RTX 5090's speed for GenAI — Flux1.dev, Stable diffusion 3.5 and HunyuanVideo tested — results are amazing

How fast is the RTX 5090 running text to image or video?

I was luck to be able to buy an RTX 5090 at MRSP (although ASUS’s ‘MRSP’ plux some import tax is alreadyat scalper level). Many people are interested in using RTX 5090 for text to image or video tasks. I did some quick benchmarking on Flux1.dev, Stable diffusion 3.5 Large and HunyuanVideo. It also has some comparisions with previous generation flagship cards RTX3090 and RTX4090.

If you can not see the full post due to paywall, please see the free video at

[embed]

Here’s the PC spec:

The CPU is a low end CPU, but the diffusion based models are mostly using GPU, so it’s powerful enough.

For installing of Ubuntu driver for 5090 or other 50 series card, see my guide here.

Flux1.dev

The workflow used is from official ComfyUI flux example page.

There are two versions:

flux1-dev-fp16:

flux1-dev-fp8: fp8 quantization reduces the checkpoint size, saves vram but slightly slower and slightly lower output quality.

Stable diffusion 3.5-large

The 3.5 large has less parameters than Flux.1-dev, so it’s slightly faster than Flux1.dev. Speed is 2.52 it/s.

HunyuanVideo

It uses lots of VRAM and GPU time for each generating. There are also two versions: fp8 and bf16. The speed are similar, but fp8 uses less VRAM.

fp8:

bf16:

Comparison with 3090

Here’s a quick summary:

Comparison with 4090

For Flux1.dev, the ComfyUI log is following

loaded completely 14841.083572097778 11350.067443847656 True 100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:12<00:00, 1.55it/s] Prompt executed in 13.69 seconds 4090

So compared with 5090, 5090 is about 40% faster.

Conclusion

The RTX 5090 is really good at text to image/video generating. Feel free to let me know what do you want to test next!

Thank you for reading!