Glossary
NVLink
NVLink is a high-bandwidth connection for fast communication between processors and AI accelerators. It helps large models share data across multiple chips more efficiently than general-purpose peripheral links alone.
· 4 min read
← Glossary · Articles · Pillar guide
Definition
NVLink is a high-speed interconnect technology from NVIDIA for moving data between processors and accelerators. In AI systems it commonly connects multiple accelerators inside a server or tightly coupled system so they can exchange data at high bandwidth.
Large AI workloads often cannot fit on one chip. A model may be split across accelerators, or a batch of work may be distributed across them. NVLink helps those accelerators communicate faster.
Why it matters
AI performance is not only chip speed. Many workloads spend time moving data between memory and accelerators. If communication is slow, expensive hardware waits. For large models, interconnect bandwidth and latency can shape what fits efficiently and how quickly outputs are produced.
How it works
NVLink provides dedicated high-speed pathways between supported NVIDIA components. Accelerators may connect directly, through NVSwitch, or to a processor in the same fabric. Software frameworks use that connectivity for model parallelism and collective operations.
NVLink usually operates inside a server or closely integrated system. Across racks, InfiniBand or high-performance Ethernet commonly handle scale-out communication.
Common misconceptions
NVLink is not the whole data center network. It is a local high-speed interconnect.
It does not automatically make every workload faster. Benefits appear when workloads are limited by accelerator-to-accelerator communication.
Related reading
See NVSwitch, InfiniBand, AI Inference, and The Latency Tax.