Launch of UniPoD S81000 X1 Super Node

On September 29, 2026, a new rack-level fully interconnected intelligent computing product, UniPoD S81000 X1 Super Node, was launched by a subsidiary of the new Ziguang Group. This product supports a maximum of 40 domestic acceleration cards per Scale-Up domain and achieves a first-level CLOS fully interconnected topology through four dedicated Scale-Up switch trays. The peer-to-peer bandwidth between any two cards reaches 448GB/s.

The S81000 features an orthogonal connection design, allowing computational nodes to connect directly to switch nodes via high-speed connectors without the need for cables and backplanes. The product employs pure electrical interconnection, eliminating optical modules, and the manufacturer claims physical reliability has improved tenfold.

In terms of memory and computing power, the 40 cards aggregate into a unified memory pool of 5760GB, supporting FP8 precision, which is commonly used for AI training and inference acceleration. The total computing power can reach 28P FLOPS.

The product is compatible with standard 19-inch racks and supports decoupled cabinet delivery, with a maximum deployment of two sets of 40-card super nodes per cabinet. Cooling is primarily achieved through liquid cooling, compatible with a hybrid air-liquid architecture.

Traditional eight-card servers rely on Ethernet stacking for networking, resulting in microsecond-level latency and bandwidth in the hundreds of Gb/s. Under the communication demands of trillion-parameter mixture of experts large models, significant GPU time is spent waiting for data. Additionally, the operation of intelligent agents 24/7, with multiple agents in series, leads to a notable composite failure rate. Traditional cluster components are complex, with slow fault localization, and operational costs grow linearly with scale.