On Thursday, a number of main tech corporations, together with Google, Intel, Microsoft, Meta, AMD, Hewlett Packard Enterprise, Cisco, and Broadcom, introduced the formation of the Extremely Accelerator Hyperlink (UALink) Promoter Group to develop a brand new interconnect commonplace for AI accelerator chips in knowledge facilities. The group goals to create an alternative choice to Nvidia’s proprietary NVLink interconnect expertise, which hyperlinks collectively a number of servers that energy at this time’s AI functions like ChatGPT.
The beating coronary heart of AI nowadays lies in GPUs, which might carry out large numbers of matrix multiplications—obligatory for working neural community structure—in parallel. However one GPU usually is not sufficient for complicated AI techniques. NVLink can join a number of AI accelerator chips inside a server or throughout a number of servers. These interconnects allow sooner knowledge switch and communication between the accelerators, permitting them to work collectively extra effectively on complicated duties like coaching massive AI fashions.
This linkage is a key a part of any fashionable AI knowledge heart system, and whoever controls the hyperlink commonplace can successfully dictate which {hardware} the tech corporations will use. Alongside these traces, the UALink group seeks to ascertain an open commonplace that enables a number of corporations to contribute and develop AI {hardware} developments as a substitute of being locked into Nvidia’s proprietary ecosystem. This method is just like different open requirements, comparable to Compute Express Link (CXL)—created by Intel in 2019—which offers high-speed, high-capacity connections between CPUs and gadgets or reminiscence in knowledge facilities.
It isn’t the primary time tech corporations have aligned to counter an AI market chief. In December, IBM and Meta, together with over 50 different organizations, formed an “AI Alliance” to advertise open AI fashions and provide an alternative choice to closed AI techniques like these from OpenAI and Google.
Given the market dominance of Nvidia—the present market chief in AI chips—it’s maybe not stunning that the corporate has not joined the brand new UALink Promoter Group. Nvidia’s latest massive financial success places it in a robust place to proceed forging its personal path. However as main tech corporations proceed to put money into their very own AI chip growth, the necessity for a standardized interconnect expertise turns into extra urgent, notably as a method to counter (or at the least steadiness) Nvidia’s affect.
Rushing up complicated AI
UALink 1.0, the primary model of the proposed commonplace, is designed to attach as much as 1,024 GPUs inside a single computing “pod,” outlined as one or a number of server racks. The usual relies on applied sciences like AMD’s Infinity Architecture and is anticipated to enhance velocity and scale back knowledge switch latency in comparison with present interconnect specs.
The group intends to type the UALink Consortium later in 2024 to handle the continuing growth of the UALink spec. Member corporations can have entry to UALink 1.0 upon becoming a member of, with a higher-bandwidth model, UALink 1.1, deliberate for launch in This fall 2024.
The primary UALink merchandise are anticipated to be out there throughout the subsequent two years, which can afford Nvidia loads of lead time to broaden proprietary lock-in because the AI knowledge heart market grows.
