Two communication paths
Scale-up communication connects devices within a system; scale-out communication joins systems across a fabric. Bandwidth, latency, contention and topology affect distributed work differently. A port-speed label alone cannot describe a multi-node job.
Storage is in the critical path
Dataset ingestion, checkpoint writes, artifact reads and model loading can expose different storage limits. Capacity in terabytes does not establish aggregate or per-job throughput. Understand cache behaviour, durability, sharing and recovery.
A useful pilot
Measure the target job with the intended network and storage configuration. Record collective communication time, data stalls and checkpoint duration. Repeat under realistic concurrency, and ask who owns diagnosis when a problem spans host, fabric and data layers.