Linux 6.2 "char/misc" Changes Land With Continued Intel Gaudi2 Enablement
Gaudi2 as the next-generation AI training and inference accelerator from Habana Labs designed to compete with -- and sharply outperform -- NVIDIA's A100 for computer vision, natural language processing, and related workloads. Gaudi2 was announced earlier this year and shortly thereafter their stellar open-source Linux team began posting new patches for extending the existing "habanalabs" Gaudi and Goya Linux kernel driver to support the Gaudi2.
The past several Linux kernel cycles have seen more Gaudi2 code upstreamed and it continues that way for Linux 6.2. With Linux 6.2 the Habana Labs driver has added a user-space API for obtaining page fault information, Gaudi2 PCI revision 2 support, support for graceful hard resets of the hardware, and a variety of other fixes and enablement items around Gaudi2.
This additional kernel driver work follows the SynapseAI Core user-space software being updated at the end of November and also now supporting Gaudi2. It's great seeing Intel / Habana Labs continue as the most open-source, Linux-upstream friendly minded AI accelerator vendor.
Intel Habana Labs Gaudi2
This Habana Labs driver work leads the char/misc changes for Linux 6.2. There isn't much else notable in this pull besides some new Ampere Computing SMpro co-processor drivers.
While the Habana Labs driver is currently part of the char/misc area, Linux 6.2 is introducing the compute "accel" accelerator subsystem as an extension of the Direct Rendering Manager (DRM) subsystem. The Habana Labs driver is among the AI/VPU accelerator drivers currently being adapted to this new subsystem/framework so it will eventually move from char/misc to this dedicated accel space.
Add A Comment
