On Thursday, leading technology companies including Google, Intel, Microsoft, Meta, AMD, Hewlett Packard Enterprise, Cisco and Broadcom announced the formation of the Ultra Accelerator Link (UALink) Promoter Group to develop a new interconnect standard for AI accelerator chips in data centers. The group aims to create an alternative to Nvidia's proprietary NVLink interconnect technology, which interconnects multiple servers that power today's AI applications such as ChatGPT.
At the heart of modern AI are GPUs, which can perform the vast number of matrix multiplications in parallel required to run neural network architectures. But a single GPU is often insufficient for complex AI systems. NVLink can connect multiple AI accelerator chips within a server or across multiple servers. These interconnects speed up data transfer and communication between accelerators, allowing them to work together more efficiently on complex tasks like training large AI models.
This link is a critical part of modern AI datacenter systems, and whoever controls the link standard can effectively determine what hardware technology companies use. In line with this, UALink Group aims to establish an open standard that allows multiple companies to contribute and develop advances in AI hardware without being tied to Nvidia's proprietary ecosystem. This approach is similar to other open standards such as Compute Express Link (CXL), created by Intel in 2019, which provides high-speed, high-capacity connections between CPUs and devices or memory in the datacenter.
advertisement
This isn't the first time tech companies have teamed up to take on AI market leaders: In December, IBM and Meta formed the “AI Alliance” along with more than 50 other organizations to promote an open AI model and offer alternatives to closed AI systems from the likes of OpenAI and Google.
Given the market power of Nvidia, the current market leader in AI chips, it may not be surprising that the company is not included in the new UALink promoter group. Nvidia has seen enormous financial success recently and is in a strong position to continue to forge its own path. But as major tech companies continue to invest in their own AI chip development, the need for standardized interconnect technology becomes more pressing, especially as a way to counter (or at least balance) Nvidia's influence.
Accelerating Complex AI
The first version of the proposed standard, UALink 1.0, is designed to connect up to 1,024 GPUs within a single computing “pod,” defined as one or more server racks. The standard is based on technologies such as AMD's Infinity Architecture, and is expected to offer increased speeds and reduced latency for data transfer compared to existing interconnect specifications.
The group plans to establish the UALink Consortium in the second half of 2024 to manage the continued development of the UALink specification. Upon joining, member companies will have access to UALink 1.0, with a higher-bandwidth version, UALink 1.1, scheduled for release in the fourth quarter of 2024.
The first UALink products are expected to launch within the next two years, potentially giving Nvidia plenty of lead time to expand its own lock-in as the AI ​​data center market grows.
