As frontier AI models surpass trillions of parameters—highlighted by recent deployments like Moonshot AI's 2.8-trillion parameter model—traditional GPU stacking is hitting physical limits. To overcome single-chip constraints and US export controls, Chinese tech firms are shifting from card-level scaling to system-level "supernodes".
At WAIC, hardware giants including Huawei, Sugon, Moore Threads, and Biren showcased multi-card interconnected supernodes scaling from 64 to over 1,024 cards per unit, utilizing optical interconnects (NPO/CPO) and liquid cooling to drastically reduce inter-card latency. Sugon even debuted its 100,000-card Dawn 8000 cluster.
Industry Perspective
This transition marks a pivotal shift in AI infrastructure. By aggregating clusters into single logical computers, Chinese manufacturers are bridging raw hardware performance gaps with architectural efficiency. The supernode strategy proves that network topology, low-latency interconnects, and cabinet-level power delivery are now as critical as the underlying silicon itself in the generative AI race.