ZCube Network Architecture Reduces Costs and Improves Performance in Large Model Inference

AIMPACT Message, May 21 (UTC+8): According to monitoring by Beating, in response to the growing structural network congestion challenges in large model PD (Prefill-Decode) separation deployment, Zhipu, Yuxun Network, and Tsinghua University have collaborated to develop and deploy the ZCube networking architecture in the GLM-5.1 coding production environment with a thousand GPUs. As long-context and PD-separated inference become mainstream, cross-node KV Cache transmission has caused severe asymmetry in inference traffic, making traditional ROFT (Rail-Optimized Fat-Tree) architectures highly prone to local hotspots and link conflicts. ZCube eliminates the Spine layer switches and adopts a fully flattened topology (2-hop network diameter), combined with a hybrid single/multi-track access mechanism, achieving end-to-end traffic load balancing across all switches in the network at the architectural level. In benchmark tests on actual production clusters, while keeping GPUs, software stacks, and applications unchanged, the ZCube architecture reduced switch and optical module hardware costs by 33% compared to traditional architectures, increased average GPU inference throughput by 15%, and reduced the P99 first-token latency (TTFT) by 40.6%. (Source: BlockBeats)