You're designing a multi-GPU system for AI training using NVIDIA GPUs with NVLink connections.

The correct answer isPXN (Peer eXchange Network).

From theNVIDIA NCCL Documentation:

"PXN enables communication between GPUs connected via NVLink and NICs by treating the GPUs as a distributed switch. This architecture improves bandwidth utilization by enabling any GPU to communicate with the NIC via the shortest path available, even if it's not directly connected to the NIC."

This enhances GPU-to-NIC and NIC-to-GPU transfers, leveraging the NVLink topology. It significantly boosts performance in multi-GPU setups where not every GPU is directly connected to the NIC.

Other options:

Adaptive Routingis a fabric-level feature for dynamic path rerouting.

Graph Search Optimizationis used internally for topology modeling in NCCL.

SHARP v2is a switch-based collective acceleration method, unrelated to PXN.