The single most useful thing to fix in your head before drawing anything is that a GPU cluster is not one network with different traffic on it. It is four networks with different physics, different failure consequences, and deliberately different blast radii. Collapsing any two of them to save switches is the decision that gets revisited during the first outage.
The east-west compute fabric carries GPU-to-GPU collective traffic and nothing else. It is bandwidth-hungry, bursty, synchronous, and intolerant of head-of-line blocking. Every design choice here — rail-optimised wiring, non-blocking bisection, adaptive routing, lossless configuration — exists to serve one traffic pattern. Nothing that is not a collective belongs on it.
The frontend, or north-south, fabric carries everything else the nodes need: provisioning and PXE, tenant virtual networks, control plane for the scheduler and orchestrator, telemetry, and — in a converged design — storage. It is a conventional leaf-spine Ethernet fabric with an EVPN-VXLAN overlay, and it is where multi-tenancy actually lives.
The storage fabric is either its own tier or a set of virtual networks on the frontend. Its traffic is throughput-sensitive but far more tolerant of jitter than collectives are. Converging it onto the frontend is a legitimate and common choice at small and mid scale; separating it is the right call when checkpoint bursts start to interfere with control-plane traffic.
The out-of-band management fabric carries BMC, console, switch management, and controller traffic. It is the lowest bandwidth and the highest consequence: it is how you recover everything else. It must be able to function when the production fabrics do not.
A worked reference runs through this chapter: 72 HGX B300 nodes, 576 GPUs, single site, single availability zone. Each node presents eight 800 Gb/s ConnectX-8 SuperNIC ports to the compute fabric and a dual-port 200 GbE adapter to the frontend. That shape — eight rails out the back for collectives, one bonded pair for everything else — is the standard Blackwell Ultra attachment and it drives every number that follows.