NVLink

NVLink is a high-bandwidth, low-latency interconnect technology developed by NVIDIA to enable fast data exchange between GPUs, and between CPUs and GPUs, in supported systems. It provides direct, multi-lane connections that can be combined to create a higher aggregate bandwidth than traditional PCIe, and can support memory-coherent access across devices for certain CUDA workloads.

The links are point-to-point and can be configured in various topologies. Individual NVLink links connect two devices, and multiple links can be grouped on a single device to increase total bandwidth. For larger systems, NVIDIA uses NVSwitch switches to create a fabric that interconnects many GPUs with scalable bandwidth.

NVLink is primarily used in data-center GPUs and accelerators aimed at high-performance computing and artificial intelligence. It enables faster parameter exchange in multi-GPU training and allows software to share device memory more efficiently through CUDA’s Unified Memory and coherent access across GPUs.

Background and evolution: NVLink was introduced in 2016 and has since evolved across NVIDIA’s server-class architectures, including Volta and subsequent generations. It remains a complement to PCIe, offering higher bandwidth for inter-GPU communication and scalable multi-GPU configurations used in DGX and HGX systems.

Limitations: NVLink requires compatible hardware and software support and is not universally enabled on all NVIDIA GPUs. It adds cost and complexity, and workloads that do not require rapid inter-GPU data transfers may see little or no benefit.