Huawei releases computing architecture for the AI era: turning millions of processors into one computer

During the Huawei Full Connectivity (HC) Conference held in Shanghai recently, Huawei Rotating Chairman Xu Zhijun and HiSilicon Chief Scientist Dr. Liao Heng conducted a Q&A exchange with some media reporters on Huawei's newly released Peerium computing architecture and Unified Bus for the AI ​​era.

Xu Zhijun said that at last year’s HC conference, the Ascend chip roadmap was released and the Lingqu protocol was opened. Behind the Ascend 950 and Atlas 950 super nodes exhibited this year is Peerium, a new computing architecture pioneered by Huawei. This architecture is based on nested parallelism, unified memory addressing, and equal interconnection. It can turn millions of processors into one computer, breaking through the serialization of the Turing paradigm.

Liao Heng pointed out that the cluster design of 256,000 cards is not random, but based on the current needs of 5 trillion parameter model training and the design considerations of China's data center power transmission system. In the next two years, China is expected to have multiple basic models with training parameter sizes ranging from 10 trillion to 40 trillion.

Regarding chip autonomy, Xu Zhijun emphasized that China’s advancement of chip independence is an inevitable path. Although there is still a gap between Shengteng and the international leading level in terms of advanced nature, solving the "have or have not" problem is the core. At present, Shengteng's domestic market demand far exceeds supply, so there are currently no plans to fully expand overseas markets.

Regarding Lingqu Bus (UB), Xu Zhijun explained that UB is not only the contribution of the network department, but a unified protocol developed by the computing department, aiming to achieve low-latency, ultra-high-speed computing interconnection and completely subvert the original master-slave architecture. Liao Heng added that UB uses a unified protocol, which can save an order of magnitude in latency compared to the protocol conversion between NVIDIA NVLink and InfiniBand.

On the optical interconnection technology route, Huawei firmly follows the NPO (onboard optics) route. Xu Zhijun believes that compared with CPO, NPO has significant advantages in cost, maintenance and yield rate. Huawei has currently released the NPO-based Hi-ONE module, which is expected to lead the industry in the next two to three years.

Finally, Xu Zhijun said that Huawei will continue to persist in innovation and build product competitiveness through intensive R&D investment. He believes that UB interconnection technology and Peerium computing architecture are the only way forward in the AI ​​era. As the industry's demand for large-scale computing power grows, more manufacturers will move in this direction in the future.