Kernel developers discuss a subsystem for AI accelerators
Summary
In 2019 kernel developers discussed a dedicated subsystem for AI accelerators, whose drivers had until then been scattered across various areas of the kernel. Olof Johansson proposed patches and wanted to maintain the subsystem together with Greg Kroah-Hartman. Developers of the graphics drivers considered the idea wrong because the differences from GPUs were minimal.
Ideas
- A shared subsystem is meant to make drivers easy to find and encourage collaboration.
- Reusable frameworks for several drivers were envisaged, but still undefined.
- Graphics developers saw AI accelerators as a variant of GPUs.
Insights
- How new hardware is classified in the kernel decides review culture and requirements.
- Separate subsystems for similar hardware easily lead to duplicate infrastructure.
Facts
- Drivers for AI accelerators had previously been spread across various areas of the Linux kernel.
- Olof Johansson proposed a shared subsystem for these accelerator drivers.
- Johansson and Greg Kroah-Hartman wanted to take over maintenance of the new subsystem.
- Dave Airlie warned of hard-to-review interfaces to proprietary user space drivers.
- Daniel Vetter saw no sufficient technical differences between AI accelerators and GPUs.
References
Critique
- A dedicated subsystem makes things easier to find, but can duplicate functions of the DRM graphics stack.
Remarks
- With Linux 6.2 in 2023, the “accel” subsystem was created as part of the DRM infrastructure, as the graphics developers had demanded.
Recommendations
- With accelerator hardware, demand open user space drivers, not just kernel modules.
- Before buying, check whether the driver is maintained in the mainline kernel under drivers/accel.
Links to the original source and the Web Archive open in a new tab.