A city data center must support nonblocking east-west bandwidth for all-reduce traffic during distributed training. What network-deployment evaluation priority should architects emphasize?

A city data center that must support nonblocking east-west bandwidth for all-reduce traffic during distributed training should emphasize sufficient bisection bandwidth so collective training traffic is not oversubscribed across the fabric. All-reduce and related collectives generate heavy GPU-to-GPU east-west flows; if the fabric's bisection capacity is oversubscribed like a classic campus design, iteration time collapses and municipal training SLAs fail. Enlarging east-west MTU sizes so collective training throughput scales within current fabric links improves per-packet efficiency but does not create additional bisection bandwidth, so all-reduce flows on oversubscribed links still congest and iteration time collapses. Tuning RoCEv2 congestion thresholds so east-west collective traffic paces smoothly across existing links smooths queue behavior on existing links but leaves bisection capacity unchanged, so all-reduce traffic still congests and iteration time collapses. Optimizing equal-cost multipath hashing so east-west flows balance evenly across available fabric paths improves utilization of available paths but does not provision the nonblocking bisection capacity required for GPU collectives. Exam caveat: AI network-deployment evaluation prioritizes east-west bisection and nonblocking behavior for distributed training collectives—not east-west MTU enlargement within current links, RoCE congestion threshold tuning across existing links, or ECMP balancing across available paths. Operational check: size and measure fabric bisection against the city's all-reduce profile, verify oversubscription ratios are AI-appropriate, and load-test a representative training job before calling the municipal GPU fabric production-ready.