Intel Ready With Habana Labs Gaudi2 Support For Linux 5.20
Gaudi2 is the second-generation Habana Labs offering for training and inference. Gaudi2 is promoted as delivering twice the AI training performance of NVIDIA A100 hardware and features 96GB of HBM2e memory, 3x for the TPCs, is now manufactured on a 7nm process, and sports 24 x 100 GbE networking. The Gaudi2 carries a 600 Watt TDP. It's an impressive piece of AI hardware and backed by open-source Linux driver support.
Gaudi2 is built off the same overall architecture of Gaudi and thus the Linux "habanalabs" driver enablement isn't too invasive -- most of the 158k lines of new code amounts to header files. Intel also published their TPC_LLVM compiler for Gaudi2 alongside the kernel patches and are working to publish their new SynapseAI Core code too in user-space.
Gaudi2 is ready for mainline with this pull request sent overnight by Oded Gabbay including the Gaudi2 support for Linux 5.20. Oded noted that all Gaudi2 features are wired up with this driver code with the exception of the networking bits.
Due to the Gaudi2 enablement, the Habana Labs driver pull request for this next kernel merge window comes in at 159,385 insertions and 1164 deletions. Greg KH has already pulled these driver changes into char-misc ahead of the 5.20 merge window opening up later this month. Intel also announced the Habana Labs Greco back in May but the Linux kernel support for that isn't yet ready.