88Z AI, 21 May — Zhipu, Yuxun Network, and Tsinghua University launched ZCube, an architecture that eliminates the Spine layer to address PD separation inference congestion. Tests show it significantly reduces hardware costs while boosting GPU throughput and response speed, providing an efficient foundation for ultra-large-scale inference clusters.
At a glance
- ZCube eliminates the Spine layer and uses flat topology to solve inference congestion
- Tests show a 33% reduction in network costs and a 15% boost in GPU throughput
The update
Zhipu AI, together with Yuxun Network and Tsinghua University, has introduced ZCube, a next-generation network architecture for large model inference. Designed to tackle congestion issues arising from prefill-decode separation, it removes the traditional Spine layer and employs a flat topology with hybrid single and multi-rail access to optimize data transmission. Real-world tests using GLM-5.1 coding demonstrate significant performance gains and cost efficiency. The results indicate a 33% reduction in switch and optical module costs, a 15% increase in GPU inference throughput, and a 40.6% drop in P99 time-to-first-token latency. This innovation provides an efficient and economical infrastructure foundation for building next-generation ultra-large-scale inference clusters.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- 智谱Source published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.