Apple Joins the UALink Consortium for AI Chip Connectivity

Device-maker Apple joined the board of a recently incorporated industry group that aims to establish open standards for directly connecting AI accelerator chip clusters in data centers.

See Also: Live Webinar | AI in the Spotlight: Exploring the Future of AppSec Evolution

The Ultra Accelerator Link Consortium developed the chip-agnostic UALink specification to wire 1,024 accelerators – GPU chips modified for specialization in AI applications – so they can directly communicate. The direct connection optimizes the parallel computing and high data throughput that make GPUs efficient at training large language models. It sidesteps CPUs as intermediaries – or as a consortium spokesperson phrased it in an email: cuts down “the number of widgets” and so reduces latency and bandwidth limits.

The consortium is widely seen as an industry attempt to take on the proprietary NVLink standard used by AI-chip giant Nvidia. It has pledged to make the specification available to the public within the first three months of this year. The specification can work with Nvidia chips, the spokesperson said, but is not compatible with NVLink.

“UALink shows great promise in addressing connectivity challenges and creating new opportunities for expanding AI capabilities and demands,” said Becky Loop, director of platform architecture at Apple.

As AI models get bigger, they need more memory and processing power, which means spreading the work across many GPUs connected in a group, or “pod.” Under the UALink specification, the GPUs are connected by UALink switches, which ensure that when one GPU needs to access memory on another GPU, it can do so quickly, with just one hop through the switch. The switch connection is faster than PCIe Gen5 speeds of 128 gigabytes on a 16-lane link, the spokesperson said.

The challenge with going through the CPU to communicate between GPUs is that each step in the process adds delay. One of these steps often becomes a bottleneck, limiting the overall speed of data transfer.

Apple shares its new board position with tech giant Alibaba and semiconductor design firm Synopsys. Intel, AMD, Google, AWS, Microsoft and Meta are also part of the over 65-member consortium.

Apple’s involvement in UALink appears to be part of its increasing investments in AI infrastructure under the Apple Intelligence banner. The Information reported that Apple is working on a new server chip that is expected to boost its data center operations for AI services.