Dragonfly Topology
Overview
Dragonfly topology is a network interconnect design that organizes compute nodes into groups (pods), where each group connects to all others via global links while also providing high-bandwidth local links within the group. The resulting structure resembles a circular multi-connected fabric (hence the "dragonfly" name). Compared to Clos (leaf-spine) topology, Dragonfly:
- Scales horizontally more efficiently: Adding groups expands the cluster without multiplying the spine layer's port requirements.
- Reduces diameter: Fewer hops between arbitrary node pairs compared to fat-tree at equivalent scale.
- Increases cabling complexity: The all-to-all inter-group links require more complex physical cabling than the structured Clos hierarchy — a significant operational challenge for ODMs and data center operators.
The Dragonfly topology is primarily used in high-performance computing (HPC) supercomputers today (e.g., Slingshot interconnect in Cray/HPE systems). For AI data centers, it is projected to become necessary for large-scale deployments starting around 2027, when cluster sizes exceed the efficient operating range of Clos without oversubscription. The roadmap projects Dragonfly eventually coexisting with quantum networking by 2029–2030.
Sign in to read the full article.
Sign In