A school district DC measures whether 100G or 400G leaf uplinks meet a training job’s bisection bandwidth needs. What is the correct evaluation approach?

Size uplinks from aggregate spine capacity using planned inference serving throughput profiles misses that distributed training places much higher east-west bisection demands than inference serving, and spine totals do not set leaf uplink requirements. Estimate uplink needs from leaf access port totals using generic enterprise oversubscription target confuses generic port math with fabric bandwidth requirements; uplinks must accommodate the actual data rate of GPU-to-GPU collectives. Ignoring bisection math and copying planned inference serving throughput profiles imports the wrong traffic model. Validate uplinks with single-flow TCP throughput and intra-server NVLink transfer measurements fails because a single TCP flow cannot validate multi-path bisection capacity and NVLink measurements stay inside the server and never cross the leaf uplinks.

Exam caveat: AI uplink evaluation is workload-driven bisection sizing (for example 100G versus 400G against training demand)—not leaf-port totals with generic enterprise oversubscription targets, aggregate spine capacity or inference serving throughput profiles, or single-flow TCP and NVLink transfer validation.

Operational check: measure or model the school district training job’s bisection demand, compare 100G and 400G uplink options against that number plus growth, and select the uplink class that keeps GPU collectives off the congestion cliff.