Habana Labs Gaudi2 Support Leads The Linux 6.0 Char/Misc Changes
Intel announced Gaudi2 at their Vision conference in May. Gaudi2 is their next-gen AI processor for training and inference. Gaudi2 aims to deliver two times better AI training performance than the popular NVIDIA A100. Compared to the original Gaudi that was manufactured at 16nm, Gaudi2 is 7nm based, triples the TPC count to 24, offers 96GB of HBM2E memory compared to 32GB HBM2, 24 x 100 GbE networking, and now has a 600 Watt TDP over 350 Watts with the original model.
Back in June Intel began publishing the Linux driver support for Gaudi2 sans the networking bits. The Gaudi2 support builds off the existing "habanalabs" mainline Linux kernel driver. The past few weeks the Gaudi2 code has been queued up in char-misc for 6.0 (nee 5.20) and now merged to mainline. Intel at Vision also announced the Greco AI processor but that isn't yet supported by this driver.
Habana Labs' self-published benchmark results for the Gaudi2 can be found at habana.ai.
The Habana Labs Gaudi2 support leads the char/misc changes especially with all the new header files leading to tons of new kernel code -- the char/misc pull is 184k lines of new code and 11.8k lines removed. But besides the Gaudi2 enablement, char/misc with Linux 6.0 also has many Industrial I/O driver updates as well as updates to SoundWire, Slimbus, FPGA drivers, and other miscellaneous changes. The full list of patches can be found via this pull.
Char/Misc could become a bit lighter moving forward with renewed discussions over a Linux kernel accelerator subsystem or the alternative proposal that AI accelerator/compute drivers should be part of the DRM subsystem.