This article was accepted into the corpus but its outbound wikilinks were never NER-processed — typical at the deepest BFS hop or when the run's entity cap was reached. No expansion funnel to show.
| NVIDIA NVLink | |
|---|---|
| Name | NVLink |
| Developer | NVIDIA |
| Introduced | 2016 |
| Type | High-speed interconnect |
| Use | GPU-to-GPU and GPU-to-CPU communication |
| Predecessors | PCI Express |
NVIDIA NVLink NVLink is a high-speed interconnect technology developed to link NVIDIA accelerators with low-latency, high-bandwidth links for heterogeneous computing. It complements PCI Express and enables tighter coupling between Tesla-class accelerators, DGX systems, and server CPUs from vendors such as IBM and hyperscalers like Google and Amazon Web Services. Designed for data-center, artificial intelligence, and high-performance computing workloads, NVLink aims to accelerate workloads used by research centers such as Lawrence Berkeley National Laboratory and projects like Human Genome Project-scale analyses.
NVLink provides point-to-point, scalable links between devices such as NVIDIA Tesla V100, NVIDIA A100, and other accelerator modules, enabling coherent memory sharing and higher aggregate bandwidth than traditional interconnects. It was announced alongside products marketed to institutions including Stanford University, MIT, ETH Zurich, Argonne National Laboratory, and corporations like Facebook (now Meta Platforms). NVLink supports system designs in enterprise offerings from Dell Technologies, HPE, Lenovo, and integrated platforms such as IBM POWER9 and later POWER10-based servers.
NVLink employs a mesh-style fabric and uses multiple physical lanes aggregated into high-bandwidth links between nodes. The design enables cache-coherent interconnects supported in collaboration with companies like IBM for CPU/GPU coherence on platforms such as NVIDIA DGX-1 and DGX Station. Electrical and power-delivery considerations align with standards bodies such as JEDEC and manufacturing partners like TSMC and Samsung Electronics. The topology options include point-to-point, ring, and hybrid meshes as seen in systems deployed at institutes such as Los Alamos National Laboratory and Oak Ridge National Laboratory.
NVLink iterations increased per-link bandwidth across generations, scaling from tens to hundreds of gigabytes per second aggregate bandwidth per device. Performance comparisons are evaluated alongside interconnects such as PCI Express 4.0, PCI Express 5.0, InfiniBand, and proprietary links used by companies like Intel and AMD. Benchmarks used by labs such as CERN, Broad Institute, and universities such as UC Berkeley often measure latency, throughput, and scaling for distributed training frameworks pioneered by organizations like OpenAI and DeepMind.
NVLink is implemented in discrete GPUs including products formerly branded under NVIDIA Tesla, NVIDIA Quadro, and newer architectures used in NVIDIA Ampere and NVIDIA Volta families. OEM systems from Supermicro, Inspur, and cloud offerings from Microsoft Azure and Oracle Cloud have incorporated NVLink-enabled hardware. Integration required collaboration with motherboard vendors such as ASUS, Gigabyte Technology, and server integrators used by NASA and biotechnology firms like Illumina.
Software stacks integrate NVLink through frameworks like CUDA, libraries such as cuDNN, NCCL, and middleware used by projects including TensorFlow, PyTorch, MXNet, and high-performance libraries maintained by institutions like Argonne National Laboratory and Sandia National Laboratories. Operating systems and drivers are provided by Red Hat, Ubuntu, and enterprise partners such as SUSE. NVLink exposes primitives for peer-to-peer access, unified memory, and remote direct memory access used in HPC codes developed by groups at Los Alamos National Laboratory and companies like Siemens.
Common use cases include distributed deep learning at organizations such as OpenAI, Google DeepMind, and research at Carnegie Mellon University; simulation workloads in projects at Lawrence Livermore National Laboratory and Princeton University; and real-time analytics in fintech firms such as Goldman Sachs and JPMorgan Chase. NVLink-enabled systems accelerate scientific visualization used at NASA Ames Research Center, computational chemistry modeled by teams at Pfizer and Merck, and weather modeling conducted by agencies like NOAA.
NVLink is often compared with interconnects such as InfiniBand (including products from Mellanox Technologies), PCI Express generations, and emerging fabric technologies from Intel (e.g., CXL) and AMD-related initiatives. Trade-offs involve topology flexibility, latency, coherence models, and vendor ecosystem support involving partnerships with HPE, Dell EMC, Lenovo, and cloud providers such as Amazon Web Services and Google Cloud Platform. Adoption decisions are influenced by procurement bodies like DARPA and collaborations between academia and industry exemplified by consortia including OpenAI collaborators and large-scale HPC centers.
Category:NVIDIA Category:Computer buses