Intel's Habana Labs Mainlining Gaudi Accelerator Support In Linux 5.8

AI startup Habana Labs, which was recently acquired by Intel, will see its Gaudi accelerator supported by the upcoming Linux 5.8 kernel.

Habana Labs even in its pre-Intel days has been a good open-source patron with punctually open-sourcing their AI accelerator kernel driver code. They began with their Goya AI last year and have continued improving the upstream kernel driver while also preparing for Gaudi support in more recent kernels/months. With Linux 5.8 it looks like their Gaudi ASIC enablement is complete.

Habana Labs Gaudi

Goya is the company's AI inference accelerator while Gaudi is intended for AI training and boasts integrated RDMA over converged Ethernet and other features to deliver leading AI training performance.

Oded Gabbay who has been overseeing the Habana Labs kernel code sent in the pull request to char/misc for pushing the Gaudi support in as part of the upcoming Linux 5.8 kernel. The kernel driver is enough for initializing the hardware and being able to run workloads on it. This Habana Labs driver update for Linux 5.8 also includes exposing more data through HWMON, support for loading the ASIC's bootloader code from host memory, better ASIC reset handling, and other improvements.