Zhipu, Yuxun Network and Tsinghua propose ZCube network architecture to break large-model inference network bottlenecks
Original title:下一代大模型推理网络架构:ZCube如何有效破解网络瓶颈?
AI summary
Zhipu, Yuxun Network and Tsinghua University jointly proposed and deployed the ZCube network architecture, which replaces traditional Clos/ROFT architectures with a flattened topology to eliminate avoidable congestion in PD-disaggregated inference at the structural level. In tests on a thousand-GPU cluster in the GLM-5.1 coding production environment, ZCube raised average GPU inference throughput by more than 15% over ROFT, cut TTFT P99 by 40.6% and reduced switch and optical module costs by about one third.
Key Takeaways
- Zhipu,Yuxun Network and Tsinghua University jointly proposed and deployed the ZCube network…
- In tests on a thousand-GPU cluster in the GLM-5.
- 1 coding production environment,ZCube raised average GPU inference throughput by more than 15%…
- 6% and reduced switch and optical module costs by about one third.
Quick Facts
- Published
- Updated
- Source
- 公众号:智谱(GLM)
- Category
- Papers & Research
- AI score
- 76
架构创新不只是省钱,ZCube 在 GLM-5.1 集群里用三分之一交换机成本换来 15% 吞吐提升和 40% 尾时延改善,为 PD 分离推理提供了一条可复用的组网路径。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: Zhipu, Yuxun Network and Tsinghua propose ZCube network architecture to break large-model inference network bottlenecks. May 21, 2026. https://new.nat6.net/en/news/cmsdu107800u2rofz68f17j53/