September 30 — H3C, under New Unisplendour Group, launched the rack-level fully interconnected UniPoD S81000 X1 super node. A single Scale-Up domain supports up to 40 domestic accelerator cards, using four dedicated Scale-Up switch trays to build a one-level CLOS full-mesh topology. P2P bandwidth between any two cards reaches 448GB/s.
The UniPoD S81000 X1 uses an orthogonal connection architecture, with compute and switch nodes directly linked via high-speed connectors, eliminating cables and backplanes. The design is pure electrical interconnect with no optical modules, which the company says improves physical reliability by 10x.
On compute and memory, 40 cards aggregate into a 5760GB unified addressable memory pool, supporting FP8 precision with up to 28P FLOPS. It fits 19-inch standard racks, supports decoupled rack delivery, and can host up to two 40-card super nodes per rack. Cooling uses a hybrid liquid-cooling-primary, air-cooling-compatible design.
Traditional 8-card servers rely on Ethernet stacking, achieving only microsecond latency and 100G bandwidth between cards. For trillion-parameter MoE models, All-to-All communication leaves GPUs waiting on data. AI agents running 24/7 also amplify failure probability through multi-agent chains. Traditional clusters have many hardware components, low troubleshooting efficiency, and rising O&M costs.
ICgoodFind: H3C's UniPoD S81000 X1 is a domestic AI super node with high-bandwidth full interconnect, built for MoE models and AI agent workloads.
ICgoodFind Summary: H3C's UniPoD S81000 X1 targets the interconnect bottleneck in large-model training and AI agent deployment, with a cable-free, all-electrical design that pushes domestic Scale-Up performance forward.
