华为于9月17日在上海发布新型计算架构Peerium，通过嵌套批量同步并行、统一内存寻址和平等互联等关键技术，实现百万级处理器作为一台计算机协同工作1。该架构突破了冯·诺依曼单机架构和传统图灵范式的限制，颠覆了长久以来的主从架构模式1。
Atlas 950超节点是该架构的首代产品，其25.6万卡规模的集群已在部署中1。华为轮值董事长徐直军表示："AI时代，华为基于开创的Peerium计算架构，持续打造满足客户训练和推理需求的超节点和集群，让算力无处不在，让智能无所不及"1。同时，基于NPU的Atlas 960系统正在测试中1。
Huawei introduced its groundbreaking Peerium computing architecture on September 17 in Shanghai, marking a significant departure from traditional computational frameworks 1. The new architecture transcends the limitations of von Neumann single-machine design and conventional Turing paradigms by enabling up to one million processors to operate cohesively as a unified computing system 1.
The architecture leverages three core technologies: Nested BSP (Nested Bulk Synchronous Parallel) processing, unified memory addressing, and peer-to-peer interconnection capabilities 1. The first commercial embodiment of this innovation is the Atlas 950 supernode, with a 256,000-card cluster already undergoing deployment 1. Additionally, an Atlas 960 system based on NPU technology is currently in testing phases 1.
Xu Zhijun, Huawei's rotating chairman, underscored the strategic significance of the breakthrough, stating: "In the AI era, Huawei, leveraging the pioneering Peerium computing architecture, continues to develop supernodes and clusters that fulfill customers' training and inference requirements, making computing power ubiquitous and intelligence boundless" 1.
评论
还没有评论，欢迎留下第一条。