On Thursday, Huawei Technologies Co. unveiled what it describes as the world’s most powerful AI computing clusters—SuperNode and SuperCluster—signifying a significant advancement in high-performance computing despite the ongoing US sanctions.
During an event in Shanghai, Huawei’s rotating chairman, Xu Zhijun, stated: “Computing power has been and will continue to be the cornerstone of artificial intelligence—particularly for the development of AI in China.”
He emphasised that Huawei aims to sustainably fulfil long-term computing needs by constructing SuperPoDs and SuperClusters utilising semiconductor manufacturing process nodes that are practically accessible within the Chinese mainland.
Despite restrictions that hinder Huawei’s access to advanced chip manufacturing, the company has successfully created large-scale AI systems that combine thousands of its Ascend processors into cohesive computing units, achieving top-tier performance through innovations in interconnect technology.
Also Read: Customs Unveils Automated Clearance System to Tackle Port Congestion
A SuperPoD functions as a single logical machine, composed of several physical machines that can learn, think, and reason collectively. Xu introduced the latest SuperPoD offerings: the Atlas 950 SuperPoD (featuring 8,192 Ascend NPUs) and the Atlas 960 SuperPoD (equipped with 15,488 Ascend NPUs).
These two SuperPoDs are set to provide industry-leading performance across various critical metrics, including the number of NPUs, overall computing power, memory capacity, and interconnect bandwidth.
Huawei asserted that, based on publicly disclosed product roadmaps from industry peers, these SuperPoDs currently represent the most powerful SuperPoDs globally and are expected to maintain this status for years ahead.
Xu also revealed the Atlas 950 SuperCluster (which includes over 500,000 Ascend NPUs) and the Atlas 960 SuperCluster (which has over one million Ascend NPUs), both of which are extensive computing clusters made up of multiple Huawei SuperPoDs. These clusters are also anticipated to surpass all other computing clusters available in the market.
