Tech Giants Unite to Pioneer Next-Gen AI Chip Components
A consortium of technology behemoths, including Intel, Google, Microsoft, and Meta, is spearheading a new venture to innovate the future of artificial intelligence (AI) computing. This collaborative effort has led to the formation of the Ultra Accelerator Link (UALink) Promoter Group, an entity committed to steering the development of components critical for linking AI accelerator chips within data centers.
The announcement, made this past Thursday, revealed that the UALink Promoter Group is not only supported by its founding members but also includes other industry powerhouses like AMD, Hewlett Packard Enterprise, Broadcom, and Cisco. Notably absent, however, is the chip manufacturer Arm. This coalition aims to establish a new industry standard for interconnecting the increasingly prevalent AI accelerator chips in server setups.
“The industry needs an open standard that can evolve swiftly, in a manner that benefits the ecosystem as a whole through the contributions of multiple companies,” explained Forrest Norrod, AMD’s General Manager of data center solutions, during a press briefing. “We’re advocating for a standard that fosters innovation at an accelerated rate, without being hindered by any single entity’s control.”
The proposed initial version, UALink 1.0, is ambitious, aiming to link up to 1,024 AI accelerators—specifically, GPUs—within a singular computing ‘pod’, with a pod defined as one or more server racks. Leveraging “open standards”, including AMD’s Infinity Fabric, UALink 1.0 aspires to enable direct memory access between AI accelerators, significantly enhancing speed and reducing data transfer latency over existing interconnect technologies.
Plans are afoot to orchestrate the UALink Consortium in the third quarter of this year to govern the ongoing evolution of the UALink specifications. The consortium’s formation will coincide with the availability of UALink 1.0 to member companies, with an expectation to introduce an upgraded, higher-bandwidth version, UALink 1.1, in the fourth quarter of 2024. Products based on UALink are anticipated to hit the market in the following years.
Conspicuously missing from the UALink Promoter Group is Nvidia, the leading AI accelerator manufacturer, controlling an estimated 80% to 95% market share. Nvidia’s current dominance is built on proprietary interconnect technology for GPUs within data centers, possibly explaining its absence from an initiative championing a rival standard. Nvidia opted not to comment on the situation.
While Nvidia rides high on its commercial success, other tech giants are not sitting idly. Companies such as Microsoft, Meta, and Google, traditionally reliant on Nvidia’s hardware for powering clouds and training AI models, are now among those most incentivized by the UALink initiative. The move is seen as a strategic pivot to reduce dependence on a single vendor, with many in the industry developing or refining their custom chips for AI applications. For instance, Google with its TPUs and Axion, Microsoft with Maia and Cobalt, and Meta with its series of accelerators.
A glimpse into Microsoft’s collaboration with OpenAI, planning a substantial investment in a supercomputer equipped with future iterations of Cobalt and Maia chips, hints that UALink could be the glue binding these innovative chips together in monumental computing tasks.
This bold step towards collaborative innovation marks a significant shift in the AI hardware landscape, with potential ramifications for vendor dominance and technological advancements. As the UALink Consortium gears up for its formative phase, the tech world waits in eager anticipation for what this collective ingenuity will unveil for the future of AI computing.
