Huawei's Atlas 960 SuperCluster to Revolutionize AI Processing in 2027
Huawei has announced plans to release the Atlas 960 SuperCluster in early 2027, a powerful supercomputer designed to revolutionise artificial intelligence processing. The new system promises unprecedented performance, outpacing its competitors.
The Atlas 960 SuperCluster will feature more than 1 million Ascend 960 NPUs, delivering up to 4 ZettaFLOPS of performance in FP4 precision. This means it can handle complex AI tasks with ease.
The supercomputer is designed to support both AI training and inference workloads. It offers up to 524 FP8 ExaFLOPS for training and up to 1 FP4 ZettaFLOPS for inference, making it a versatile tool for AI developers.
Huawei has also revealed that the Atlas 950 SuperCluster, its predecessor, consists of 64 Atlas 950 SuperPoDs. Each of these runs hundreds of thousands of Ascend 950DT neural processing units (NPUs). This system is capable of handling AI models with hundreds of billions to tens of trillions of parameters.
The Atlas 950 SuperCluster supports both RoCE and Huawei's proprietary UBoE protocols, ensuring compatibility with a wide range of systems.
Huawei's upcoming Atlas 960 SuperCluster promises to be a game-changer in AI processing. With its release in early 2027, it aims to compete with Nvidia's Rubin-based systems, offering unparalleled performance and versatility.
