Clos Topology

Overview

Clos topology (commonly called leaf-spine in data center contexts) is a hierarchical network topology that provides non-blocking connectivity at scale by connecting leaf switches (ToR/access layer) to a full mesh of spine switches. It is the dominant topology for current data center networking, including AI cluster deployments.

Key properties:

• ECMP load balancing: Multiple equal-cost paths between any two hosts distribute traffic across all spine switches.
• Radix limitation: Increasing the number of connected hosts requires either increasing switch port count (radix) or adding hierarchical layers (fat-tree). Adding layers increases complexity, hop count, and latency.
• Predictable architecture: Well-understood cabling, management tooling, and failure modes make Clos the default choice for operators building AI infrastructure today.

For large-scale AI clusters, the Clos topology's radix limitation becomes a constraint as cluster sizes grow into tens of thousands of GPUs, requiring either oversubscription or addition of aggregation layers. The Dragonfly topology is projected to emerge as an alternative for very large AI deployments around 2027.