Question

A research team is designing a new AI cluster. They want the eight GPUs inside each server to share activations and gradients over the highest-bandwidth, lowest-latency direct interconnect, while connecting 32 separate servers together into a larger training fabric. Which pairing correctly matches the interconnect to each level of the design?

Accepted Answer

NVLink/NVSwitch for the GPU-to-GPU connections within each server, and InfiniBand for the server-to-server fabric across the 32 nodes

Answer

InfiniBand for the GPU-to-GPU connections within each server, and NVLink for the server-to-server fabric across the 32 nodes

Answer

PCIe for the GPU-to-GPU connections within each server, and NVLink for the server-to-server fabric across the 32 nodes

Answer

Standard 10GbE Ethernet for both the intra-server GPU links and the inter-server fabric