Almost anything; they won’t GPL their driver like they ought to, because they made unwise decisions surrounding the use of material covered by software patents and non-Free licenses, and thus are unable to do so.

Aren't they re-architecting & creating a new, free software driver that's only responsible for memory management & conveying requests from the user through to to the _real_ computer running on the GPU?

Don't clutter up this thread with actual facts and information.

Unlikely. AI projects are tied to Nvidia not the Linux. They just port their driver to FreeBSD or Solaris or whatever. The AI crowd will follow. But they can also just release patches that as part of the driver install patches the end user's kernel to remove this janky stuff.

i'm out of the loop on this one. i did a good deal of HPC -- specifically electromagnetics simulation -- about a year ago and Spir-V worked OOTB for GPGPU on AMD processors. hell, i wrote my kernels in _Rust_ and they compiled and ran just fine (Embark Studio's `rust-gpu` target), which is about as edge-casey as i can imagine.

how is AI stuff tied to nvidia? what would make it difficult to port (or to be portable)?

AFAIK the vast majority of ML training/inference on the GPU runs on CUDA, and has done so for many years, while other vendor-specific GPU backends are slowly gaining support.

It doesn't help that AMD's equivalent open toolchain, ROCm, is barely even supported by their own GPU lineup.

FWIW, ROCm GPU support is getting better, particularly in the distro-provided packages. e.g. https://salsa.debian.org/rocm-team/community/team-project/-/...