Huawei Unveils Peerium Computing Architecture to Scale AI Workloads Across Millions of Processors
Huawei announced Peerium, a new computing architecture designed for the AI era that enables up to one million processors to operate together as a single computer. The architecture relies on nested parallel execution, unified memory addressing, and an open peer-to-peer interconnect protocol called Lingqu (UnifiedBus). By moving away from conventional master-slave hierarchies and overcoming Von Neumann single-machine limits, Peerium provides a new paradigm to scale massive AI training and inference tasks across unprecedented cluster sizes. This could allow Huawei to build ultra-large supercomputing nodes to meet rapidly growing AI compute demand. The architecture introduces a Nested Bulk Synchronous Parallel (Nested BSP) model for recursive task management and is first embodied in the Atlas 950 superpod, which is deploying a 256,000-card cluster. Huawei is also testing the next-generation Atlas 960 system, which incorporates Near-Package Optics (NPO) technology.
## BACKGROUND
Traditional computers follow the Von Neumann architecture, where a processing unit executes instructions sequentially under a master-slave design, creating memory and bandwidth bottlenecks when scaling up thousands of chips. Bulk Synchronous Parallel (BSP) is a classic parallel programming framework that divides computations into stages separated by global synchronization barriers. Near-Package Optics (NPO) places optical transceivers immediately adjacent to the switch ASIC on the board, significantly boosting bandwidth and energy efficiency in high-density supercomputers.