NVIDIA's Open GPU Linux Kernel Driver Will Soon Be The Default For Turing & Newer GPUs

While we are all waiting for the NVIDIA R555 series Linux driver beta that is expected to debut as soon as next week based on prior information with Wayland improvements (explicit sync) and more, with the NVIDIA R560 series Linux driver successor is a very interesting change: NVIDIA is planning on defaulting to using their open-source GPU kernel driver by default for GeForce RTX 2000 "Turing" GPUs and newer.

Back when NVIDIA unveiled their open-source, out-of-tree GPU kernel drivers for Linux in May of 2022 the initial focus was on workstation / data center GPUs and with time that open-source alternative to their long-used proprietary kernel drivers has been improved. With consumer GeForce GPUs on recent NVIDIA Linux kernel drivers, the open kernel modules perform similar to the proprietary code and in general working well besides a few feature caveats such as some power management differences.

While the NVIDIA R555 Linux driver branch is heading up next in the near term, with the R560 Linux driver series later this year they are planning on switching to the open-source GPU kernel drivers by default for Turing / RTX 20 series and later.

In an updated NVIDIA posting around their feature deprecation schedule they noted today:

"Starting in the release 560 series, it will be recommended to use the open flavor of NVIDIA Linux Kernel Modules 9 wherever possible (Turing or later GPUs, or Ada or later when using GPU virtualization).

If installing from the .run file, installation will detect what GPUs are present and default to installing the open kernel modules if all NVIDIA GPUs in the system can be driven by the open kernel modules. Distribution-specific repackaging of the NVIDIA driver may require additional steps, specific to that packaging, to choose the open flavor.

In the release 560 series, it will still be possible to configure the .run file to install the proprietary flavor of kernel modules, with the --kernel-module-type=proprietary command line option. However, in the future, some GPUs may only be supported with the open flavor."

Great to see! Presumably it's with next-gen/future GPUs where they may end up only enabling those future GPU architectures along the "open flavor" kernel driver and leave the proprietary driver for supporting existing/legacy products.

As a reminder, this open-source GPU kernel driver support is currently just around their kernel driver. This kernel driver code remains out-of-tree / distributed as part of a GitHub repository or their packaged Linux driver releases. NVIDIA did recently hire the former Nouveau kernel driver maintainer and another NVIDIA engineer has begun contributing to the open-source NVK driver but so far there's no indications of major shifts around going open-source on the user-space driver side, especially around their walled CUDA compute garden.