The Ultra Accelerator Link (UALink) Consortium, founded last summer, brings together some of the biggest names in semiconductors and data centres, including AMD, Intel, and Microsoft, to develop standards to enhance connecting multiple GPUs.
The group has now officially ratified the UALink 200G 1.0 Specification, which allows up to 1,024 AI accelerators to communicate with each other at high speeds within computing clusters.
The Consortium hopes the standard will enhance the performance of interconnected chips running AI workloads while reducing the reliance on NVLink, Nvidia’s proprietary interconnect system.
“As the demand for AI compute grows, we are delighted to deliver an essential, open industry standard technology that enables next-generation AI/ML applications to the market,” said Kurtis Bowman, board chair of the UALink Consortium and director for architecture and strategy at AMD.
“The groundbreaking performance made possible with the UALink 200G 1.0 Specification will revolutionise how Cloud Service Providers, System OEMs, and IP/Silicon Providers approach AI workloads.”
The UALink standard creates what is effectively a switch ecosystem for accelerators, providing inter-hardware communication across system nodes to create multi-node systems for AI applications.
The standard is designed to drastically reduce latency within computing clusters while providing a simple load/store protocol that has the same raw speed as Ethernet but with the latency of PCIe switches.
“With the release of the UALink 200G 1.0 Specification, the UALink Consortium’s member companies are actively building an open ecosystem for scale-up accelerator connectivity,” said Peter Onufryk, an Intel Fellow and president of the UALink Consortium.
“We are excited to witness the variety of solutions that will soon be entering the market and enabling future AI applications.”
The group welcomed Alibaba Cloud, Apple, and Synopsys as its newest members in January, adding to existing participants, including AWS, Cisco, Google, and Meta.
RELATED STORIES
Apple, Alibaba Cloud join board of group developing GPU interconnect standards
