A research team is deploying a cluster to train a trillion-parameter model that requires very high all-to-all GPU bandwidth across 32 GPUs spanning four DGX H100 nodes. They want these 32 GPUs to behave as a single large scale-up domain with NVLink-class bandwidth between GPUs in different chassis, rather than relying on the node-internal NVSwitch fabric only. Which technology enables extending the NVLink domain across multiple physical nodes?