For years, there has been talk of an accelerator subsystem for the Linux kernel, given that as of now, most AI training / inference accelerator drivers end up getting lodged in. the “char / misc” area of ​​the kernel. This discussion of the accelerator subsystem has been reignited with discussions of whether to have such a subsystem or move these drivers into the GPU / DRM subsystem space.
Arising from the recent controversy around the Habana Labs AI driver code changes (although now at least partially resolved with the release of an open source compiler and user space library) and how some changes were circumvented in going through “char / misc” and facing less scrutiny than if the same changes were attempted to go through the GPU / DRM tree, the discussion once again takes place over where these throttle drivers should reside in. the core.
Direct Rendering Manager pilots even want to move throttle pilots within drivers / gpu given how similar they tend to be with non-display GPU drivers sharing many interfaces and having common goals. Living in the hypothetical drivers / acceleration was also brought up again with DRM maintainers still interested in overseeing this new area of ​​the kernel.
Greg KH, who oversees char / misc, welcomes all new reviewers and helps manage the code, but questioned the pilot / gpu housing proposal and more firmly determined some of the new throttle pilot requirements for the ‘upstream.
Nothing has been decided for sure yet on an accelerator subsystem or whether GPU / DRM maintainers will oversee these accelerator drivers in the future. Those interested in the latest discussion can view this kernel thread.
