Huawei Unveils Record-Breaking Atlas 960E SuperPoD AI Computing Cluster

Huawei has unveiled its latest AI computing cluster, the Atlas 960E SuperPoD, which connects 4,096 of its Ascend neural processing units into a single logical machine with unified memory addressing.

The system allows for seamless communication between the chips, enabling it to train models with up to 10 trillion parameters. This is a significant increase from previous Atlas generations, which delivered between 2.3 and 4 times the training and inference throughput for large models.

Huawei's Hi-ONE Near-Packaged Optics engine plays a crucial role in this achievement, pushing 7.2 terabits per second per module and reducing power consumption by more than 550 kilowatts per pod.