Intel Habana Labs AI Driver Moving To New «Accel» Subsystem For Linux 6.3


The Intel «habanalabs» AI driver is moving to the new accelerator «accel» subsystem with the upcoming Linux 6.3 kernel cycle.

The Habana Labs driver has been living in the catch-all «char/misc» area of the accelerator subsystem since it was mainlined several years ago. The habanalabs driver has lived there for the lack of having a dedicated subsystem for AI devices / accelerators. But finally during the back half of 2022, the «accel» subsystem finally came together for the AI accelerator drivers working their way to the mainline kernel and in agreement of the different upstream kernel developer parties. This accel subsystem builds off the Direct Rendering Manager (DRM) subsystem given all the commonality between today’s AI accelerators and GPUs, especially when it comes to a number of interfaces around DMA-BUF and others that can be shared/re-used with these drivers.

The compute accelerator subsystem was merged in Linux 6.2 but that just established the area and the initial framework code to be used by the accelerator drivers. Now for Linux 6.3 we’ll see the first drivers living within this newly-established area of the kernel.

As I wrote about a few days ago, the Intel Versatile Processing Unit (iVPU) driver was submitted as the first to go into this new area. Now on Thursday the pull request was sent out for adapting the Intel Habana Labs driver to move from char/misc to accel.

This pull request from maintainer Oded Gabbay transitions the driver to the new accelerator subsystem. But still there is more integration work that will happen over future kernel cycles.

Intel Habana Labs Gaudi2 accelerator card.
Intel Habana Labs Gaudi2

Besides moving the Habana Labs driver to the accel subsystem for Linux 6.3, there is more Gaudi2 enablement work and various other minor enhancements.