- cross-posted to:
- [email protected]
- cross-posted to:
- [email protected]
LibreCUDA is a project aimed at replacing the CUDA driver API to enable launching CUDA code on Nvidia GPUs without relying on the proprietary CUDA runtime. It achieves this by communicating directly with the hardware via ioctls, (specifically what Nvidia’s open-gpu-kernel-modules refer to as the rmapi), as well as QMD, Nvidia’s MMIO command queue structure. LibreCUDA is capable of uploading CUDA ELF binaries onto the GPU and launching them via the command queue.
Nvidia’s grip on CUDA needs breaking, one way or another.
Nvidia can keep CUDA, we just need libraries to be written with the option to use Intel or AMD - I can’t speak for AMD but it generally takes 10 minutes to port libraries that explicitly use cuda in pytorch to intel oneapi so it’s not even that hard to do, I can’t speak for AMD though