Recommended way to use NVIDIA GPUs and CUDA on Ubuntu Core

[mauricio.pena]

What’s the recommended approach to handle NVIDIA drivers and CUDA applications on Ubuntu Core?

I understand that GPU drivers come through content snaps like nvidia-core22, but how should developers:

• ensure driver/library version compatibility,
• and package CUDA-based applications as snaps?

Any official documentation or working examples would be very helpful.

Thanks!

[astrojuanlu]

Hi @mauricio.pena ! In case it helps, Canonical has recently published silicon-optimized inference snaps that make heavy use of GPU, you might draw some inspiration from it:

• source https://github.com/canonical/deepseek-r1-snap/blob/19d8a05/snap/snapcraft.yaml
• relevant talk at the Ubuntu Summit 25.10 https://www.youtube.com/live/bEEamxJ60aI?si=kgO5L4KeofAnj5N1&t=23843

Does this help fill the gap?

[mauricio.pena]

Thanks for this, the example and the Snap docs are a solid starting point. But I’m still figuring out how to keep the NVIDIA driver snap and CUDA versions in sync on Ubuntu Core, especially for a headless setup. The docs kind of assume you’re on a desktop, so I’ll need to piece together the rest. I’ll check the video too, maybe it’ll fill in the gaps.