AMD, Broadcom, Cisco, Google, HPE, Intel, Meta, and Microsoft are uniting in Ultra Accelerator Link (UALink). The collaboration aims to counterbalance Nvidia’s NVLink Interconnect technology.
UALink describes the goal as follows: “Develop a new industry standard that serves to advance high-speed, low-latency communications for scalable AI systems interconnected in data centres.” This can be understood as a desire to develop an alternative to Nvidia’s NVLink Interconnect technology.
The Nvidia technology is used in data centres to join the forces of different servers. Specifically, it involves linking different AI accelerator chips. That way, AI applications such as ChatGPT can spit out answers at high speed, but bundling is even more important for the intensive task of LLM training.
New industry standard
The group that unites Nvidia competitors, among others, will devise an industry standard to bundle server capabilities in a new way. UALink wants the new standard to put OEMs, IT professionals and system integrators in control.
The first version of the industry standard is expected in the third quarter of 2024, which is in the fall. This specification should manage to connect up to 1,024 accelerators. The standard will be made available to companies that join the UALink Consortium.
Nvidia dominates
Nvidia has made good money in recent months from its monopoly on AI hardware for data centres. The company’s decision not to join UALink will also be driven by its strong market position. However, this would have been better news for the development of the standard. Now, it remains to be seen whether the industry standard catches on and lifts Nvidia’s domination.
Also read: Nvidia’s market cap is hot on Apple’s heels