From ATOP to ZCube: Automated Topology Optimization Pipeline and a Cost-Effective Topology for Large Model Training#

Zihan Yan, Dan Li et al. (Tsinghua, Zhongguancun Lab, ByteDance)

Why: DC topology design for training clusters is core physical-networking work, and an automated search plus a concrete new topology is more useful than another hand-tuned fabric.

What: A pipeline that treats topology as hyperparameters to search at tens-of-thousands-of-GPU scale, yielding ZCube, which claims 3–7% faster training and 26–46% lower hardware cost vs ROFT/Rail-only/HPN.

How: Models topology as an optimization problem over training traffic, collective performance, fault tolerance, and cost; evaluated in build/optimize/expand scenarios and validated on a real testbed (25% cost cut vs rail-optimized).

Where applicable: General topology-design methodology for GPU clusters; results are simulation-heavy with a testbed check.