Intel Begins Publishing Habana Labs Gaudi2 Linux Driver Code

Last month for their Vision conference, Intel announced the Habana Labs Gaudi2 accelerator. Gaudi2 is the second-generation offering from Intel-owned Habana Labs for training and inference. The open-source Linux kernel driver support for Gaudi2 is now getting underway along with accompanying open-source user-space software stack support.

Intel introduced Gaudi2 as delivering two times better AI training performance than the NVIDIA A100 solutions. Gaudi2 is manufactured on a 7nm process rather than 16nm with the original Gaudi, a 3x improvement for the TPCs, 96GB of HBM2e memory up from 32GB of HBM2, 48MB SRAM cache that is double from the original, 24 x 100 GbE, and has a 600 Watt TDP (350 Watt for the original).

Just over one month since announcing the Habana Labs Gaudi2, the open-source driver work is working its way out publicly. The Gaudi2 support builds off the existing "habanalabs" in-kernel driver that supports the existing Gaudi and Goya accelerators.

Learn more about the Gaudi2 architecture via the Habana.ai whitepaper.

Gaudi2 from the driver perspective is quite close to the original Gaudi and re-affirms the same overall architecture. Enabling Gaudi2 is around 158k lines of new kernel code, but much of that is header files -- just as we are used to the massive header files with the AMDGPU kernel driver. New header files by far represent the bulk of the code churn for the Gaudi2 enablement for this mainline driver.

In addition to the kernel driver work, also published today was TPC_LLVM 1.1 as the newest release of their open-source LLVM-based compiler for targeting Habana Labs accelerators. TPC_LLVM 1.1 adds the compiler support for Gaudi2 as well as the new Greco ASIC. They are also working on updating their open-source SynapseAI Core for Gaudi2 as well.

See this patch series for the new Gaudi2 kernel driver code. Given its evolving off the same Gaudi architecture, hopefully the code review will go smoothly and we could see this Intel Gaudi2 support added as soon as the v5.20 cycle if all goes well.

Intel Vision 2022 also saw the announcement of the Habana Labs Greco as their deep learning inference efficiency-optimized card. Greco is a big upgrade over Goya. Aside from the TPC_LLVM compiler support, we haven't yet seen the Habana Labs kernel driver support for Greco but presumably that will be out soon too.