Huawei unveils Atlas 950 SuperCluster – promises 1 ZettaFLOPS FP4 performance (www.tomshardware.com)

At Huawei Connect 2025, Huawei announced the Atlas 950 SuperCluster, a data-center scale AI system that stitches together 524,288 Ascend 950DT NPUs across 64 Atlas 950 SuperPoDs (each PoD contains 8,192 chips and ~160 cabinets). Huawei claims peak throughput of up to 524 FP8 ExaFLOPS for training and up to 1 MXFP4 ZettaFLOPS for inference, supported by optical interconnects across more than 10,240 cabinets and dual networking options: industry-standard RoCE and Huawei’s proprietary UBoE (promised lower idle latency and fewer switches/modules). The SuperPoD footprint is large—about 1,000 m² per PoD—so a full SuperCluster would occupy roughly 64,000 m², requiring substantial power/cooling infrastructure. The announcement is significant because it signals Huawei’s strategy of achieving top-tier AI performance via extreme scale rather than per-chip superiority, positioning the Atlas 950 to challenge Rubin-era GPU clusters from Nvidia by 2026–2027. Key implications: the system targets training and inference of models from hundreds of billions to tens of trillions of parameters and promises high interconnect bandwidth for dense and sparse workloads, but the numbers are theoretical peak FLOPS and real-world efficiency, utilization, and power/performance trade-offs remain open questions. Huawei also previewed an Atlas 960 roadmap (2027) scaling to >1M NPUs and multi-zettaFLOPS peaks, underscoring continued focus on massive-scale AI fabrics.