Apple has secured a position on the board of the Ultra Accelerator Link Consortium, which is comprised of over 65 members dedicated to developing cutting-edge AI accelerator architecture. This collaboration will enable Apple to play a crucial role in shaping the new standard and advocating for its widespread implementation. Essentially, Apple is set to contribute to the establishment and endorsement of a standard for interconnecting numerous GPUs in data centers facilitating AI operations.
“UALink shows significant potential in overcoming connectivity hurdles and fostering new avenues for enhancing AI functionalities and requirements,” remarked Becky Loop, the Director of Platform Architecture at Apple. “With a longstanding tradition of leading and collaborating on groundbreaking innovations that advance our industry, we are eager to become part of the UALink Board of Directors.”
So, what is UALink? In simple terms, UALink represents an industry initiative to challenge Nvidia’s supremacy in AI hardware. The objective is to create a standard that competes with Nvidia’s proprietary NVLink technology. An article from Tom’s Hardware last fall elaborates:
UALink aims to establish itself as the open standard for scaling up connections among multiple AI accelerators, effectively positioning itself against Nvidia’s proprietary NVLink. NVLink facilitates communication between GPUs in servers or clusters of servers through Infiniband, another technology predominantly owned by Nvidia, enabling advanced scaling. Infiniband faces competition from the newcomer Ultra Ethernet, a major consortium of tech giants working on an open standard to counter Nvidia’s market control.
With support from AMD and Intel, the UALink Consortium includes key members such as AWS, Google, Meta, and Microsoft. Apple has now joined the ranks.
“The initial 1.0 version 200Gbps UALink specification permits the connection of up to 1,000 accelerators within an AI pod and is based on the IEEE P802.3dj PHY Layer,” as stated by the UALink Consortium. “The specification will be accessible to Contributor Members in 2024 and will be publicly released in the first quarter of 2025.”
ʟᴀᴛᴇꜱᴛ ᴀᴘᴘʟᴇ ᴀᴄᴄᴇꜱꜱᴏʀʏ ʀᴇᴄᴏᴍᴍᴇɴᴅᴀᴛɪᴏɴꜱ
: . More.