Huawei Introduces Peerium Computing Architecture for AI Era

Huawei announced on September 17 in Shanghai its new Peerium computing architecture, tailored for the artificial intelligence era. The company states that this architecture allows millions of processors to collaborate as a single computational system, addressing the growing need for AI processing power.

The Peerium architecture is built upon principles of nested parallelism, unified memory addressing, and peer-to-peer interconnectivity. This design aims for high scalability, enabling the coordination of millions of processors. It introduces "Nested BSP" (Batch Synchronous Parallel), a computational paradigm intended to overcome the limitations of traditional von Neumann and Turing machine architectures.

Huawei claims the Peerium architecture redefines the conventional master-slave model where components execute tasks sequentially. Instead, it facilitates an equal interconnection between various computing elements, including CPUs, NPUs (Neural Processing Units), memory, SSD storage, and network devices, ensuring a more integrated approach.

The core enabling technology for this architecture is "Lingqu" (灵衢), an interconnect technology based on an open protocol. It is designed for virtually unlimited expansion, connecting CPUs, NPUs, memory, SSDs, network cards, and switches to achieve an equitable connection between computation, storage, and networking. The first product based on the Peerium architecture is the "Atlas 950" supernode. Huawei reports that a cluster of 256,000 Atlas 950 cards is currently being deployed.

"In the AI era, we continue to build supernodes and clusters that meet customer needs for training and inference with our pioneering Peerium computing architecture, making computing power ubiquitous and intelligence all-encompassing," said Huawei's Rotating Chairman, Eric Xu.