For years, utilizing hardware acceleration for machine learning at the edge meant being tethered to a vendor’s Board Support Package (BSP): a world of stagnant kernels, proprietary binary blobs, and zero auditability. The companies that wanted to run a modern ML workload had to accept their outdated software or spend months fighting it, in both situations risking vendor lock-in.

That era is ending. Thanks to recent work in the Linux accel subsystem and Mesa, a truly open-source AI stack is now a reality. This talk goes over what is currently supported by the mainline stack and the existing four hardware drivers: Etnaviv (Vivante), Rocket (Rockchip), Ethos-U (Arm), and Thames (TI C7x). I will also explain what is missing and what is coming next.