At the HUAWEI CONNECT 2026 event in Shanghai on September 17, Huawei introduced the Atlas 960E SuperPoD. This AI computing cluster integrates 4,096 Ascend neural processing units into a single logical machine with unified memory addressing, designed to support models with up to 10 trillion parameters.
The system provides 8 EFLOPS at FP8 precision and 16 EFLOPS at FP4 precision, with a total capacity of 1 petabyte of high-bandwidth memory. To manage data exchange, Huawei implemented the Hi-ONE Near-Packaged Optics engine, which achieves 7.2 terabits per second per module and reduces power consumption by over 550 kilowatts per pod.
The 960E architecture offers 2.3 to 4 times the throughput of previous Atlas generations for large models, with a reported operational availability of 99.8%. Huawei also announced the OceanStor M900 storage solution and upgrades to the TaiShan 950 SuperPoD.
The company accelerated the release of the Ascend 960DT training chip to Q1 2027, moving the launch three quarters ahead of the original schedule. Huawei also stated a long-term objective to scale its infrastructure to 1 million NPUs using multi-rail topology.
