This morning someone with an RTX 5090 posted a fix on r/comfyui. MiniMax H3, image to video, a 20-second clip at 0.8 megapixels. It ran out of memory. They had installed PyTorch for CUDA 12.8, the way they had a month earlier. Other people in the community told them to reinstall it for CUDA 13.0 and update ComfyUI. After that, a 20-second clip at 0.98 megapixels, larger than the one that failed, rendered without trouble.
That is one person on one machine, and we have not reproduced it. The reason they gave is easier to check, and it holds up.
What ComfyUI does on CUDA 12.8
ComfyUI ships with a library called comfy-kitchen, a set of fast GPU routines for the compressed model formats most people now run (fp8, nvfp4, int8) and for a few operations that every modern image and video model uses. It has a fast CUDA version of each routine and a slower plain PyTorch version as a fallback.
When ComfyUI starts, it checks which CUDA your PyTorch was built for. If it is older than 13, it switches the fast CUDA version off and uses the fallback for everything. It also prints a warning, in capitals, saying you need PyTorch for CUDA 13.0 or higher, and that on a 20 series card or newer this is required.
The README says the same thing, and comfy-kitchen’s own prebuilt files need CUDA 13.0 and an NVIDIA driver from the 580 series or later. None of this is new. It has been in ComfyUI since at least August. What catches people is that it fails quietly: ComfyUI keeps working, just without the fast path.
Why people end up on 12.8
When the 50 series launched, CUDA 12.8 was the first PyTorch build that supported the cards, so well over a year of guides, forum answers and install scripts told 5090 owners to use it. Plenty of them are still up. If you followed one and never revisited it, you are probably on 12.8 now.
Whether the fallback costs memory on every workflow, as well as speed, we have not measured. On that one 5090 with H3 it was the difference between a clip that rendered and one that did not.
How to check yours
Three ways, quickest first.
Scroll up through the console window when ComfyUI starts. If you see the warning about needing PyTorch with cu130 or higher, you are on the slow path.
Or look a little lower in the same startup text for the line starting Found comfy_kitchen backend cuda. On a correct install it says 'available': True, 'disabled': False.
Or ask PyTorch directly, using the same Python that runs ComfyUI:
python -c "import torch; print(torch.__version__, torch.version.cuda)"
If that prints 12.8 or anything below 13, the fix is the one in ComfyUI’s README: reinstall PyTorch from the cu130 index, and update your NVIDIA driver first if it is older than the 580 series. The desktop app and the portable build have their own update routes, so use those if that is how you installed.
Ours
We checked before writing this. The Python environments behind our ComfyUI instances are all CUDA 13.0 builds, the driver is well past 580, and the startup logs show the fast CUDA path enabled. So this was not our problem, and we cannot tell you from our own renders how much it costs you.
We can tell you it costs you something, because ComfyUI says so every time it starts.
A daily note from our news radar. The week gets the full treatment in the DIY AI Brief every Monday.
