Benchmarking Stable Diffusion on my AMD GPU

Motivation

I have been using stable diffusion since it released qutie regularly and last few months I have seen massive improvement on my AMD GPU, both in terms of performance and size of images that I can generate locally.

The tom’s Hardware Benchmark for stable diffusion has pretty much become a one of the go to places to check Stable Diffusion performance of a GPU. I feel it has become really out of date espcially when considering AMD GPU’s so much so that the performance of my GPU is more than twice than as shown in the graph.

This is just me trying to show to show the performance improvement that has occurred on AMD GPU’s in the period of time.

Methodology

The tests were conducted on the latest rocm/pytorch docker container (as of 9 September 2023) , which comes preinstalled with pytorch version rocm5.4.2 and python version 3.8.16.

On this docker container ComfyUI was installed, using which the tests were then conducted.

First 1 image generation was queued, the result of which was discarded, following this 10 images were added to the queue whose result was then averaged.

The Efficiency Nodes for ComfyUI module was used for the purpose of this test. (click here for module github)

Image prompt Details

These are basically a copy paste of what tom's Hardware used for their benchmark.

Positive Prompt:

postapocalyptic steampunk city, exploration, cinematic, realistic, hyper detailed, photorealistic maximum detail, volumetric light, (((focus))), wide-angle, (((brightly lit))), (((vegetation))), lightning, vines, destruction, devastation, wartorn, ruins

Negative Prompt:

(((blurry))), ((foggy)), (((dark))), ((monochrome)), sun, (((depth of field)))

Steps:

100

Classifier Free Guidance:

15.0

Sampling Algorithm:

euler_ancestral

Workflow File

The workflow file can be found by clicking here

Computer Specs

1
2
3
4
5


GPU: AMD Radeon RX 6700 XT 12GB VRAM
CPU: Intel Core i7-12700
RAM: 64GB 3200MHz
OS: NixOS 23.05
Kernel: 6.4.1-zen1

Results

The results are an average of 10 runs, conducted on ComfyUI.

No Overclocking

Model	Dreamshaper_8 (SD_1.5)	DreamShaper_XL (SDXL)
Image Size	512x512 px	1024x1024 px
No Flags	4.78 it/s	1.31 s/it
Flag: `--dont-upcast-attention`	6.58 it/s	1.04 s/it

My try to Overclock GPU myself with CoreCtrl

Model	Dreamshaper_8 (SD_1.5)	DreamShaper_XL (SDXL)
Image Size	512x512 px	1024x1024 px
No Flags	5.17 it/s	1.22 s/it
Flag: `--dont-upcast-attention`	7.10 it/s	1.04 it/s

This was the first time I overclocked anything in my whole life and I do think there is room for improvement in the results. There were runs where I got much better speeds (~7.35 it/s on Dreamshaper_8) but they were not sustainable and usually resulted in errors crashing ComfyUI.

Conclusion

While the 6700 XT is not the most top of the line GPU, it does hold itself very well in terms of both price and performance, I would put it between the RTX 3060ti and RTX 3060 in terms of performance. This is a great leap in performance especially when you consider that NVIDIA gpus haven’t really gained much or anything in terms of performance.