How Huawei plans to bypass US chip curbs with its new UnifiedBus technology

Huawei Technologies announced on Thursday a new computing architecture that could potentially link up to 1 million processors in a single computer, thanks to its UnifiedBus interconnect technology. The Peerium Computing Architecture leverages nested parallelism, unified memory addressing, and peer-to-peer interconnect, with UnifiedBus acting as the high-speed fabric that connects processors, memory, storage, and networking equipment.

An Atlas 950 SuperCluster containing 256,000 computing cards is already operational, while a next-generation Atlas 960 system using near-packaged optics (NPO) is undergoing testing. Known in Chinese as Lingqu, UnifiedBus has become pivotal to Huawei's strategy to increase computing power at the system level, a necessity given US restrictions that limit the company's access to cutting-edge semiconductor manufacturing.

The technology utilizes a single open protocol to link CPUs, NPUs, memory, solid-state drives, network interface cards, and switches, aiming to unify a system containing up to 1 million processors. Unlike traditional AI clusters that rely on conventional network connections like Ethernet, UnifiedBus promises higher bandwidth and lower latency between processors, memory, and storage modules.

This allows for more efficient data sharing and communication through a peer-to-peer approach. Huawei's motivation for developing UnifiedBus stems from US restrictions, but company officials emphasize it's a forward-thinking solution rather than a workaround. Huawei's rotating chairman, Eric Xu, described UnifiedBus as a "path for the future" in the AI era, stating that the technology is open to all and that other industry players will likely follow suit.

Despite facing limitations due to the availability of sophisticated process nodes in China, Huawei believes its approach can compete at the system level, an area where it holds global leadership.