NVLink is a wire-based serial multi-lane near-range communications link developed by Nvidia. It is a high-speed connection for GPUs and CPUs that enables fast data transfer and improved scalability for multi-GPU and multi-node systems.
NVLink uses a proprietary high-speed signaling interconnect (NVHS) and a robust software protocol to provide low latency and high bandwidth communication between processors.
NVLink can be used for various purposes, such as:
- Accelerating AI and high-performance computing (HPC) workloads by enabling faster and larger model training, inference, and data processing.
- Enhancing GPU memory capacity and bandwidth by enabling GPU memory pooling and coherent access.
- Improving system performance and efficiency by enabling in-network computing engines, such as NVIDIA Scalable Hierarchical Aggregation and Reduction Protocol (SHARP)™, for collective operations and multicast acceleration.
NVLink has evolved through several generations, each offering higher data rates and improved features. The latest generation, NVLink 4.0, provides 100 Gbit/s per lane and supports up to 18 NVLink connections per GPU, resulting in a total bandwidth of 900 GB/s per GPU.
NVLink can also be combined with NVIDIA NVSwitch ™, which provides all-to-all GPU communication at full NVLink speed within a single node and between nodes.
