The primary goal is to reduce latency and maximize bandwidth. Minimizing hops between frequently communicating GPUs is essential for achieving this. Server placement for cooling is a secondary concern, and topology choice should prioritize performance over simplicity. Subnet configuration and server placement relative to the network edge are less critical for NVLink-specific performance.