The formation of the Ultra Accelerator Link (UALink) Promoter Group by major tech companies, including Google, Intel, Microsoft, Meta, AMD, Hewlett Packard Enterprise, Cisco, and Broadcom, marks a significant development in the AI accelerator chip industry. The purpose of this group is to establish a new interconnect standard for AI accelerator chips in data centers, challenging Nvidia’s proprietary NVLink interconnect technology.
At the heart of AI systems today are GPUs, which are capable of running parallel matrix multiplications necessary for neural network architecture. However, the complexity of AI systems often requires multiple GPUs. This is where interconnect technologies like NVLink come into play, enabling the connection of multiple AI accelerator chips within a server or across multiple servers. These interconnects facilitate faster data transfer and communication between the accelerators, enhancing their efficiency in complex tasks such as training large AI models.
The establishment of an open standard for interconnects is crucial in the AI data center industry. The company that controls the link standard has significant influence over the choice of hardware used by tech companies. The UALink group aims to break away from Nvidia’s proprietary ecosystem and create an open standard that encourages multiple companies to contribute and develop AI hardware advancements. This approach aligns with other open standards like Intel’s Compute Express Link (CXL), which provides high-speed connections in data centers.
This move by tech companies to counter the dominance of Nvidia is not unprecedented. In December, IBM and Meta, along with over 50 other organizations, formed an “AI Alliance” to promote open AI models and offer an alternative to closed AI systems. Nvidia’s current market leadership and financial success make it understandable why the company has not joined the UALink Promoter Group. However, as other tech companies continue to invest in their own AI chip development, the need for a standardized interconnect technology becomes increasingly important to balance Nvidia’s influence.
The proposed UALink 1.0 standard aims to connect up to 1,024 GPUs within a single computing “pod,” which can include one or several server racks. Based on technologies like AMD’s Infinity Architecture, UALink 1.0 is expected to improve speed and reduce data transfer latency compared to existing interconnect specifications. The UALink Consortium, planned to be formed later in 2024, will oversee the ongoing development of the UALink specification. Member companies will have access to UALink 1.0 upon joining, with a higher-bandwidth version, UALink 1.1, scheduled for release in Q4 2024.
While the first UALink products are expected to be available within the next two years, Nvidia still holds a significant position in the AI data center market. The company may have ample time to expand its proprietary lock-in as the market continues to grow. However, the formation of the UALink Promoter Group signals a shift in the industry toward open standards and collaboration among tech companies, ultimately giving customers more choices and fostering innovation in AI hardware.
Source link
