# RX 6800 (RDNA2): 3x faster images in ComfyUI (int8 convrot) — retained record

> Retained reddit record from reddit:comfyui.

Canonical: https://highsignal.app/data/records/7d2690b4d2e08cef

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:comfyui)  published 2026-11-11 00:59 ingested 2026-11-12 02:42

[original source ↗](https://www.reddit.com/r/comfyui/comments/1xt4690/rx_6800_rdna2_3x_faster_images_in_comfyui_int8/)

## retained text



```
I was not happy with my RX 6800 speeds on newer diffusion modes like Qwen Image 2.1 for example and decided to do something about it using Deepsek V4 Flash 1017. Personally I don't know much details about what is going on here, but gains are insane, so I thought this might be useful for RDNA2 owners to keep their cards alive longer. Buying new shiny NVIDIA in current economy is just rough.

We with Deepseek been working on ComfyUI speed on an RX 6800 (ROCm, Linux) and got about 3x on images.

Qwen Image 2.1 (int8 convrot), 1024x1024, same seed, per image, text encode and VAE included:

* base, 25 steps: **180 s → 60 s**
* turbo, 8 steps: **70 s → 20 s**

How to use it

`git clone -b rdna2-all` [`https://github.com/KornienkoDev/ComfyUI.git`](https://github.com/KornienkoDev/ComfyUI.git)

`cd ComfyUI`

*# set up your venv and install requirements as usual, then:*

`pip install --force-reinstall --no-deps \`

[`https://github.com/KornienkoDev/comfy-kitchen/releases/download/rdna2-int8-simt/comfy_kitchen-0.2.41-cp312-abi3-linux_x86_64.whl`](https://github.com/KornienkoDev/comfy-kitchen/releases/download/rdna2-int8-simt/comfy_kitchen-0.2.41-cp312-abi3-linux_x86_64.whl)

The second step matters: the fast INT8 matmul lives in a separate package, so the branch alone only gets you the attention fixes. The wheel covers gfx1030/1031/1032 (6800 / 6800 XT / 6900, 6700, 6600).

Start it with --disable-mmap --disable-pinned-memory --highvram.

What's in it

1. INT8 was switched off on these cards. ComfyUI asked the GPU whether it could do INT8 math, got a "no" back, and quietly used a slower path. It can do it, so now it does.
2. Attention was moving a lot of data around every block for no good reason. Fixed, and it no longer wastes that time or memory.
3. The INT8 matmul that shipped was built for cards with special AI hardware. Ours don't have that, so Deepseek wrote one that runs on the regular shader cores.
```



document fetched 2026-11-12 00:43 · canonical https://reddit.com/r/comfyui/comments/1xt4690/rx_6800_rdna2_3x_faster_images_in_comfyui_int8
