A team is planning an InfiniBand fabric for a large AI/HPC cluster and wants to reduce the volume of data traversing the network during distributed training's collective communication phases. Which InfiniBand capability directly addresses this by offloading those operations to the switch network?