Correct for inference - the main use of the interconnect is RDMA requests between GPUs to fit models that wouldn't otherwise fit. Not really correct for traini

Correct for inference - the main use of the interconnect is RDMA requests between GPUs to fit models that wouldn't otherwise fit. Not really correct for training - training has a lot of all-to-all problems, so hierarchical reduction is useful but doesn't really solve the incast problem - Nvlink _bandwidth_ is less of an issue than perhaps the SHARP functions in the NVLink switch ASICs.