Skip to main content
Back to Glossary
ai-infrastructure

InfiniBand

A high-speed, low-latency network fabric used in AI supercomputer clusters to connect GPUs across nodes, enabling RDMA at speeds up to 400 Gb/s per port.

Unit: Gb/s per port

Full Definition

InfiniBand is a high-performance interconnect technology standardised by the InfiniBand Trade Association (IBTA). It provides sub-microsecond latency and high throughput via Remote Direct Memory Access (RDMA), allowing GPU memory to be read and written across nodes without CPU involvement — critical for distributed AI training.

Current generations include HDR (200 Gb/s) and NDR (400 Gb/s) InfiniBand. NVIDIA's Quantum-2 NDR switches are the backbone of the world's largest AI supercomputers, including NVIDIA DGX SuperPOD, Meta's AI Research SuperCluster, and national HPC facilities. At rack scale, NVLink handles intra-rack GPU communication; InfiniBand carries inter-rack traffic.

Also Known As

IBHDR InfiniBandNDR InfiniBand

Source Reference

InfiniBand Trade Association (IBTA) specification; NVIDIA Quantum-2 NDR InfiniBand Switch Datasheet

Related Terms