On Tuesday, the Ultra Accelerator Link Consortium (UALink) announced that Alibaba Cloud, Apple, and Synopsys have joined its board of directors. This inclusion allows these tech giants to play a significant role in shaping UALink, a technology aimed at connecting AI and high-performance computing (HPC) accelerators, positioning itself as a rival to Nvidia’s NVLink. The move suggests Apple’s growing interest in AI datacenter connectivity, hinting at possible plans for a dedicated AI datacenter acceleration product.
Set to release in early 2025, UALink Version 1.0 will facilitate the linkage of as many as 1,024 accelerators in an AI computing pod. This architecture will operate within a high-speed, low-latency network reaching speeds up to 200 Gb/s, essential for AI training tasks as it allows for seamless data transfers between processor-attached memory. Several prominent companies including AMD, Broadcom, Cisco, Google, HPE, Intel, Meta, and Microsoft support this open industry standard, all of which are influential in developing AI hardware and software. In total, 65 companies are supporting UALink.
Apple’s Director of Platform Architecture, Becky Loop, stated, “UALink holds immense potential to overcome connectivity hurdles and broaden AI capabilities and application scopes. Apple has consistently been at the forefront of innovation and collaboration that propels our industry forward. We’re thrilled to be part of the UALink board.”
While Apple has not officially confirmed any projects to create datacenter-class AI processors, there has been ongoing speculation about such developments. Apple could potentially emulate its tech counterparts like Google and Facebook by collaborating with Broadcom to produce custom datacenter accelerators for both training and inference tasks. This strategy would allow Apple to customize its AI capabilities to meet specific requirements, potentially reducing dependency on costly external accelerators from companies like AMD and Nvidia, as well as cutting down on energy and software expenses. Although still speculative, Apple’s involvement in the UALink Consortium strongly suggests interest in developing proprietary datacenter-grade processors.
For Alibaba Cloud, joining the UALink Consortium aligns with its ambitions to craft its own AI hardware, a common goal among major cloud service providers. Qiang Liu, VP of Alibaba Cloud and GM of Alibaba Cloud Server Infrastructure, remarked, “At Alibaba Cloud, we see great value in advancing AI accelerator interconnection technology by identifying and addressing core needs from a cloud computing and application standpoint. UALink, as a frontrunner in AI accelerator interconnects, has united pivotal AI infrastructure players to establish protocols inherently suited for AI accelerators, fostering innovation and enhancing AI workload performance. This collaboration is key to developing an open, innovative industry ecosystem.”
While Synopsys does not produce AI accelerators, the firm licenses physical IP to its clients, making its involvement in UALink crucial. Richard Solomon, a UALink Board Member and Sr. Staff Product Manager at Synopsys, explained, “UALink will play a vital role in meeting the performance and bandwidth demands of hyperscale data centers, enabling the necessary high-speed interconnects to expand pods and clusters. As a leader in top-tier interface IP solutions, Synopsys is dedicated to leveraging our expertise within the UALink Consortium to establish standards for the world’s most advanced AI accelerator architectures.”