Question

An AI infrastructure team is designing a cluster of eight DGX H100 systems for large-model training. Within each DGX node, all eight GPUs must exchange gradients at maximum bandwidth during every training step, and the eight nodes must also communicate for multi-node all-reduce operations. Which interconnect strategy correctly matches each communication requirement?

Accepted Answer

Use NVLink/NVSwitch for GPU-to-GPU communication inside each node, and InfiniBand for communication between the eight nodes.

Answer

Use InfiniBand for GPU-to-GPU communication inside each node, and NVLink for communication between the eight nodes.

Answer

Use standard PCIe for intra-node GPU communication and 10GbE Ethernet for inter-node communication to reduce cost.

Answer

Use NVLink for both intra-node and inter-node communication by cabling NVLink directly between separate DGX chassis.