In case others are interested on the general compute aspect, e.g inference for self hosted AI, here is something related I found :
- ZLUDA, for CUDA everywhere, https://github.com/vosen/ZLUDA/ unfortunately in the process of a major rewrite https://github.com/vosen/ZLUDA/discussions/261 due to AMD (somehow?!) not wanting their code being used in there (maybe they did some shady RE on NVIDIA work?)
- AMD ROCm https://github.com/ROCm/ROCm and HIP https://github.com/ROCm/HIP
Well... a distribution IS a selection of packages and a way to keep them working together. Arguably the "only" innovation in that context is HOW to do that and WHICH packages to rely on. For the first, the "latest" real change could be considered immutable distributions, as on the SteamDeck, and declarative setup, e.g. NixOS. For the second... well I don't actually know if anybody is doing that, maybe things like PrimTux for kids at schools in France?
Anyway, I agree but I think it's tricky to be innovative there so let me flip the question, what would YOU expect from an innovative distribution?