The Intel Habana Labs driver continues seeing a lot of work around the Gaudi2 hardware especially. Though somewhat surprising is not seeing much driver activity around Greco that was also announced last year at the same time as Gaudi2.
An initial pull request was sent out on Monday with the initial Habana Labs driver updates ready for DRM-Next that will target the Linux 6.4 merge window. The Habana Labs accelerator driver for Linux 6.4 is adding a graceful reset mechanism for compute reset, continued additions around Gaudi2 hardware support, letting the user fetch the device memory current usage even when undergoing a compute reset, providing the user with the specific reason why a device is still in use, and adapting to the latest device firmware.
The Habana Labs accelerator driver with Linux 6.4 includes a number of new user-space API additions. Among those changes are being able to stall/resume specific engines within Gaudi2 as part of power testing/measurements during training with different topologies. There is also a new interface to query the amount of device memory the driver and firmware have reserved for themselves. There is also new user-space API additions for exposing Gaudi2's bit-mask for the available rotator engines, exposing the register's address of the firmware that should be used to trigger interrupts from within the user's code running in the compute engines, and new opcodes to fetch information on hardware and firmware events.
Overall the "habanalabs" accel driver continues moving along and paired with their open-source user-space stack and compiler remains the best AI accelerator solution for having an upstream kernel driver with fully open-source software stack besides the usual firmware caveat.