AI Networking & Fabrics Guide

How GPU servers talk to each other: InfiniBand versus RoCE Ethernet, ConnectX adapter part numbers, PCIe bottlenecks, HDR100 and NDR splits, and the lossless-Ethernet settings that decide tail latency. Pair it with the interconnect matrices for Hopper, Ampere and Volta.

Start here

Adapters and fabric switches

Reference tables & compatibility

Every AI Networking & Fabrics article