Thanks - makes sense.
> Nvidia is the only one that wants vendor-specific code.
Isn't that because CUDA is better and that tight software/hardware integration is more powerful?
If it wasn't then presumably people would be using AMD GPUs for their deep learning, but they're not.
It all comes down to there always being two ways to do things when interacting with a GPU under Linux: The FOSS/vendor-neutral way, and the Nvidia way.
The machine learning crowd has largely gone the Nvidia way. Good luck getting your CUDA codebase working on any other GPU.
The desktop Linux crowd has largely gone the FOSS route. They have software that works well with AMD, Intel, VIA, and other manufacturers. Nvidia is the only one that wants vendor-specific code.