UB-Mesh: a Hierarchically Localized nD-FullMesh Datacenter Network Architecture

The scaling of Large-scale Language Models (LLMs) demands unprecedented computational power and bandwidth. We present UB-Mesh, an innovative AI datacenter network architecture that enhances scalability, performance, and cost-efficiency through a hierarchical nD-FullMesh topology. Unlike traditional symmetrical designs, UB-Meshoptimizes LLM training by prioritizing localized data movement and minimizing switch usage. The architecture features UB-Mesh-Pod, a physical implementation of 4D-FullMesh using custom hardware including NPUs, CPUs, Low/High-Radix Switches (LRS/HRS), and NICs, interconnected via our Unified Bus (UB) technology for dynamic resource allocation. For network optimization, we introduce All-Path-Routing (APR) to efficiently manage data traffic. Combined with topology-aware performance tuning and robust reliability mechanisms like 64 + 1 backup, UB-Meshachieves 2.04× better cost-efficiency and 7.2% higher availability than Clos networks. These innovations address the critical challenges of building practical, high-performance AI infrastructure at scale.