H3C Unveils UniPoD S81000 X1 Supernode Supporting 40 Domestic GPUs
H3C launched the UniPoD S81000 X1, a rack-scale supernode that aggregates up to 40 domestic GPUs into a single scale-up domain with 448 GB/s peer-to-peer (P2P) bandwidth. The architecture features a single-tier CLOS topology and cableless orthogonal connectors to deliver 28 PFLOPS of FP8 compute power and a 5,760 GB unified memory pool. By expanding GPU scale-up domains without relying on optical modules, this hardware architecture addresses latency bottlenecks in All-to-All communication for trillion-parameter Mixture-of-Experts (MoE) models and multi-agent workflows. It provides a dense, reliable, domestic hardware platform to support China's growing AI infrastructure needs. The system utilizes a direct orthogonal connection between compute and switch nodes using high-speed electrical connectors, eliminating backplanes and cables to reportedly boost physical reliability by 10x. A standard 19-inch rack can deploy up to two 40-GPU supernodes using a liquid-cooling dominant architecture.
## BACKGROUND
Traditional AI server clusters stack standard 8-GPU servers connected over Ethernet networks, which creates microsecond-level latencies and bandwidth bottlenecks during large-scale LLM training. CLOS network topologies and orthogonal connection designs are advanced hardware techniques used to enable high-density, non-blocking interconnects between chips while minimizing signal loss and cable clutter.