Every large-scale supercomputer built before 2015 connected its nodes in a torus. IBM's Blue Gene machines, Cray's XT and XE lines, Fujitsu's K computer -- the topology that defined a generation of scientific computing was not fat-tree. It was a grid of nodes, each wired to its nearest neighbours in multiple dimensions, no switches required.
Today's AI training clusters overwhelmingly use fat-tree. The torus didn't lose because the engineers got bored. It lost because the workload changed -- and understanding why it lost is the clearest possible explanation of what fat-tree actually provides, and what it costs you when you choose something different.
This chapter covers the three topology families that matter for context: torus (and its descendants), dragonfly, and Google's TPU Pod variant. Along the way you'll see how topology choice is always a bet on traffic pattern -- and why the all-to-all synchronisation barrier at the heart of transformer training makes that bet decisive.