Huawei unveiled the Atlas 960E SuperPoD computing cluster at the HUAWEI CONNECT 2026 event in Shanghai on September 17. The system integrates up to 4,096 Ascend Neural Processing Units using a unified memory-addressing architecture to eliminate data transfer delays.
The cluster provides 8 EFLOPS for FP8 and 16 EFLOPS for FP4 computations, supported by 1 petabyte of High Bandwidth Memory. Huawei reports that the new model offers 2.3 to 4 times higher performance for training and inference tasks compared to the Atlas 950.
The design replaces 48,000 conventional 800G optical modules with 5,500 Hi-ONE Near-Packaged Optics units, reducing power consumption by over 550 kilowatts per pod. The system achieves 99.8% availability and doubles the mean time between failures.
Huawei aims to scale its AI framework to 256,000 nodes, with future SuperCluster configurations planned to exceed one million NPUs. Larger versions of the Atlas 960 platform are scheduled for release in the fourth quarter of 2027.
