Glossary
InfiniBand
InfiniBand is a specialized networking technology used in high-performance computing and AI clusters. It is designed for fast, predictable communication between servers.
· 4 min read
← Glossary · Articles · Pillar guide
Definition
InfiniBand is a high-performance networking technology used to connect servers in demanding compute clusters. In AI infrastructure it is often the scale-out network that lets many servers work together on training, inference, or data-processing tasks.
Why it matters
Distributed AI workloads are often limited by communication. A slow network leaves accelerators idle. For expensive systems, idle time is costly. InfiniBand is designed for high throughput, low latency, and predictable behavior under heavy load.
How it works
InfiniBand uses specialized adapters, switches, cables, and software. It commonly supports remote direct memory access (RDMA), which lets machines exchange data with low CPU overhead. Topology, oversubscription, and routing still determine real performance.
Common misconceptions
InfiniBand is not simply faster Ethernet. It is a separate technology with its own hardware and operational model. Ethernet can also serve AI clusters, but the design tradeoffs differ.
The fastest network does not solve scaling alone. Software efficiency and local accelerator interconnects still matter.
Related reading
See The Latency Tax, NVLink, and Metro AI Infrastructure.