Three topology families, three distinct tradeoffs. The decision reduces to answering four questions about your workload:

1. What is the dominant communication pattern?

Nearest-neighbour stencil computation (CFD, climate, FEA, molecular dynamics) favours torus. The hop count for the majority of communication is 1, and the topology exploits locality perfectly. Random all-to-all AllReduce in data-parallel training favours fat-tree, which provides maximum bisection bandwidth regardless of which nodes are talking.

2. What is the scale, and what is the budget for interconnect hardware?

At petascale (100,000+ nodes), the switch cost of fat-tree becomes enormous. Torus and dragonfly both reduce link count significantly. This was the primary reason torus dominated 2004-2015 scientific HPC. At DGX cluster scale (32-2048 nodes), fat-tree switch cost is trivial relative to GPU cost, and the bisection bandwidth advantage dominates.

3. How important is fault tolerance and redundancy?

Fat-tree with ECMP and adaptive routing provides multiple equal-cost paths between any pair of nodes. A link failure triggers rerouting through alternate paths with no topology change. Torus and dragonfly have limited alternate paths; failures in specific locations can create hotspots or partitions. For always-on AI training clusters where a job may run for weeks, this matters.

4. What is the latency tolerance of the collective operations?

If your training loop runs AllReduce barriers synchronously (the common case with data parallelism), then barrier latency determines step time. Fat-tree typically provides lower all-to-all latency than torus at the same node count, because shorter worst-case hop count. Dragonfly with adaptive routing can approach fat-tree latency for within-group traffic but degrades for cross-group AllReduce.

The verdict for AI training clusters at DGX scale is unambiguous: fat-tree. The bisection bandwidth guarantee, the predictable AllReduce latency, the fault tolerance, and the ECMP load balancing under variable traffic patterns all point in the same direction. The switch cost is affordable at this scale.

The verdict for very-large-scale scientific HPC at tens of thousands of nodes is more nuanced: dragonfly (Frontier) and torus (TPU Pod) are both valid depending on whether the dominant workload is random-permutation collective operations or spatially-local stencil computations.

What you should carry forward: when you read about a new interconnect and someone claims it outperforms fat-tree, the first question is for which traffic pattern. The second question is at what scale. The answer tells you whether the comparison is meaningful for your workload.