×

注意!页面内容来自https://www.nvidia.cn/data-center/nvlink/,本站不储存任何内容,为了更好的阅读体验进行在线解析,若有广告出现,请及时反馈。若您觉得侵犯了您的利益,请通知我们进行删除,然后访问 原网页

 NVLink Switch Chip

NVIDIA NVLink 和 NVLink 交换机

纵向扩展网络实现 GPU 到 GPU 的高带宽通信,助力 AI 训练、推理和其他要求严苛的机架级 GPU 加速工作负载。

<> #container-b73b8a29aa { } /* Mobile: up to 639px */ @media only screen and (max-width: 639px) { #container-b73b8a29aa { } } /* Tablet: 640px to 1023px */ @media only screen and (min-width: 640px) and (max-width: 1023px) { #container-b73b8a29aa { } } /* Laptop: 1024px to 1349px */ @media only screen and (min-width: 1024px) and (max-width: 1349px) { #container-b73b8a29aa { } } /* Desktop: 1350px and up */ @media only screen and (min-width: 1350px) { #container-b73b8a29aa { } }
<> @media screen and (max-width: 639px){ #nv-separator-3510aa90ac{ height:calc(56.25vw); } }
<> @media screen and (max-width: 639px){ #nv-separator-3bfb7cf703{ height:30px; } }
<> @media (max-width: 639px){ #bm-uf{height:auto!important;} #bm-uf>.nv-video-as-bg{height:auto!important;} #bm-uf>.nv-video-as-bg>video{bottom: 0;width: 100%;} #bm-uf>.nv-img-as-bg>picture>img{bottom: 0;width: 100%;} }
<> #bm-uf { } /* Mobile: up to 639px */ @media only screen and (max-width: 639px) { #bm-uf { } } /* Tablet: 640px to 1023px */ @media only screen and (min-width: 640px) and (max-width: 1023px) { #bm-uf { background-image: url('/content/dam/en-zz/Solutions/Data-Center/nvlink/[email protected]'); background-image: -webkit-image-set( url('/content/dam/en-zz/Solutions/Data-Center/nvlink/nvidia-nvlink-switch-bm-md420-offset-t.jpg') 1xurl('/content/dam/en-zz/Solutions/Data-Center/nvlink/[email protected]') 2x ); background-image: image-set( url('/content/dam/en-zz/Solutions/Data-Center/nvlink/nvidia-nvlink-switch-bm-md420-offset-t.jpg') 1xurl('/content/dam/en-zz/Solutions/Data-Center/nvlink/[email protected]') 2x ); } } /* Laptop: 1024px to 1349px */ @media only screen and (min-width: 1024px) and (max-width: 1349px) { #bm-uf { background-image: url('/content/dam/en-zz/Solutions/Data-Center/nvlink/[email protected]'); background-image: -webkit-image-set( url('/content/dam/en-zz/Solutions/Data-Center/nvlink/nvidia-nvlink-switch-bm-md460-offset-left-l.jpg') 1xurl('/content/dam/en-zz/Solutions/Data-Center/nvlink/[email protected]') 2x ); background-image: image-set( url('/content/dam/en-zz/Solutions/Data-Center/nvlink/nvidia-nvlink-switch-bm-md460-offset-left-l.jpg') 1xurl('/content/dam/en-zz/Solutions/Data-Center/nvlink/[email protected]') 2x ); } } /* Desktop: 1350px and up */ @media only screen and (min-width: 1350px) { #bm-uf { background-image: url('/content/dam/en-zz/Solutions/Data-Center/nvlink/[email protected]'); background-image: -webkit-image-set( url('/content/dam/en-zz/Solutions/Data-Center/nvlink/nvidia-nvlink-switch-bm-md460-offset-left-d.jpg') 1xurl('/content/dam/en-zz/Solutions/Data-Center/nvlink/[email protected]') 2x ); background-image: image-set( url('/content/dam/en-zz/Solutions/Data-Center/nvlink/nvidia-nvlink-switch-bm-md460-offset-left-d.jpg') 1xurl('/content/dam/en-zz/Solutions/Data-Center/nvlink/[email protected]') 2x ); } }

速度更快的纵向扩展互连

最新的 AI 模型追求最卓越的性能,需要在整个服务器机架内实现无缝、高吞吐量的 GPU 到 GPU 通信。第六代 NVIDIA NVLink™ 和 NVLink Switch 支持超低延迟、海量网络带宽和全连接,专为加速训练和推理而设计,显著提高推理和代理式 AI 工作负载的性能。

<> #introduction { } /* Mobile: up to 639px */ @media only screen and (max-width: 639px) { #introduction { } } /* Tablet: 640px to 1023px */ @media only screen and (min-width: 640px) and (max-width: 1023px) { #introduction { } } /* Laptop: 1024px to 1349px */ @media only screen and (min-width: 1024px) and (max-width: 1349px) { #introduction { } } /* Desktop: 1350px and up */ @media only screen and (min-width: 1350px) { #introduction { } }
<> #nvlink { background-color:#f7f7f7; } /* Mobile: up to 639px */ @media only screen and (max-width: 639px) { #nvlink { } } /* Tablet: 640px to 1023px */ @media only screen and (min-width: 640px) and (max-width: 1023px) { #nvlink { } } /* Laptop: 1024px to 1349px */ @media only screen and (min-width: 1024px) and (max-width: 1349px) { #nvlink { } } /* Desktop: 1350px and up */ @media only screen and (min-width: 1350px) { #nvlink { } }

NVLink 大规模性能

与上一代产品相比,NVIDIA Rubin 中的第六代 NVIDIA NVLink 可将 GPU 到 GPU 的通信带宽提高 2 倍,从而助力最新的 AI 模型架构加快训练和推理速度。

<> #container-fed58c343e { background-color:#f7f7f7; } /* Mobile: up to 639px */ @media only screen and (max-width: 639px) { #container-fed58c343e { } } /* Tablet: 640px to 1023px */ @media only screen and (min-width: 640px) and (max-width: 1023px) { #container-fed58c343e { } } /* Laptop: 1024px to 1349px */ @media only screen and (min-width: 1024px) and (max-width: 1349px) { #container-fed58c343e { } } /* Desktop: 1350px and up */ @media only screen and (min-width: 1350px) { #container-fed58c343e { } }

借助 NVLink 通信提高逻辑推理吞吐量

<> #nvswitch { background-color:#ffffff; } /* Mobile: up to 639px */ @media only screen and (max-width: 639px) { #nvswitch { } } /* Tablet: 640px to 1023px */ @media only screen and (min-width: 640px) and (max-width: 1023px) { #nvswitch { } } /* Laptop: 1024px to 1349px */ @media only screen and (min-width: 1024px) and (max-width: 1349px) { #nvswitch { } } /* Desktop: 1350px and up */ @media only screen and (min-width: 1350px) { #nvswitch { } }
<> #container-c20866f4a9 { } /* Mobile: up to 639px */ @media only screen and (max-width: 639px) { #container-c20866f4a9 { } } /* Tablet: 640px to 1023px */ @media only screen and (min-width: 640px) and (max-width: 1023px) { #container-c20866f4a9 { } } /* Laptop: 1024px to 1349px */ @media only screen and (min-width: 1024px) and (max-width: 1349px) { #container-c20866f4a9 { } } /* Desktop: 1350px and up */ @media only screen and (min-width: 1350px) { #container-c20866f4a9 { } }

使用 NVIDIA NVLink 和 NVLink 交换机完全连接 GPU

NVLink 是一种 3.6 TB/s 的双向 GPU 到 GPU 直接互连技术,可扩展服务器内的多 GPU 输入和输出 (IO) 。NVIDIA NVLink Switch 芯片可连接多个 NVLink,在整个机架内基于 NVLink 实现全线速 all-to-all GPU 通信。

为了支持高速集合运算,每个 NVLink 交换机都配有 NVIDIA SHARP™ 技术引擎,可用于网络内归约和组播加速。

<> #container-3bc9bed5ec { } /* Mobile: up to 639px */ @media only screen and (max-width: 639px) { #container-3bc9bed5ec { } } /* Tablet: 640px to 1023px */ @media only screen and (min-width: 640px) and (max-width: 1023px) { #container-3bc9bed5ec { } } /* Laptop: 1024px to 1349px */ @media only screen and (min-width: 1024px) and (max-width: 1349px) { #container-3bc9bed5ec { } } /* Desktop: 1350px and up */ @media only screen and (min-width: 1350px) { #container-3bc9bed5ec { } }
<> #container-279f1c8d79 { background-color:#ffffff; } /* Mobile: up to 639px */ @media only screen and (max-width: 639px) { #container-279f1c8d79 { } } /* Tablet: 640px to 1023px */ @media only screen and (min-width: 640px) and (max-width: 1023px) { #container-279f1c8d79 { } } /* Laptop: 1024px to 1349px */ @media only screen and (min-width: 1024px) and (max-width: 1349px) { #container-279f1c8d79 { } } /* Desktop: 1350px and up */ @media only screen and (min-width: 1350px) { #container-279f1c8d79 { } }
<> #container-97f1d66e9e { } /* Mobile: up to 639px */ @media only screen and (max-width: 639px) { #container-97f1d66e9e { } } /* Tablet: 640px to 1023px */ @media only screen and (min-width: 640px) and (max-width: 1023px) { #container-97f1d66e9e { } } /* Laptop: 1024px to 1349px */ @media only screen and (min-width: 1024px) and (max-width: 1349px) { #container-97f1d66e9e { } } /* Desktop: 1350px and up */ @media only screen and (min-width: 1350px) { #container-97f1d66e9e { } }

借助 NVLink 交换机系统加速万亿参数模型的测试时逻辑推理

借助 NVLink Switch,NVLink 连接可以跨节点扩展,创建无缝、高带宽、多节点的 GPU 集群,从而有效地形成数据中心大小的 GPU。NVIDIA NVLink 交换机可在一个 NVIDIA Vera Rubin NVL72 系统中实现 260 TB/s 的 GPU 聚合带宽,从而加速大型模型并行计算。采用 NVLink 的多服务器集群可在计算能力显著增加的情况下平衡 GPU 通信,因此与单个 8 GPU 系统相比,NVIDIA Vera Rubin NVL72 可支持 9 倍的 GPU 数量。

<> #container-d3dd385f95 { } /* Mobile: up to 639px */ @media only screen and (max-width: 639px) { #container-d3dd385f95 { } } /* Tablet: 640px to 1023px */ @media only screen and (min-width: 640px) and (max-width: 1023px) { #container-d3dd385f95 { } } /* Laptop: 1024px to 1349px */ @media only screen and (min-width: 1024px) and (max-width: 1349px) { #container-d3dd385f95 { } } /* Desktop: 1350px and up */ @media only screen and (min-width: 1350px) { #container-d3dd385f95 { } }
<> #container-7b1ec0aa84 { } /* Mobile: up to 639px */ @media only screen and (max-width: 639px) { #container-7b1ec0aa84 { } } /* Tablet: 640px to 1023px */ @media only screen and (min-width: 640px) and (max-width: 1023px) { #container-7b1ec0aa84 { } } /* Laptop: 1024px to 1349px */ @media only screen and (min-width: 1024px) and (max-width: 1349px) { #container-7b1ec0aa84 { } } /* Desktop: 1350px and up */ @media only screen and (min-width: 1350px) { #container-7b1ec0aa84 { } }
<> #container-e2fc2ed622 { } /* Mobile: up to 639px */ @media only screen and (max-width: 639px) { #container-e2fc2ed622 { } } /* Tablet: 640px to 1023px */ @media only screen and (min-width: 640px) and (max-width: 1023px) { #container-e2fc2ed622 { } } /* Laptop: 1024px to 1349px */ @media only screen and (min-width: 1024px) and (max-width: 1349px) { #container-e2fc2ed622 { } } /* Desktop: 1350px and up */ @media only screen and (min-width: 1350px) { #container-e2fc2ed622 { } }
<> #nvlink-fusion { background-color:#000000; } /* Mobile: up to 639px */ @media only screen and (max-width: 639px) { #nvlink-fusion { } } /* Tablet: 640px to 1023px */ @media only screen and (min-width: 640px) and (max-width: 1023px) { #nvlink-fusion { } } /* Laptop: 1024px to 1349px */ @media only screen and (min-width: 1024px) and (max-width: 1349px) { #nvlink-fusion { } } /* Desktop: 1350px and up */ @media only screen and (min-width: 1350px) { #nvlink-fusion { } }
<> #nv-separator-23c0c5b938 .cmp-separator__separator{ background-color: #cccccc; } @media screen and (min-width:1350px){ #nv-separator-23c0c5b938{ height:2px; } } @media screen and (min-width:1024px) and (max-width:1349px){ #nv-separator-23c0c5b938{ height:2px; } } @media screen and (min-width:640px) and (max-width:1023px){ #nv-separator-23c0c5b938{ height:2px; } } @media screen and (max-width: 639px){ #nv-separator-23c0c5b938{ height:2px; } }

从企业级计算扩展至百亿亿次级 (Exascale) 计算

<> #container-925dd25a78 { } /* Mobile: up to 639px */ @media only screen and (max-width: 639px) { #container-925dd25a78 { } } /* Tablet: 640px to 1023px */ @media only screen and (min-width: 640px) and (max-width: 1023px) { #container-925dd25a78 { } } /* Laptop: 1024px to 1349px */ @media only screen and (min-width: 1024px) and (max-width: 1349px) { #container-925dd25a78 { } } /* Desktop: 1350px and up */ @media only screen and (min-width: 1350px) { #container-925dd25a78 { } }

通过完全连接实现非凡性能

Vera Rubin NVL72 中的 NVLink Switch 芯片可实现 72 个 GPU 的全连接无阻塞计算网络。NVLink Switch 能以惊人的 3.6 TB/s 速度互连每对 GPU。支持全线速 all-to-all 通信。Vera Rubin NVL72 中的 72 个 GPU 可整体用作单个高性能加速器,支持高达 3.6 exaFLOPS 的 AI 计算能力。

<> #container-96d3953c78 { } /* Mobile: up to 639px */ @media only screen and (max-width: 639px) { #container-96d3953c78 { } } /* Tablet: 640px to 1023px */ @media only screen and (min-width: 640px) and (max-width: 1023px) { #container-96d3953c78 { } } /* Laptop: 1024px to 1349px */ @media only screen and (min-width: 1024px) and (max-width: 1349px) { #container-96d3953c78 { } } /* Desktop: 1350px and up */ @media only screen and (min-width: 1350px) { #container-96d3953c78 { } }

功能强大的 AI 和 HPC 平台

NVLink 和 NVLink 交换机是整个 NVIDIA 数据中心解决方案的必要基础模组,该解决方案融合了来自 NVIDIA AI Enterprise 软件套件和 NVIDIA NGC™ 目录中的硬件、网络、软件、库以及优化的 AI 模型和应用。作为强大的端到端 AI 和高性能计算平台,它让研究人员能够更快地得到真实的结果,并能将解决方案部署到生产环境中,从而在各种规模上实现出色的加速。

<> #container-e54f1f2cd6 { } /* Mobile: up to 639px */ @media only screen and (max-width: 639px) { #container-e54f1f2cd6 { } } /* Tablet: 640px to 1023px */ @media only screen and (min-width: 640px) and (max-width: 1023px) { #container-e54f1f2cd6 { } } /* Laptop: 1024px to 1349px */ @media only screen and (min-width: 1024px) and (max-width: 1349px) { #container-e54f1f2cd6 { } } /* Desktop: 1350px and up */ @media only screen and (min-width: 1350px) { #container-e54f1f2cd6 { } }
<> #container-4e7e8d8566 { } /* Mobile: up to 639px */ @media only screen and (max-width: 639px) { #container-4e7e8d8566 { } } /* Tablet: 640px to 1023px */ @media only screen and (min-width: 640px) and (max-width: 1023px) { #container-4e7e8d8566 { } } /* Laptop: 1024px to 1349px */ @media only screen and (min-width: 1024px) and (max-width: 1349px) { #container-4e7e8d8566 { } } /* Desktop: 1350px and up */ @media only screen and (min-width: 1350px) { #container-4e7e8d8566 { } }
<> #nvlink-switch-specifications { background-color:#f7f7f7; } /* Mobile: up to 639px */ @media only screen and (max-width: 639px) { #nvlink-switch-specifications { } } /* Tablet: 640px to 1023px */ @media only screen and (min-width: 640px) and (max-width: 1023px) { #nvlink-switch-specifications { } } /* Laptop: 1024px to 1349px */ @media only screen and (min-width: 1024px) and (max-width: 1349px) { #nvlink-switch-specifications { } } /* Desktop: 1350px and up */ @media only screen and (min-width: 1350px) { #nvlink-switch-specifications { } }

深入了解 NVIDIA Rubin 平台。

<> #container-74c05d02d8 { background-color:#1A1A1A; } /* Mobile: up to 639px */ @media only screen and (max-width: 639px) { #container-74c05d02d8 { } } /* Tablet: 640px to 1023px */ @media only screen and (min-width: 640px) and (max-width: 1023px) { #container-74c05d02d8 { } } /* Laptop: 1024px to 1349px */ @media only screen and (min-width: 1024px) and (max-width: 1349px) { #container-74c05d02d8 { } } /* Desktop: 1350px and up */ @media only screen and (min-width: 1350px) { #container-74c05d02d8 { } }
<> .unibrow-subnav{ position: relative; top: auto !important; } <> .global-nav.gtc .geo-locator { display: none; } :lang(zh-cn){ .nv-button.nv-button-text.nv-button-caret{ a[href]:not([href*="nvidia.com"]):not([href^="/"]):not([href^="#"]):not([href^="."]):not([href^="javascript"]):not([target^="_modal"]):not(.cmp-image__link):not(.cmp-teaser__link-entire-card){ position:relative; &:after{ content: "" !important; } } } .nv-teaser .nv-teaser-text-link:not([href*="nvidia.com"]):not([href^="/"]):not([href^="#"]):not([href^="."]):not([href^="javascript"]):not([target^="_modal"]):not(.cmp-image__link):not(.cmp-teaser__link-entire-card), .nv-overlay .nv-teaser-text-link:not([href*="nvidia.com"]):not([href^="/"]):not([href^="#"]):not([href^="."]):not([href^="javascript"]):not([target^="_modal"]):not(.cmp-image__link):not(.cmp-teaser__link-entire-card){ position:relative; &:after{ content: "" !important; } } }