Speaker
Luis Chamberlain
(Samsung)
Description
We have different GPU Linux drivers, but for AI we just need GPU compute. Do we really need different GPU drivers for GPU compute? Could we learn from what NVMe folks have done for GPU compute? Could we collaborate on a unified GPU compute driver? What would that look like? In lieu of any proposals out there I have put one together for folks to review and I'd like community discussions around this. The folks on CC are required for the session.
https://github.com/mcgrof/unified-gpu-compute-driver
Things to discuss:
- Why would we do this?
- Is it possible, could we do it?
- A path to it
Primary author
Luis Chamberlain
(Samsung)