In a 512-GPU training cluster, engineers observe poor scaling efficiency when moving from single-node to multi-node runs, even though the compute nodes themselv...