Question

A research organization is building an AI cluster of 32 DGX H100 nodes for large-scale distributed training. The network architect must design the compute fabric so that inter-node GPU-to-GPU collective operations (like all-reduce during gradient synchronization) are not the bottleneck. Which fabric design best meets this requirement?

Accepted Answer

A dedicated non-blocking InfiniBand compute fabric with a rail-optimized topology, providing a separate high-speed link per GPU across nodes

Answer

A single shared 25 GbE Ethernet network carrying compute, storage, and management traffic together

Answer

PCIe cabling between the nodes to extend the intra-node GPU bus across the cluster

Answer

Relying solely on NVLink and NVSwitch to connect all 32 nodes into one memory domain