Skip to content
Search

LibreCuda

lastmod 2024-08-16created 2024-08-161 edits

Run CUDA code on Nvidia GPUs without the proprietary runtime

Features

  • A project that aims to replace the CUDA driver API so CUDA code can run on Nvidia GPUs without depending on the proprietary CUDA runtime
  • Talks to the hardware directly through ioctls (specifically what Nvidia's open GPU kernel modules call rmapi) and QMD, Nvidia's MMIO command queue structure
  • Can upload CUDA ELF binaries to the GPU and launch them through the command queue
  • Current features
    • Allocating and freeing GPU memory, and mapping memory so the CPU can access it
    • Can upload CUDA ELF binaries to the GPU
    • Launching CUDA kernels through the command queue
  • MIT license

GN⁺'s summary

  • LibreCUDA is a project that lets you run CUDA code on Nvidia GPUs without the proprietary CUDA runtime
  • Only basic features are implemented so far, and it is not ready for real-world use
  • A project with similar functionality is ROCm, which provides CUDA-like functionality on AMD GPUs

Hacker News comments

  • LibreCUDA – Launch CUDA code on Nvidia GPUs without the proprietary runtime | Hacker News
  • The point of an open CUDA is to be able to run it on GPUs other than NVIDIA's
    • Meaningless if you still have to buy an NVIDIA GPU
    • With real competition, it would be easy to buy devices with more VRAM
    • Someday we'll be able to run Llama 405b locally
  • Limited experience with CUDA, but wondering whether this would help with the CUDA/CUDNN dependency version issues that come up when running various ML libraries (e.g. TensorFlow, ONNX)
  • Very nice
    • This is all I want from a CUDA runtime
    • Should make it possible to run the LLVM libc unit tests
    • This could justify an AMD library that uses the same direct system call approach
  • For projects that depend on CUDA, moving from LibreCUDA to HiP should be the first step toward platform freedom
  • Wondering what a CUDA ELF file is
    • Wondering whether it is binary SASS code and whether an open-source PTXAS alternative is needed
  • As a CUDA beginner, wondering what problem this solves
  • Seems like a missed chance to call it CUDA Libre
  • Wondering whether you still have to run the proprietary NVIDIA graphics driver, or whether it's entirely unrelated
  • The authors should start thinking about trademark infringement notices
  • Since ZLUDA was taken down at AMD's request, a ZLUDA replacement is needed as a general way to break the CUDA dependency
    • Something not tied to NVIDIA hardware is needed

See also

Favorite site

author Zer0 sitemap