I've seen a lot of claims about AMD performance, but not enough hard numbers. Most benchmarks are still on NVIDIA. I ran some quick SD 1.5 txt2img tests on my setup.
My config:
* Card: RX 7900 XTX (24GB)
* OS: Linux (Arch)
* Backend: Automatic1111 with `--use-directml` flag (ROCm was more hassle than it was worth)
* Settings: 512x512, Euler a, 20 steps, batch count 4, batch size 1
Results:
```
Average time per image: 2.8 seconds
Total for 4 images: 11.4 seconds
VRAM usage: ~7GB peak
```
For comparison, I found an old result from a 3090 (24GB) with similar settings:
* Average time per image: 1.9 seconds
The AMD setup is slower, but the gap isn't as huge as I expected. Has anyone done more rigorous tests with SDXL or different schedulers? I'm particularly interested in seeing if the newer `torch-directml` builds change anything.
- bench_beast
Benchmarks don't lie.