NVLink

NVLink, developed by Nvidia, is a high-speed, direct GPU-to-GPU interconnect that enhances multi-GPU communication within servers. NVLink represents a significant advancement in GPU interconnect technology, offering high bandwidth and low latency for data-intensive applications. Its development reflects Nvidia’s commitment to meeting the growing demands of advanced computing, including AI, deep learning, and HPC[1][3][6][8].

It’s designed to facilitate faster and more efficient data transfer between GPUs, and between GPUs and CPUs, compared to traditional technologies like PCIe. NVLink achieves this by providing a significantly higher bandwidth, up to 80 gigabytes per second or more, which is at least five times that of PCIe connections[1][3]. This technology is crucial for high-performance computing (HPC), deep learning, and large-scale AI applications, where rapid data exchange between multiple GPUs is essential for processing large datasets and complex computations.

Introduced with Nvidia’s Pascal GPU generation in 2016, NVLink has evolved alongside Nvidia’s GPU architectures, playing a pivotal role in the Summit and Sierra pre-exascale supercomputers commissioned by the U.S. Department of Energy. These systems leverage NVLink for efficient CPU-to-GPU and GPU-to-GPU communications, enabling considerable application speedups compared to systems using PCIe interconnects[3].

NVLink’s architecture allows for multiple connections between devices, using a mesh networking approach rather than a central hub. This design reduces bottlenecks and improves data throughput, with recent iterations offering bandwidths up to 900 GB/s. NVLink supports a mesh topology, enabling versatile and numerous connections between GPUs for more efficient parallel processing[8][11].

Despite its advantages, NVLink requires compatible hardware and specific physical bridges for connection, which can add to the cost and complexity of system setup. It’s primarily used in professional and data center GPUs, like the Nvidia Quadro and Tesla series, and is less common in consumer-grade GPUs[2][5].

Citations:

[1] https://www.nvidia.com/en-us/data-center/nvlink/

[2] https://www.nvidia.com/en-us/design-visualization/nvlink-bridges/

[3] https://info.nvidianews.com/rs/nvidia/images/NVIDIA%20NVLink%20High-Speed%20Interconnect%20Application%20Performance%20Brief.pdf

[4] https://www.servethehome.com/dual-nvidia-geforce-rtx-3090-nvlink-performance-review-asus-zotac/

[5] https://www.pugetsystems.com/labs/articles/nvidia-nvlink-2021-update-and-compatibility-chart-2074/

[6] https://blogs.nvidia.com/blog/what-is-nvidia-nvlink/

[7] https://www.youtube.com/watch?v=Z0IWrcmJvYQ

[8] https://www.storagereview.com/review/from-sli-to-nvlink-the-evolution-of-gaming-and-multi-gpu-technologies-and-their-influence-on-ai

[9] https://www.reddit.com/r/LocalLLaMA/comments/13t09kj/how_much_performance_increase_does_using_nvlink/

[10] https://www.reddit.com/r/HPC/comments/qiewx2/to_nvlink_or_to_not_nvlink_that_is_the_question/

[11] https://en.wikipedia.org/wiki/NVLink

[12] https://www.electronicshub.org/nvlink-vs-sli/

[13] https://www.pny.com/professional/software-solutions/about-nvidia-gpus/nvlink

[14] https://huggingface.co/transformers/v4.9.2/performance.html

[15] https://www.naddod.com/blog/nvlink-infiniband-and-roce-in-ai-gpu-interconnect-technologies

[16] https://www.fibermall.com/blog/what-is-nvidia-nvlink.htm

[17] https://support.chaos.com/hc/en-us/articles/4406882847505-NVLink-FAQ

[18] https://forums.tomshardware.com/threads/is-nvlink-actually-useful-for-anything-and-can-it-be-used-with-two-different-model-gpus.3804090/

[19] https://www.fabricatedknowledge.com/p/nvlinkswitch-and-platform-wars-micron

[20] https://community.fs.com/article/an-overview-of-nvidia-nvlink.html

[21] https://www.naddod.com/blog/unveiling-the-evolution-of-nvlink

[22] https://en.wikichip.org/wiki/nvidia/nvlink

[23] https://www.exxactcorp.com/blog/Components/what-is-nvlink-and-how-does-it-work