Huawei announces Atlas 950 SuperCluster, boasting a 1 ZettaFLOPS FP4 performance and comprising numerous 950DT APUs numbering in the hundreds of thousands.
Huawei Unveils Atlas 950 SuperCluster: A New Powerhouse for AI
Huawei has made a significant stride in the realm of artificial intelligence (AI) with the unveiling of its Atlas 950 SuperCluster. The supercomputer, showcased at the Huawei Connect 2025 conference, is set to redefine AI capabilities.
The Atlas 950 SuperCluster, spanning across 1,000 square meters, is a colossal structure housing 160 total cabinets. Of these, 128 are dedicated to computation, while 32 are for communications. The supercomputer offers up to 524 FP8 ExaFLOPS for AI training and an impressive 1 FP4 ZettaFLOPS for AI inference.
The supercomputer's design is geared towards supporting AI models with hundreds of billions to tens of trillions of parameters. The Atlas 950 SuperCluster employs a brute force approach, utilising hundreds of thousands of AI accelerators to achieve 1 ZettaFLOPS.
The heart of the Atlas 950 SuperCluster is the Atlas 950 SuperPoD, which offers 8 FP8 ExaFLOPS and 16 FP4 ExaFLOPS of compute performance. Each SuperPoD integrates 8,192 Ascend 950DT chips, marking a 20-fold increase in processing units compared to the Atlas 900 A3 SuperPoD.
Looking ahead, Huawei plans to launch the Atlas 960 SuperCluster and Atlas 960 SuperPoD in the fourth quarter of 2027. The Atlas 960 SuperCluster will scale up to more than 1 million Ascend 960 NPUs and provide 2 FP8 ZettaFLOPS and 4 MXFP4 ZettaFLOPS of performance.
The total site footprint of the Atlas 960 SuperCluster could be significantly larger than 64,000 m2, as it will require additional space for power rooms, chillers/cooling towers, battery/UPS systems, and support offices.
In terms of competition, Huawei's SuperCluster is expected to compete with Nvidia's Rubin-based systems in late 2026. The peak performance figures for the Atlas 950 SuperCluster are just behind leading-edge AI supercomputers, such as Oracle's OCI Supercluster.
However, it remains to be seen whether the Atlas 950 SuperCluster can achieve its peak performance numbers in real life. The supercomputer supports both RoCE and the proprietary UBoE protocols for communication, with the Atlas 960 SuperCluster expected to support UBoE for improved latency and uptime metrics.
In conclusion, the Atlas 950 SuperCluster is one of the most powerful supercomputers for artificial intelligence on the planet, setting a new benchmark for AI capabilities.
Read also:
- U Power's strategic collaborator UNEX EV has inked a Letter of Intent with Didi Mobility to deploy UOTTA(TM) battery-swapping electric vehicles in Mexico.
- Commercial-grade hydrogen enhancement systems manufacturing initiated by H2i Technology
- Gold nanorod market to reach a value of USD 573.3 million by 2034, expanding at a compound annual growth rate (CAGR) of 11.7%
- China proposes establishing 39 AI data centers equipped with 115,000 limited Nvidia Hopper GPUs, sparking concerns about procurement and the potency of trade restrictions.
