Glossary

NVSwitch

NVSwitch extends NVLink by creating a high-bandwidth switching fabric among multiple accelerators so large AI systems can communicate as a tightly connected group.

· 4 min read

← Glossary · Articles · Pillar guide

Definition

NVSwitch connects multiple accelerators through an NVLink fabric. Where NVLink provides high-speed links, NVSwitch helps many accelerators communicate through a switched topology rather than only point-to-point connections.

Why it matters

Large AI models often need more memory and compute than one accelerator can provide. Splitting work creates a communication problem. A switched fabric can improve aggregate communication and utilization for training and for inference of partitioned models.

How it works

NVSwitch devices sit between accelerators and route traffic across an NVLink fabric. Software uses this topology for collective operations and model partitioning. NVSwitch does not replace the cluster network. Scale-out networking still connects servers and racks.

Common misconceptions

NVSwitch and NVLink are not interchangeable. NVLink is the link technology. NVSwitch is the switch fabric that uses those links.

It does not eliminate all bottlenecks. Memory bandwidth, software scheduling, storage, and scale-out networks still matter.

See NVLink, InfiniBand, and AI Inference.