news.volyx.in

LibreCUDA – Launch CUDA code on Nvidia GPUs without the proprietary runtime (github.com)

643 points by rrampage · 731 days ago · 145 comments on HN

Article summary

LibreCUDA is a project that aims to replace the CUDA driver API, allowing users to launch CUDA code on Nvidia GPUs without relying on the proprietary CUDA runtime. It achieves this by communicating directly with the hardware via ioctls and Nvidia's MMIO command queue structure. The project is in its early stages and currently implements only rudimentary CUDA functions, making it not yet ready for production use. LibreCUDA provides a way to upload CUDA ELF binaries onto the GPU and launch them via the command queue.

Main themes

  • CUDA alternative
  • Nvidia GPU support
  • Open-source development
  • GPU computing
  • Proprietary software

What commenters say

  • The development of LibreCUDA could potentially lead to the creation of a corresponding library for AMD hardware, promoting open-source GPU computing.
  • Some commenters believe that the project solves the problem of relying on proprietary Nvidia software, while others think it does not address the issue of bloated official packages.
  • There are concerns that large corporations may be able to embed licensing checks and other restrictions at the firmware level, potentially limiting the use of open-source alternatives like LibreCUDA.
  • The project's ability to launch CUDA code without the proprietary runtime is seen as a step towards breaking the CUDA lock-in and promoting open-source GPU computing.
  • Some argue that using Vulkan Compute is not a viable alternative to CUDA due to its limited semantics and the need for a proprietary Vulkan driver.
  • Others believe that the development of LibreCUDA and similar projects is crucial for promoting open-source GPU computing and reducing dependence on proprietary software.
  • The discussion also touches on the topic of AI and machine learning, with some commenters expressing concerns about the potential for corporations to use AI for their own interests and restrict access to certain information.
  • There are differing opinions on the potential impact of LibreCUDA on the CUDA/CUDNN dependency version nightmare that comes with running various ML libraries.