On Thursday, leading technology firms including Google, Intel, Microsoft, Meta, AMD, Hewlett-Packard Enterprise, Cisco, and Broadcom revealed the formation of the Ultra Accelerator Link (UALink) Promoter Group. This alliance aims to develop a new interconnect standard for AI accelerator chips in data centers, offering an alternative to Nvidia’s exclusive NVLink technology, which currently connects multiple servers for advanced AI applications like ChatGPT.
In today’s AI landscape, GPUs play a crucial role by performing extensive matrix multiplications necessary for neural network architectures. Complex AI systems often require more than one GPU, and Nvidia’s NVLink connects multiple AI accelerator chips within a server or across servers. These connections facilitate faster data transfer and communication, enhancing efficiency in tasks such as training large AI models.
Interconnects are vital to modern AI data centers, and the entity controlling the interconnect standard can significantly influence hardware choices in the tech industry. The UALink group aims to establish an open standard, enabling multiple companies to collaborate and innovate in AI hardware, rather than being confined to Nvidia’s proprietary ecosystem. This initiative mirrors other open standards like Compute Express Link (CXL), introduced by Intel in 2019, which enables high-speed, high-capacity connections between CPUs and devices or memory in data centers.
This isn’t the first effort to counteract a dominant player in the AI market. In December, IBM and Meta, along with over 50 other organizations, formed an “AI Alliance” to promote open AI models, providing alternatives to closed systems from companies like OpenAI and Google.
Given Nvidia’s dominance in the AI chip market, it’s not surprising that the company has opted out of the new UALink Promoter Group. Nvidia’s recent financial success has fortified its market position, allowing it to continue advancing its proprietary technologies. However, as more tech giants invest in AI chip development, the necessity for a standardized interconnect technology grows, offering a means to counterbalance Nvidia’s market influence.
Enhancing AI Performance
The initial version of the proposed standard, UALink 1.0, is designed to connect up to 1,024 GPUs within a single computing pod, which can consist of one or several server racks. Drawing on technologies like AMD’s Infinity Architecture, this standard is expected to enhance speed and reduce data transfer latency compared to existing specifications. The group plans to establish the UALink Consortium later in 2024 to oversee the ongoing development of the UALink specification. Member companies will gain access to UALink 1.0 upon joining, with a higher-bandwidth version, UALink 1.1, scheduled for release in Q4 2024.
The first UALink products are anticipated to be available within the next two years, providing Nvidia with a window to expand its proprietary ecosystem as the AI data center market continues to grow.