H3C, a subsidiary of New Tsinghua Unigroup, has launched the UniPoD S81000 X1 super node, a rack-scale intelligent computing product. The system supports up to 40 domestic GPU accelerator cards within a single scale-up domain. It utilizes four dedicated scale-up switching trays to establish a CLOS full-interconnect topology, achieving a P2P bandwidth of 448 GB/s between any two cards.
The hardware features an orthogonal connection design that links computing and switching nodes via high-speed connectors, eliminating the need for cables or backplanes. By using pure electrical interconnection without optical modules, the manufacturer claims a tenfold increase in physical reliability. The system is compatible with standard 19-inch racks and supports decoupled delivery, allowing up to two 40-card nodes per rack.
The 40-card configuration provides a unified addressable memory pool of 5760 GB. It supports FP8 precision for AI training and inference, delivering a total computing power of 28P FLOPS. The cooling architecture primarily relies on liquid cooling, while also maintaining compatibility with air-cooled systems to manage thermal loads effectively.
H3C developed this solution to address the limitations of traditional 8-card servers, which rely on Ethernet stacking and suffer from microsecond-level latency and lower bandwidth. The company notes that AI agents now place significant pressure on computing resources, with token consumption per task reaching hundreds of thousands to millions, far exceeding traditional conversational AI requirements.
