Graphics Processing Unit
GPUBack to Glossary
ai-infrastructure
NVLink
NVIDIA's proprietary high-bandwidth interconnect for GPU-to-GPU and GPU-to-CPU communication within a server or rack-scale system.
Unit: TB/s (bandwidth)
Full Definition
NVLink is a high-speed electrical interconnect developed by NVIDIA to replace PCIe for GPU-to-GPU and GPU-to-CPU data transfers within a node. NVLink 4.0 (used in H100 systems) provides 900 GB/s of bidirectional bandwidth per GPU — approximately 7× the bandwidth of PCIe 5.0 ×16.
In the GB200 NVL72 architecture, NVLink is extended via NVLink Switch chips to create a rack-scale fabric connecting 72 Blackwell GPUs with a combined 130 TB/s of all-to-all bandwidth. This enables the rack to function as a single large GPU for model parallelism without inter-node InfiniBand hops for intra-rack communication. NVLink is a key differentiator of NVIDIA AI compute versus commodity GPU server designs.
Also Known As
NVLink SwitchNVSwitch
Source Reference
NVIDIA NVLink and NVSwitch Technical Overview; NVIDIA GB200 NVL72 Architecture Whitepaper