~/HIGH SPEED I/sugon-highlights-high-speed-interconnects-for-scaling-ai-computing-clusters

Sugon Highlights High-Speed Interconnects for Scaling AI Computing Clusters

Sugon's chief engineer highlighted high-speed interconnects as vital for scaling AI clusters, backed by their self-developed scaleFabric 400G lossless network which achieves a sub-net scale of up to 11.4 thousand cards. As AI model sizes grow rapidly, overcoming communication bottlenecks between computing nodes is critical for the efficient operation of large-scale supercomputing clusters. Sugon's homegrown infrastructure provides a viable domestic alternative that reduces networking costs and improves cluster efficiency. The scaleFabric 400G network features a port bandwidth of 400Gbps, an end-to-end latency of 0.9 microseconds, and link failure recovery times under 1 millisecond. It has already been deployed at the Zhengzhou core node of the National Supercomputing Internet, supporting three 10,000-card scaleX clusters.

## BACKGROUND

RDMA (Remote Direct Memory Access) is a key networking technology that allows computers to exchange data in memory without involving the operating system or CPU, significantly reducing latency and boosting throughput. SerDes (Serializer/Deserializer) IP cores are essential hardware components used in semiconductor designs to enable high-speed serial data communication across chips and devices.

## REFERENCES

## KEYWORDS

#High-Speed Interconnect#AI Infrastructure#Supercomputing#RDMA#Sugon

$ subscribe --daily