NVIDIA is advancing its 2022 initiative by progressively substituting proprietary GPU kernels with open-source alternatives. This development marks a pivotal shift, though it falls short of a complete open-source GPU driver, focusing solely on the kernel. Despite this, the move signifies a significant departure for NVIDIA.
Closed-source software has drawn criticism, particularly due to concerns that it could lead to discontinued hardware support when companies decide to phase out older architectures. While NVIDIA does not have a comprehensive open-source driver stack in its immediate plans, the decision to open its GPU kernel for GeForce and workstation GPUs represents a notable step forward.
Two years after initiating the open-sourcing of the Linux GPU kernel, NVIDIA reports that their open-source kernel matches or exceeds the performance of the closed-source counterpart. Moreover, new functionalities such as Heterogeneous Memory Management (HMM), confidential computing, and coherent memory architectures for the Grace architecture have been introduced.
According to NVIDIA, the upcoming R560 driver release will complete the transition to fully embrace open-source GPU kernel modules. However, it’s important to note that the open-source GPU kernel is not universally compatible with all architectures. It is mandatory and supported on Grace Hopper and Blackwell, while newer architectures like Turing, Ampere, Ada, and Hopper offer user choice. Unfortunately, open-source kernels are not supported on Maxwell, Pascal, and Volta architectures.
Currently, NVIDIA has introduced drivers from the R560 branch, but these are available exclusively for Windows, not Linux.
Related topics:
Google Data and Machine Learning: Transforming Insights into Innovations