Tuesday, September 24, 2024

What is Intel Gaudi 3?

Intel's Enterprise Tech Tour in Oregon brought more than just glimpses of the new Xeon 6900 "Granite Rapids" server processors. The spotlight also shone on the Gaudi 3, their AI inference accelerator, which has generated a great deal of anticipation. However, one question has been lingering: When will the much-needed Linux driver support arrive?

The Gaudi 3 was initially unveiled back in April. Since then, many have been eagerly awaiting the release of the Linux kernel driver patches for the "habanalabs" accel driver. This wait is particularly perplexing considering the swiftness with which Intel provided driver support for its predecessor, the Gaudi 2. Announced in May 2022, the Gaudi 2 saw its Linux driver patches surface just a month later. Six months on, however, there's still no sign of the Gaudi 3 kernel driver patches. This delay becomes even more concerning given the extended release cycles of the Linux kernel.

At the Enterprise Tech Tour, Intel offered a glimmer of hope, stating that the Gaudi 3 Linux driver support should be released with their October software update. While this is good news, the timing still raises questions. An October release would mean that the Gaudi 3 driver patches would be upstreamed to Linux v6.13 at the earliest, which is not ideal. With the Linux 6.13 merge window opening in November following the Linux 6.12 release, the stable v6.13 kernel isn't expected before February. This means that even if the Gaudi 3 support is published next month, reviewed promptly, and goes smoothly, it might still be early next year before it becomes available as a stable kernel and finds its way into distributions like Ubuntu 25.04. Furthermore, the October deadline leaves just a few weeks for code review and any revisions, potentially causing further delays.

The ideal scenario would have seen the code released weeks ago, potentially making its way into the habanalabs driver for Linux 6.12, this year's Long Term Support (LTS) kernel version. Adding to the complexity of the Gaudi 3 upstreaming plans, earlier this year, the long-time driver maintainer, Oded Gabbay, stepped away from Intel. Just two months later, the new maintainer also left the company. This turnover in personnel could have impacted the driver development timeline.

Intel has announced "select availability" of the Gaudi 3 accelerator in Q4, with general availability scheduled for 2025. The specific date in 2025 has not been disclosed, so if the Gaudi 3 support does manage to make it into Linux 6.13, the timing might be suitable for those using the latest kernel. It's also likely that Intel will offer a DKMS module or similar for those running older enterprise kernels.

Despite the delays, Intel's commitment to open-source driver support for their accelerators has been a positive aspect. The habanalabs driver remains a flagship example of Intel's contributions to establishing the "accel" subsystem within the Linux kernel. The hope is that Gaudi 3 will follow suit and provide a robust and reliable open-source driver experience.

While the delayed arrival of the Gaudi 3 Linux driver may cause frustration, Intel's commitment to open-source support remains a crucial factor for the success of this AI inference accelerator. The coming months will be crucial for ensuring a seamless integration with the Linux ecosystem, and Intel's timeline will undoubtedly be closely watched by developers and users alike.

0 comments:

Post a Comment