Huawei's Xu Zhijun: Peerium architecture and Lingqu bus turn a million processors into a single computer.

During Huawei Connect in Shanghai, Huawei Rotating Chairman Xu Zhijun and HiSilicon Chief Scientist Liao Heng held a Q&A with media reporters on Huawei's newly released Peerium computing architecture for the AI era and the Lingqu bus. During this year's HC conference, Huawei showcased the Ascend 950 and Atlas 950 super nodes. Xu Zhijun said that "super nodes" are now widely mentioned, but implementations vary greatly among vendors, and the real challenge lies in making thousands or tens of thousands of processors form a single computer. Huawei named this new computing architecture Peerium, which can make a million processors become one computer. Currently, Huawei is deploying and testing the Atlas 950 SuperPoD super node at a scale of 256,000 cards. Liao Heng believes that within the next two years, China is expected to see about 6 to 7 frontier AI laboratories, all aiming to train foundation language models with parameter scales between 10 trillion and 40 trillion. These numbers roughly match the memory capacity and scale of super nodes, so infrastructure of this scale is needed for training.