A wastewater plant’s AI uplink cannot tolerate lossy drops that stall RDMA training iterations. What network property should evaluation require?

A wastewater plant’s AI uplink that cannot tolerate lossy drops that stall RDMA training iterations should require lossless or tightly controlled loss behavior appropriate for RDMA/AI training traffic. RDMA-based training is sensitive to loss: drops trigger expensive recovery and stall iterations, so evaluation must demand congestion management and loss behavior suited to GPU flows rather than relying on sustained high-throughput capacity sized for large checkpoint transfers on the plant fabric. Sustained high-throughput capacity sized for large checkpoint transfers will not prevent the stalls the plant cannot accept if loss is uncontrolled. Consistently low round-trip latency measured across GPU fabric uplink paths optimizes the wrong metric when the real requirement is controlled loss for RDMA. Detailed telemetry and flow monitoring visibility across AI fabric links for operations is helpful for operations but does not itself enforce lossless behavior and is not an evaluation requirement for production RDMA training. Exam caveat: Lossless or carefully controlled loss for RDMA/AI training is a network property under evaluation—not raw throughput, latency targets, or telemetry by themselves. Operational check: validate PFC/ECN or equivalent controls on the wastewater AI uplink, measure loss under training load, and refuse cutover if sustained high-throughput capacity for checkpoint transfers is present but GPU flows still see uncontrolled loss that stalls iterations.