Big Tech computing companies formed a consortium to define a new open standard for interconnecting AI accelerators. NVIDIA was not invited to be part of the group even though it is the largest supplier of AI GPUs by far.
AI data centers need to move massive amounts of data with very low latency. High-bandwidth data processing on GPUs happens extremely fast, but the challenge is to transfer data within and between clusters of these AI accelerators within data centers.
NVIDIA created NVLink, its proprietary high-speed interconnect specifically designed for communication between its GPUs. The problem is that NVLink is proprietary, so it only works with NVIDIA GPUs.
AMD, Broadcom, Cisco, Google, Hewlett Packard Enterprise (HPE), Intel, Meta, and Microsoft announced that they have formed the Ultra Accelerator Link Promoter Group. The group aims to define and promote an open standard called Ultra Accelerator Link, or UALink.
The idea is to have UALink adopted by the industry as the standard solution for high-bandwidth, low-latency data transfer between AI accelerators in data centers.
Similar efforts to standardize protocols have been essential for the tech industry in the past. Because we have open standards like the PCI Bus, Ethernet, or TCP/IP, hardware and software from different manufacturers can be connected to each other.
This may be part of the reason why NVIDIA wasn’t invited to the party. If the consortium of tech companies can agree on an open industry networking standard that isn’t influenced by NVIDIA’s tech then it could work to break the near monopoly NVIDIA seems to have.
AMD and Intel are direct competitors of NVIDIA in the GPU market, and Microsoft and Google are both developing their own AI hardware.
“An industry specification becomes critical to standardize the interface for AI and Machine Learning, HPC (high-performance computing), and Cloud applications for the next generation of AI data centers and implementations,” the consortium said in a statement.
Version 1.0 of UALink is expected to be ready by Q3 2024 and will be made available to companies that join the Ultra Accelerator Link (UALink) Consortium.
The absence of NVIDIA doesn’t necessarily mean they’re permanently excluded. The consortium could decide to welcome them in the future, and NVIDIA could choose to adopt UALink if there’s widespread industry acceptance.
