Huawei Unveils Peerium Computing Architecture for AI Era, Aiming to Unite a Million Processors
Huawei announced Peerium, an AI-era computing architecture it says can make a million processors act as one computer. The company also introduced Lingqu interconnect and said the Atlas 950 supernode is its first product.
Huawei said Peerium is based on nested parallelism, unified memory addressing and equal interconnection, and is designed for strong scaling across a million processors. It breaks from the Turing paradigm with Nested BSP, and it breaks from the von Neumann single-machine architecture, the company said. The architecture also overturns the long-standing master-slave model, allowing a million processors to truly become one larger computer.
Lingqu is the key interconnect technology for Peerium. According to Huawei, Lingqu is an open protocol that provides an infinitely scalable high-speed bus connecting CPUs, NPUs, memory, SSDs, network cards and switches. With Lingqu, computing, storage and networking can be interconnected on an equal basis, the company said.
Atlas 950 supernode is the first-generation product of the Peerium computing architecture. Huawei said a 256,000-card Atlas 950 supernode cluster is already being deployed, and an NPO-based Atlas 960 system is in testing.
Huawei rotating chairman Xu Zhijun said: “In the AI era, Huawei, based on the pioneering Peerium computing architecture, will continue to build supernodes and clusters that meet customers’ training and inference requirements, making computing ubiquitous and intelligence accessible everywhere.”