Shanghai: TECHz Magazine – News Desk
Huawei has officially unveiled the Atlas 960E SuperPoD, a massive artificial intelligence computing cluster built to support the training and inference of next-generation AI models exceeding 10 trillion parameters. Unveiled at Huawei Connect 2026 in Shanghai, the system represents a significant shift in hyperscale AI architecture by focusing on highly integrated, large-scale system efficiency rather than individual accelerator performance.
The Atlas 960E is the industry’s first SuperPoD based on near-packaged optics (NPO). It relies on Huawei’s new UnifiedBus architecture and the Hi-ONE optical engine to alleviate the extreme interconnect bottlenecks inherent in training massive models. A single Atlas 960E SuperPoD links up to 4,096 Ascend NPUs, delivering 8 EFLOPS of FP8 compute performance and 16 EFLOPS at FP4 precision.
By integrating 5,500 proprietary Hi-ONE optical engines – the industry’s first mass-produced NPO with a built-in light source – the system eliminates the need for roughly 48,000 conventional 800G optical modules. This streamlining cuts system power consumption by over 550 kilowatts and doubles fault-free operating time, achieving 99.8% system availability.
Beyond single units, multiple Atlas 960E clusters can be networked via UnifiedBus into a SuperCluster. Using a standard two-tier Clos architecture, the infrastructure can connect up to 512,000 NPUs, while advanced multi-rail topologies are designed to eventually push that limit to one million interconnected processors.
