Habana Labs Preps More Linux Code For Their AI Accelerators With The 5.7 Kernel

Habana Labs, the AI accelerator start-up being acquired by Intel, has more driver improvements on tap for Linux 5.7.

Habana Labs has been a good open-source supporter with punctually working on their mainline Linux driver enablement for their products. Their upstream Linux driver work started off at the start of 2019 with their for their Goya inference accelerator and increasing work on their Gaudi AI training product. They have been aiming to land their Gaudi enablement in Linux 5.7~5.8 but now it's looking like that will be the latter kernel if not longer.

For Linux 5.7 is more work on prepping their driver for Gaudi and future products but for now are not supported. The Habana Labs Linux driver code for the 5.7 kernel is seeing MMU improvements, better handling for future ASIC support, a new DebugFS interface, improved HWMON sysfs support with sensor reading, printing to the kernel log when clock throttling happens due to power/thermal issues, and various other fixes and improvements.

More details on the Habana Labs driver work for Linux 5.7 via this now honored pull request. Given Intel's long open-source track record, we have no doubt about Habana Labs' open-source work continuing once Intel's acquisition of the Israeli start-up is complete. In fact, we are looking forward to it being integrated within oneAPI and the other aspects of Intel's software ecosystem.

Last year there was talk of introducing a hardware accelerator subsystem for these AI accelerators rather than just living within char-misc, though at least for the moment that seems to have stalled. When seeing more AI accelerator drivers for the mainline kernel moving forward, perhaps then they'll receive their own official area of the kernel.