InfiniBand & AI switches are high-speed interconnect devices used to build the low-latency, high-bandwidth fabrics that GPU clusters, HPC systems, and AI training infrastructure depend on. This category covers InfiniBand switches across HDR (200 Gb/s) and NDR (400 Gb/s) generations from vendors such as NVIDIA Mellanox, as well as purpose-built Ethernet AI switches optimized for RDMA over Converged Ethernet (RoCE) workloads. Port counts range from compact top-of-rack units to large-scale director-class chassis supporting hundreds of ports in a single hop.

Key buying considerations include fabric generation (HDR vs. NDR vs. emerging XDR), port density, radix (the number of ports per switch directly affects fat-tree or dragonfly topology scale-out), and compatibility with your installed HCAs and cables. These switches are commonly deployed in GPU SuperPOD configurations, Beowulf-style HPC clusters, distributed deep-learning training environments, and high-frequency financial computing fabrics where microsecond latency and MPI all-reduce throughput are critical constraints.