InfiniBand
InfiniBand is a high-throughput, low-latency networking standard used to interconnect servers and storage inside data centers, most commonly for high-performance computing and AI training clusters. Unlike standard Ethernet, InfiniBand offers native remote direct memory access (RDMA), letting GPUs and CPUs exchange data without involving the host CPU, cutting latency to single-digit microseconds. NVIDIA's Quantum-series InfiniBand switches, inherited from its 2020 acquisition of Mellanox, dominate the market and are the default fabric linking NVLink-connected GPU clusters such as those built around H100, B200, and GB200 systems, though Ethernet-based alternatives — including specifications from the Ultra Ethernet Consortium — are gaining ground as a lower-cost option for large AI fabrics. InfiniBand's per-port cost and power draw are material line items in AI factory build budgets.
Where this comes up
Related terms