The single leaf switch you have been diagnosing for six chapters now has a place in a much larger picture.
You know what bisection bandwidth is and why full bisection is required for AllReduce at line rate. You know the k-ary fat-tree formula -- how switch radix and stage count determine maximum cluster size and whether the fabric can support non-blocking all-to-all communication.
You know the BasePOD (32 nodes, 2 stages, 8 leaf + 32 spine) and SuperPOD (256 nodes, 3 stages) as concrete reference designs with specific port allocations, switch counts, and hop distances. You know how the extra stages in a SuperPOD double the AllReduce hop count and why that matters for JCT.
You know oversubscription as a calculable number -- downlink bandwidth divided by uplink bandwidth -- and why a 2:1 oversubscription does not give you 50% throughput in an AI cluster: it gives you PFC storms on every AllReduce barrier because statistical multiplexing does not apply when every GPU sends simultaneously.
You know ROD and RUD as wiring philosophies -- ROD for pure training (clean rail diagnostics, one-hop AllReduce), RUD for mixed workloads (higher per-node bandwidth, less diagnostic clarity). You know the buffer depth trade-off: deep buffers absorb AllReduce bursts at the cost of base latency; shallow buffers require perfect load balancing or they overflow before PFC can respond.
And you know the cabling constraints -- connector types, reach limitations, and the critical importance of correct rail assignment -- because one wrong cable can create a topology asymmetry that no diagnostic counter will directly reveal.