interconnect

The transfer looks like gigabytes per request until you divide by prefill time. Then the prompt length cancels, and the real question becomes which models and which GPUs break the link.

Collectives run at the speed of their slowest link, so one NIC that trained at a quarter of its rate slows every GPU in the instance by the same amount. The answer shows the arithmetic that makes a 4x slower link a 2x slower step, then bisects the cluster to the culprit.