seed 5 times to get average

Results

average it/s

average pixels*it/s

[Dampfinchen]

How can I install Xformers in Comfy UI? I am using the standalone package. I've tried to install it via pip install but it still doesn't use it. Help would be appreciated.

[jbeougher7924]

There are so many reason why this will not give you accurate results. Neither system config is the same. Rerun your test with the same workstation but only change out the video card and see what the results are. Have each GPU perform the same test at least 3 times and average the results.