High-speed interconnectivity is the critical enabler for scaling modern computing infrastructure, according to Wань Вей, chief engineer of Sugon’s high-speed network solutions department. Speaking at the 8th China Conference on Supercomputing (ChinaSC 2026), he emphasized that systemic innovations in this domain are essential for enabling significantly larger computational resources to collaborate effectively.
Wань Вей highlighted Sugon’s proprietary scaleFabric high-speed network, which integrates network architecture with communication optimization. This solution aims to reduce data exchange latency between compute resources and minimize the impact of network failures on task execution through rapid fault recovery. These advancements are crucial for the efficient operation of massive intelligent computing clusters.
Earlier, Sugon unveiled its first fully independently developed 400G lossless high-speed network, scaleFabric. The company asserts that all components, from the foundational 112G SerDes IP, switching chips, and network cards to the complete systems, drivers, and management software, were developed in-house.
Official specifications for the scaleFabric 400G network card indicate a port bandwidth of 400 Gbps and an end-to-end latency of up to 0.9 microseconds. The switch offers a single port bandwidth of 800 Gbps, with a system-wide bidirectional throughput of 64 Tbps and a switching latency around 260 nanoseconds. Connection recovery time after a failure is reported to be less than 1 millisecond.
Sugon also claims that the scalability of a single sub-network connection can reach up to 114,000 cards, a 2.33x increase compared to traditional InfiniBand networks. Furthermore, switch port density has reportedly increased by 25%, potentially reducing the overall cost of network construction by approximately 30%.
Additionally, in October, it was announced that under Sugon’s leadership, the development of the ‘Technical Regulations for High-Speed Network Management of Intelligent Computing Clusters’ was approved. This standard has become the first national RDMA (Remote Direct Memory Access) standard for high-speed networks to receive approval in China.
The scaleFabric network has already been deployed in the core of the national supercomputing internet in Zhengzhou, supporting three scaleX intelligent computing clusters, each comprising 10,000 cards, for a total capacity of 30,000 cards.
