Huawei Ascend 960 SuperPoD uses NPO technology to power next-generation AI infra

Huawei unveiled the Ascend 960 SuperPoD, a next-generation AI infrastructure platform built to address the communication and scaling limits of ultra‑large models. Traditional multi‑server setups face growing bottlenecks as inter‑chip traffic rises - communication can consume over 40% of training time in very large clusters - so the SuperPoD architecture ties many compute nodes into a single, unified system with shared memory and high‑speed interconnects. The Ascend 960 SuperPoD supports up to 4,096 NPU cards, delivers up to 8 exaFLOPS (FP8) of compute and up to 1 PB of HBM, reframing AI performance as a systems engineering problem rather than just a chip contest.

A key technical leap is Near‑Packaged Optics (NPO), which moves optical components closer to chips to shorten high‑speed electrical paths and improve density, latency and power efficiency. Huawei’s Hi‑ONE NPO engines provide 7.2 Tb/s per unit; roughly 5,500 Hi‑ONE units can replace about 48,000 conventional 800G optical modules, cutting power draw by more than 550 kW while doubling mean time between failures and reaching 99.8% availability. The focus on NPO, UnifiedBus interconnects and integrated storage underscores a shift from isolated chip metrics to end‑to‑end cluster design, enabling ambitions such as million‑card deployments and continuous, pervasive AI across devices and services.