{"id":10270,"date":"2024-07-10T08:56:12","date_gmt":"2024-07-10T08:56:12","guid":{"rendered":"https:\/\/www.fibermall.com\/blog\/?p=10270"},"modified":"2024-07-10T09:08:00","modified_gmt":"2024-07-10T09:08:00","slug":"hardware-configuration-and-network-design-for-large-scale-gpu-clusters","status":"publish","type":"post","link":"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm","title":{"rendered":"Hardware Configuration and Network Design for Large-Scale GPU Clusters"},"content":{"rendered":"\n<div id=\"ez-toc-container\" class=\"ez-toc-v2_0_76 ez-toc-wrap-left counter-hierarchy ez-toc-counter ez-toc-grey ez-toc-container-direction\">\n<div class=\"ez-toc-title-container\">\n<p class=\"ez-toc-title\" style=\"cursor:inherit\">Table of Contents<\/p>\n<span class=\"ez-toc-title-toggle\"><a href=\"#\" class=\"ez-toc-pull-right ez-toc-btn ez-toc-btn-xs ez-toc-btn-default ez-toc-toggle\" aria-label=\"Toggle Table of Content\"><span class=\"ez-toc-js-icon-con\"><span class=\"\"><span class=\"eztoc-hide\" style=\"display:none;\">Toggle<\/span><span class=\"ez-toc-icon-toggle-span\"><svg style=\"fill: #999;color:#999\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\" class=\"list-377408\" width=\"20px\" height=\"20px\" viewBox=\"0 0 24 24\" fill=\"none\"><path d=\"M6 6H4v2h2V6zm14 0H8v2h12V6zM4 11h2v2H4v-2zm16 0H8v2h12v-2zM4 16h2v2H4v-2zm16 0H8v2h12v-2z\" fill=\"currentColor\"><\/path><\/svg><svg style=\"fill: #999;color:#999\" class=\"arrow-unsorted-368013\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\" width=\"10px\" height=\"10px\" viewBox=\"0 0 24 24\" version=\"1.2\" baseProfile=\"tiny\"><path d=\"M18.2 9.3l-6.2-6.3-6.2 6.3c-.2.2-.3.4-.3.7s.1.5.3.7c.2.2.4.3.7.3h11c.3 0 .5-.1.7-.3.2-.2.3-.5.3-.7s-.1-.5-.3-.7zM5.8 14.7l6.2 6.3 6.2-6.3c.2-.2.3-.5.3-.7s-.1-.5-.3-.7c-.2-.2-.4-.3-.7-.3h-11c-.3 0-.5.1-.7.3-.2.2-.3.5-.3.7s.1.5.3.7z\"\/><\/svg><\/span><\/span><\/span><\/a><\/span><\/div>\n<nav><ul class='ez-toc-list ez-toc-list-level-1 ' ><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-1\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\/#Background\" >Background<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-2\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\/#Relevant_Components\" >Relevant Components<\/a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-3\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\/#GPUs\" >GPUs<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-4\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\/#HGX_Servers\" >HGX Servers<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-5\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\/#Networking\" >Networking<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-6\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\/#Data_Center_Network_DCN_Topology\" >Data Center Network (DCN) Topology<\/a><\/li><\/ul><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-7\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\/#NVIDIA_DGX_SuperPod_%E2%80%93_A100\" >NVIDIA DGX SuperPod &#8211; A100<\/a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-8\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\/#DGX_A100_System\" >DGX A100 System<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-9\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\/#SuperPod_SU\" >SuperPod SU<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-10\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\/#Spine_Rack\" >Spine Rack<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-11\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\/#DGX_SuperPod_100-node\" >DGX SuperPod 100-node<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-12\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\/#DGX_SuperPod_140-node\" >DGX SuperPod 140-node<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-13\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\/#Storage_Rack\" >Storage Rack<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-14\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\/#DGX_SuperPod_Storage_Fabric\" >DGX SuperPod Storage Fabric<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-15\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\/#Additional_Configurations\" >Additional Configurations<\/a><\/li><\/ul><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-16\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\/#NVIDIA_DGX_SuperPod_%E2%80%93_H100\" >NVIDIA DGX SuperPod &#8211; H100<\/a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-17\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\/#DGX_H100_System\" >DGX H100 System<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-18\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\/#SuperPod_SU-2\" >SuperPod SU<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-19\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\/#Management_Rack\" >Management Rack<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-20\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\/#DGX_SuperPod_127-node\" >DGX SuperPod 127-node<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-21\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\/#Additional_Configurations-2\" >Additional Configurations<\/a><\/li><\/ul><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-22\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\/#Industry_GPU_Training_Cluster_Solutions\" >Industry GPU Training Cluster Solutions<\/a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-23\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\/#Two-Level_Fat-Tree_Topology\" >Two-Level Fat-Tree Topology<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-24\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\/#FiberMall_Two-Level_Fat-Tree_Topology\" >FiberMall Two-Level Fat-Tree Topology<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-25\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\/#Three-Level_Fat-Tree_Topology\" >Three-Level Fat-Tree Topology<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-26\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\/#Related_Posts\" >Related Posts<\/a><\/li><\/ul><\/li><\/ul><\/nav><\/div>\n<h2 class=\"wp-block-heading\" id=\"h-background\"><span class=\"ez-toc-section\" id=\"Background\"><\/span><strong>Background<\/strong><strong><\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p>Since OpenAI introduced ChatGPT, large language models (LLMs) have gained significant attention and rapid development. Many enterprises are investing in LLM pre-training to keep up with this trend. However, training a 100B-scale LLM typically requires substantial computational resources, such as clusters equipped with thousands of GPUs. For example, the Falcon series model trains a 180B model on a 4096 A100 GPU cluster, taking nearly 70 days for 3.5T tokens. As data scales continue to grow, the demand for compute power increases. Meta, for instance, trained its LLaMA3 series model using 15T tokens across two 24K H100 clusters.<\/p>\n\n\n\n<p>In this article, we delve into the components and configurations involved in building large-scale GPU clusters. We\u2019ll cover different GPU types, server configurations, network devices (such as network cards, switches, and optical modules), and data center network topologies (e.g., 3-Tier, Fat-Tree). Specifically, we\u2019ll explore NVIDIA\u2019s DGX A100 SuperPod and DGX H100 SuperPod configurations, as well as common topologies used in multi-GPU clusters.<\/p>\n\n\n\n<p>Keep in mind that constructing ultra-large GPU clusters is an extremely complex endeavor, and this article only scratches the surface. In practical cluster deployment, storage networks, management networks, and other aspects come into play, but we won\u2019t delve into those details here. Additionally, network topology designs vary based on different application scenarios. Our focus will be on tree-based topologies commonly used in large-scale AI GPU clusters. Lastly, we won\u2019t cover critical components like power systems and cooling systems, which are essential for maintaining and operating GPU clusters.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"h-relevant-components\"><span class=\"ez-toc-section\" id=\"Relevant_Components\"><\/span><strong>Relevant Components<\/strong><strong><\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-gpus\"><span class=\"ez-toc-section\" id=\"GPUs\"><\/span><strong>GPUs<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>The chart below illustrates Ampere, Hopper, and the latest Blackwell series GPUs. Notice that memory capacity, computational power, and NVLink capabilities are gradually improving:<\/p>\n\n\n\n<p>A100 -&gt; H100: FP16 dense compute increases by over 3x, while power consumption only rises from 400W to 700W.<\/p>\n\n\n\n<p>H200 -&gt; B200: FP16 dense compute doubles, with power consumption increasing from 700W to 1000W.<\/p>\n\n\n\n<p>B200 FP16 dense compute is approximately 7x that of A100, while power consumption is only 2.5x higher.<\/p>\n\n\n\n<p>Blackwell GPUs support FP4 precision, offering twice the compute power of FP8. Some comparisons between FP4 and Hopper\u2019s FP8 architecture show even more significant acceleration.<\/p>\n\n\n\n<p>Note that GB200 uses the full B200 chip, while B100 and B200 are corresponding cut-down versions.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/GB200-uses-the-full-B200-chip-1024x338.png\" alt=\"GB200 uses the full B200 chip\" class=\"wp-image-10273\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/GB200-uses-the-full-B200-chip-1024x338.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/GB200-uses-the-full-B200-chip-300x99.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/GB200-uses-the-full-B200-chip-768x253.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/GB200-uses-the-full-B200-chip.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-hgx-servers\"><span class=\"ez-toc-section\" id=\"HGX_Servers\"><\/span><strong>HGX Servers<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>HGX is a high-performance server from NVIDIA, usually containing 8 or 4 GPUs, typically paired with Intel or AMD CPUs, and using NVLink and NVSwitch to achieve full interconnection (8 GPUs are usually the upper limit of NVLink full interconnection, except for NVL and SuperPod).<\/p>\n\n\n\n<p>From HGX A100 -&gt; HGX H100 and HGX H200, the FP16 dense computing power increased by 3.3 times, while the power consumption is less than 2 times.<\/p>\n\n\n\n<p>From HGX H100 and HGX H200 -&gt; HGX B100 and HGX B200, the FP16 dense computing power increased by about 2 times, while the power consumption is similar, at most not more than 50%.<\/p>\n\n\n\n<p>It should be noted that:<\/p>\n\n\n\n<p>The network of HGX B100 and HGX B200 has basically not been upgraded, and the IB network card is still 8x400Gb\/s.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/HGX-B100-and-HGX-B200-1024x558.png\" alt=\"HGX B100 and HGX B200\" class=\"wp-image-10274\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/HGX-B100-and-HGX-B200-1024x558.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/HGX-B100-and-HGX-B200-300x163.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/HGX-B100-and-HGX-B200-768x418.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/HGX-B100-and-HGX-B200.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>NVIDIA DGX and HGX are two high-performance solutions designed for deep learning, artificial intelligence, and large-scale computing needs. However, they differ in design and target applications:<\/p>\n\n\n\n<p>DGX:<\/p>\n\n\n\n<p>Targeted at general consumers.<\/p>\n\n\n\n<p>Provides plug-and-play high-performance solutions.<\/p>\n\n\n\n<p>Comes with comprehensive software support, including NVIDIA\u2019s deep learning software stack, drivers, and tools.<\/p>\n\n\n\n<p>Typically pre-built and closed systems.<\/p>\n\n\n\n<p>HGX:<\/p>\n\n\n\n<p>Targeted at cloud service providers and large-scale data center operators.<\/p>\n\n\n\n<p>Suitable for building custom high-performance solutions.<\/p>\n\n\n\n<p>Offers modular design, allowing customers to customize hardware based on their requirements.<\/p>\n\n\n\n<p>Usually provided as a hardware platform or reference architecture.<\/p>\n\n\n\n<p>Regarding networking:<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-networking\"><span class=\"ez-toc-section\" id=\"Networking\"><\/span><strong>Networking<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"h-network-cards\"><strong>Network Cards<\/strong><\/h4>\n\n\n\n<p>We\u2019ll focus on ConnectX-5\/6\/7\/8, which are high-speed network cards from Mellanox.<\/p>\n\n\n\n<p>These cards support both Ethernet and InfiniBand (IB).<\/p>\n\n\n\n<p>ConnectX-5 was released in 2016, followed by ConnectX-6 in 2020, ConnectX-7 in 2022, and ConnectX-8, which was introduced by Jensen Huang during the 2024 GTC conference (though detailed specifications are not yet available).<\/p>\n\n\n\n<p>Each generation roughly doubles the total bandwidth, and the next generation is estimated to reach 1.6 Tbps.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Network-Cards.png\" alt=\"Network Cards\" class=\"wp-image-10276\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Network-Cards.png 949w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Network-Cards-300x92.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Network-Cards-768x236.png 768w\" sizes=\"(max-width: 949px) 100vw, 949px\" \/><\/figure>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"h-switches\"><strong>Switches<\/strong><\/h4>\n\n\n\n<p>NVIDIA also offers switches for both Ethernet and <a href=\"https:\/\/www.fibermall.com\/collection\/infiniband.htm\" target=\"_blank\" rel=\"noreferrer noopener\">InfiniBand<\/a> (IB). These switches often have dozens or even hundreds of ports, corresponding to a total throughput (Bidirectional Switching Capacity) calculated as the maximum bandwidth multiplied by the number of ports, with the \u201c2\u201d\u00a0indicating bidirectional communication.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Spectrum-X-Series-Ethernet-Switches-1-1024x292.png\" alt=\"Spectrum-X Series Ethernet Switches\" class=\"wp-image-10277\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Spectrum-X-Series-Ethernet-Switches-1-1024x292.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Spectrum-X-Series-Ethernet-Switches-1-300x86.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Spectrum-X-Series-Ethernet-Switches-1-768x219.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Spectrum-X-Series-Ethernet-Switches-1.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><strong>Spectrum-X Series Ethernet Switches<\/strong><strong><\/strong><\/p>\n\n\n\n<p>Quantum-X Series InfiniBand Switches:<\/p>\n\n\n\n<p>These switches deliver 400 Gb\/s throughput.<\/p>\n\n\n\n<p>They excel in high-performance computing (HPC), AI, and hyperscale cloud infrastructures.<\/p>\n\n\n\n<p>Quantum-X switches offer robust performance while minimizing complexity and cost.<\/p>\n\n\n\n<p>In addition to Mellanox switches, many data centers now adopt modular switches (such as the Arista 7800 series) alongside traditional options. For instance, Meta recently built two GPU clusters with 24K H100 GPUs, utilizing Arista 7800 switches. The 7800 series includes modular switches like the 7816LR3 and 7816R3, which can provide 576 ports of 400G high-speed bandwidth. These switches use efficient internal buses or switch backplanes for low-latency data transmission and processing.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Modular-Switch-1024x872.png\" alt=\"Modular Switch\" class=\"wp-image-10278\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Modular-Switch-1024x872.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Modular-Switch-300x256.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Modular-Switch-768x654.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Modular-Switch.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"h-optical-modules\"><strong>Optical Modules<\/strong><\/h4>\n\n\n\n<p>Optical modules play a crucial role in optical fiber communication. They convert electrical signals into optical signals, which are then transmitted through optical fibers. These modules offer higher transmission rates, longer distances, and greater immunity to electromagnetic interference. Typically, an optical module consists of a transmitter (for converting electrical to optical signals) and a receiver (for converting optical to electrical signals).<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Optical-Modules.png\" alt=\"Optical Modules\" class=\"wp-image-10279\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Optical-Modules.png 800w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Optical-Modules-300x106.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Optical-Modules-768x272.png 768w\" sizes=\"(max-width: 800px) 100vw, 800px\" \/><\/figure>\n\n\n\n<p>Two commonly used optical module interface types are:<\/p>\n\n\n\n<p>SFP (Small Form-factor Pluggable): SFP modules usually operate as single transmission channels (using one fiber or a pair of fibers).<\/p>\n\n\n\n<p>QSFP (Quad Small Form-factor Pluggable): QSFP modules support multiple transmission channels. QSFP-DD (Double Density) further enhances port density by using 8 channels.<\/p>\n\n\n\n<p>Recently, the OSFP (Octal Small Form-factor Pluggable) packaging has emerged, specifically designed for high-bandwidth scenarios like 400Gbps and 800Gbps. OSFP modules have 8 channels and are slightly larger than QSFP-DD. They are not compatible with SFP and QSFP interfaces and require converters. The diagram below illustrates <a href=\"https:\/\/www.fibermall.com\/store-21951-400g-osfp.htm\" target=\"_blank\" rel=\"noreferrer noopener\">400Gbps OSFP modules <\/a>for different transmission distances (100m, 500m, 2km, and 10km).<\/p>\n\n\n\n<p>For various distances, consider the following module choices:<\/p>\n\n\n\n<p>Between Core and Spine layers: Use 10km 400G LR4 or 800G 2xLR4.<\/p>\n\n\n\n<p>Between Spine and Leaf layers: Opt for 2km 400G FR4.<\/p>\n\n\n\n<p>Between Leaf and ToR (Top of Rack): Choose 500m 400G DR modules.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/400G-DR-1024x793.png\" alt=\"400G DR\" class=\"wp-image-10280\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/400G-DR-1024x793.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/400G-DR-300x232.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/400G-DR-768x595.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/400G-DR.png 1085w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-data-center-network-dcn-topology\"><span class=\"ez-toc-section\" id=\"Data_Center_Network_DCN_Topology\"><\/span><strong>Data Center Network (DCN) Topology<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"h-basic-concepts\">Basic Concepts<\/h4>\n\n\n\n<p>North-South Traffic: Refers to traffic coming from outside the data center. It includes not only internet-related traffic but also traffic between different data centers.<\/p>\n\n\n\n<p>East-West Traffic: Refers to traffic within the same data center. For example, it encompasses communication between different servers within the data center. In modern data centers, this type of traffic typically constitutes a significant portion, often accounting for 70% to 80% of the total.<\/p>\n\n\n\n<p>Common data center network (DCN) topologies are illustrated in the diagram below.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Data-Center-Network.png\" alt=\"Data Center Network\" class=\"wp-image-10281\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Data-Center-Network.png 1021w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Data-Center-Network-300x191.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Data-Center-Network-768x488.png 768w\" sizes=\"(max-width: 1021px) 100vw, 1021px\" \/><\/figure>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"h-multi-tier-dcn-architecture\"><strong>Multi-Tier DCN Architecture<\/strong><\/h4>\n\n\n\n<p>Multi-tier DCN architectures are prevalent, especially the 3-tier DCN architecture. This tree-based structure primarily manages North-South traffic and consists of three layers:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Core Layer: The core layer typically comprises high-capacity routers or switches.<\/li>\n\n\n\n<li>Aggregation Layer (Distribution Layer): Responsible for connecting access layer devices and providing routing, filtering, and traffic engineering between them.<\/li>\n\n\n\n<li>Access Layer: The access layer is where end-user devices connect directly to the network, facilitating the connection of user devices to the data center network.<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Multi-Tier-DCN-Architecture.png\" alt=\"Multi-Tier DCN Architecture\" class=\"wp-image-10282\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Multi-Tier-DCN-Architecture.png 769w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Multi-Tier-DCN-Architecture-300x201.png 300w\" sizes=\"(max-width: 769px) 100vw, 769px\" \/><\/figure>\n\n\n\n<p>In this architecture, it is generally assumed that not all access devices communicate simultaneously at maximum bandwidth. Therefore, a common practice is to allocate smaller total bandwidth as we move up the hierarchy. For example, the total bandwidth at the access layer might be 20 Gbps, while the distribution layer\u2019s total bandwidth could be only 1 Gbps. In extreme cases, if all devices communicate at maximum bandwidth, this can lead to blocking, increased latency, and unpredictable delays. This situation is often referred to as oversubscription, with the ratio (e.g., 20:1) indicating the oversubscription rate.<\/p>\n\n\n\n<p>Within this architecture, redundancy or backup mechanisms are typically present. Switches between the core and distribution layers may interconnect, potentially creating loops. To avoid loops, spanning tree protocols (such as the Spanning Tree Protocol, STP) are used. However, this can also result in wasted bandwidth due to redundancy.<\/p>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"h-clos-networks\"><strong>CLOS Networks<\/strong><\/h4>\n\n\n\n<p>CLOS networks are a multi-stage switching network structure initially proposed by Charles Clos in 1953. Although originally used for telephone exchanges, their principles and design are now widely applied in data centers and high-performance computing. The core idea is to provide high bandwidth and low-latency network services through a multi-stage interconnected structure while maintaining scalability.<\/p>\n\n\n\n<p>As shown in the diagram below, CLOS networks typically consist of three layers:<\/p>\n\n\n\n<p>Ingress Layer: Responsible for receiving external input signals.<\/p>\n\n\n\n<p>Middle Layer: Connects the ingress layer to the egress layer switches.<\/p>\n\n\n\n<p>Egress Layer: Responsible for sending data to the final destination.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/CLOS-networks.png\" alt=\"CLOS networks\" class=\"wp-image-10283\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/CLOS-networks.png 815w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/CLOS-networks-300x193.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/CLOS-networks-768x494.png 768w\" sizes=\"(max-width: 815px) 100vw, 815px\" \/><\/figure>\n\n\n\n<p>CLOS networks offer the following features and advantages:<\/p>\n\n\n\n<p>Non-Blocking: Ideally, a CLOS network design is non-blocking (without convergence), meaning that data transmission delays or losses do not occur due to switch bottlenecks.<\/p>\n\n\n\n<p>Scalability: By adding more layers and switches, CLOS networks can easily scale to support additional input and output connections without sacrificing performance.<\/p>\n\n\n\n<p>Redundancy: The design\u2019s multiple paths allow data to be transmitted via alternative routes even if certain switches or connections fail, enhancing overall network reliability.<\/p>\n\n\n\n<p>Flexibility: CLOS networks support various configurations to accommodate different system sizes and performance requirements.<\/p>\n\n\n\n<h4 class=\"wp-block-heading\" id=\"h-fat-tree-topology\"><strong>Fat-Tree Topology<\/strong><\/h4>\n\n\n\n<p>The Fat-Tree data center network (DCN) architecture is a specialized form of the CLOS network. It is widely used in high-performance computing and large-scale data centers.<\/p>\n\n\n\n<p>Charles Leiserson introduced this network topology in 1985. Unlike the traditional 3-tier tree networks, Fat-Tree topology has some unique features:<\/p>\n\n\n\n<p>All layer switches are replaced with low-cost switches.<\/p>\n\n\n\n<p>As we move upward in the hierarchy, the links \u201cthicken,\u201d&nbsp;maintaining consistent total bandwidth between layers to avoid bottlenecks.<\/p>\n\n\n\n<p>The number of switches and their connections are symmetric at each layer, ensuring balanced paths for devices and minimizing single points of failure.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/fat-tree-topologies-1024x545.png\" alt=\"fat-tree topologies\" class=\"wp-image-10284\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/fat-tree-topologies-1024x545.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/fat-tree-topologies-300x160.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/fat-tree-topologies-768x409.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/fat-tree-topologies.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>Maximizing End-to-End Bandwidth: The primary goal of Fat-Tree architecture is to maximize end-to-end bandwidth. It achieves a 1:1 oversubscription ratio, resulting in a non-blocking network.<\/p>\n\n\n\n<p>Switch Count and Port Configuration:<\/p>\n\n\n\n<p>In a K-port Fat-Tree network topology (where K is the number of ports per switch), all switches typically have the same number of ports.<\/p>\n\n\n\n<p>Let\u2019s explore the 2-layer and 3-layer Fat-Tree topologies:<\/p>\n\n\n\n<p><strong>2-Layer Fat-Tree Topology:<\/strong><strong><\/strong><\/p>\n\n\n\n<p>Spine Switches: K\/2 switches, each with K*(K\/2) ports.<\/p>\n\n\n\n<p>Leaf Switches: K switches, each with K*K ports.<\/p>\n\n\n\n<p>This configuration allows for a maximum of KK\/2 servers in a non-blocking network, requiring 3K\/2 network switches.<\/p>\n\n\n\n<p><strong>3-Layer Fat-Tree Topology:<\/strong><strong><\/strong><\/p>\n\n\n\n<p>Core Switches (Super Spine Switches): (K\/2)^2 switches, each with K*(K\/2)^2 ports.<\/p>\n\n\n\n<p>Spine Switches: 2*(K\/2)^2 switches, each with K2(K\/2)^2 ports.<\/p>\n\n\n\n<p>Leaf Switches: 2*(K\/2)^2 switches, each with K2(K\/2)^2 ports.<\/p>\n\n\n\n<p>This design supports a maximum of K2(K\/2)^2\/2 = K^3\/4 servers in a non-blocking network, requiring 5*K^2\/4 switches.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/2-layer-and-3-layer-Fat-Tree-topologies-1024x435.png\" alt=\"2-layer and 3-layer Fat-Tree topologies\" class=\"wp-image-10285\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/2-layer-and-3-layer-Fat-Tree-topologies-1024x435.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/2-layer-and-3-layer-Fat-Tree-topologies-300x128.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/2-layer-and-3-layer-Fat-Tree-topologies-768x326.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/2-layer-and-3-layer-Fat-Tree-topologies.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>For both 2-layer and 3-layer Fat-Tree topologies, the switch counts and port configurations follow specific patterns.<\/p>\n\n\n\n<p>Note that there are variations in terminology (e.g., Fat-Tree vs. Spine-Leaf), but we\u2019ll consider them all under the Fat-Tree umbrella.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"h-nvidia-dgx-superpod-a100\"><span class=\"ez-toc-section\" id=\"NVIDIA_DGX_SuperPod_%E2%80%93_A100\"><\/span><strong>NVIDIA DGX SuperPod &#8211; A100<\/strong><strong><\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-dgx-a100-system\"><span class=\"ez-toc-section\" id=\"DGX_A100_System\"><\/span><strong>DGX A100 System<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>The DGX A100 System, as shown in the diagram below, is a 6U configuration with the following components:<\/p>\n\n\n\n<p>8*A100 GPUs: Each GPU offers 600 GB\/s NVLink bandwidth.<\/p>\n\n\n\n<p>Total NVSwitch Bandwidth: The system achieves a total of 4.8 TB\/s NVSwitch bandwidth, with 640 GB of HBM2 memory (80 GB per GPU).<\/p>\n\n\n\n<p>Compute Connections (IB): There are 8 ConnectX-6 network cards, providing a combined total bandwidth of 8 * 200 Gbps.<\/p>\n\n\n\n<p>Storage Connections (IB): 2 connections for storage.<\/p>\n\n\n\n<p>In-Band Connection (Ethernet): Used for internal communication.<\/p>\n\n\n\n<p>Out-Band Connection (Ethernet): For management purposes.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/DGC-A100-system-1024x542.png\" alt=\"DGC A100 system\" class=\"wp-image-10286\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/DGC-A100-system-1024x542.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/DGC-A100-system-300x159.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/DGC-A100-system-768x406.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/DGC-A100-system.png 1136w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>Notably, the NVLink bandwidth is measured in bytes, while network bandwidth typically uses bits. In this system, the internal bandwidth reaches 4.8 TB\/s, whereas the overall network bandwidth is 1.6 Tbps, resulting in a 24-fold difference.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/DGX-A100-System-1024x467.png\" alt=\"DGX A100 System\" class=\"wp-image-10287\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/DGX-A100-System-1024x467.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/DGX-A100-System-300x137.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/DGX-A100-System-768x351.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/DGX-A100-System.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-superpod-su\"><span class=\"ez-toc-section\" id=\"SuperPod_SU\"><\/span><strong>SuperPod SU<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>The SuperPod SU (Scalable Unit), depicted in the figure, serves as the fundamental building block for constructing the DGX-SuperPod-A100. Here are its key components:<\/p>\n\n\n\n<p>Each SU includes 5 Compute Racks and 1 Leaf Network Rack.<\/p>\n\n\n\n<p>Each Compute Rack houses 4 DGX A100 Systems and 2 3U Power Distribution Units (PDUs), totaling 32 A100 GPUs per Compute Rack. Thus, an SU comprises 160 A100 GPUs.<\/p>\n\n\n\n<p>The Leaf Network Rack contains 8 Compute Switches (1U) and 2 Storage Switches (1U).<\/p>\n\n\n\n<p>Compute Switches utilize QM8790 200 Gb\/s IB switches, resulting in a total of 320 ports:<\/p>\n\n\n\n<p>160 ports connect to the ConnectX-6 network cards in the Compute Racks, providing 200 Gbps per GPU.<\/p>\n\n\n\n<p>The remaining 160 ports connect to the Spine Rack.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/SU-rack-elevations-1024x675.png\" alt=\"SU rack elevations\" class=\"wp-image-10288\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/SU-rack-elevations-1024x675.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/SU-rack-elevations-300x198.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/SU-rack-elevations-768x506.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/SU-rack-elevations.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>Some scenarios may also use Top-of-Rack (ToR) Switches within a cabinet for simpler cabling. However, this approach may lead to port wastage. For instance, due to power constraints and cooling challenges, GPU Servers are often limited to a single cabinet, reducing the number of network cards.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/GPU-Servers.png\" alt=\"GPU Servers\" class=\"wp-image-10289\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/GPU-Servers.png 737w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/GPU-Servers-300x151.png 300w\" sizes=\"(max-width: 737px) 100vw, 737px\" \/><\/figure>\n\n\n\n<p>Please note that while some industrial scenarios may use fewer network cards (e.g., 4&#215;200 Gbps) within an 8*A100 System, the overall network topology remains similar.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-spine-rack\"><span class=\"ez-toc-section\" id=\"Spine_Rack\"><\/span><strong>Spine Rack<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>As shown in the figure, a Spine Rack contains 20 1U Compute Switches, specifically QM8790 200 Gb\/s IB switches, totaling 800 ports. The remaining Out-of-band Switch and In-band Switch can be used for network management.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/spine-rack-elevations-1024x846.png\" alt=\"spine rack elevations\" class=\"wp-image-10290\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/spine-rack-elevations-1024x846.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/spine-rack-elevations-300x248.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/spine-rack-elevations-768x635.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/spine-rack-elevations-1536x1269.png 1536w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/spine-rack-elevations.png 1590w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-dgx-superpod-100-node\"><span class=\"ez-toc-section\" id=\"DGX_SuperPod_100-node\"><\/span><strong>DGX SuperPod 100-node<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>The figure below illustrates a 100-node DGX-SuperPOD, comprising 5 SUs and an additional Spine Rack.<\/p>\n\n\n\n<p>Each SU includes 8 Leaf Compute Switches (QM7890, 200 Gbps).<\/p>\n\n\n\n<p>Each Node\u2019s 8 ConnectX-6 NICs connect to 8 Leaf Compute Switches, with each ConnectX-6 corresponding to 1 GPU.<\/p>\n\n\n\n<p>The Leaf Compute Switches have 20 ports connecting to 20 Nodes within the SU, and an additional 20 ports connecting to the 20 Spine Compute Switches in the Spine Rack.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/compute-fabric-topology-for-100-node-stsyem-1024x521.png\" alt=\"compute fabric topology for 100 node stsyem\" class=\"wp-image-10291\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/compute-fabric-topology-for-100-node-stsyem-1024x521.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/compute-fabric-topology-for-100-node-stsyem-300x153.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/compute-fabric-topology-for-100-node-stsyem-768x391.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/compute-fabric-topology-for-100-node-stsyem.png 1187w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>This topology achieves a non-blocking network for 800 GPUs (any two GPUs can communicate):<\/p>\n\n\n\n<p>GPUs from different SUs connect via: ConnectX-6 -&gt; Leaf Switch -&gt; Spine Switch -&gt; Leaf Switch -&gt; ConnectX-6.<\/p>\n\n\n\n<p>GPUs within the same SU but different Nodes connect via: ConnectX-6 -&gt; Leaf Switch -&gt; ConnectX-6.<\/p>\n\n\n\n<p>GPUs within the same Node communicate via NVLink.<\/p>\n\n\n\n<p>The practical limit for 800 GPUs (each GPU corresponding to a 200 Gbps NIC port) using QM8790 is a 2-level Fat-Tree network. Beyond 800 GPUs, a 3-level Fat-Tree would be needed, allowing up to 16,000 GPUs.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-dgx-superpod-140-node\"><span class=\"ez-toc-section\" id=\"DGX_SuperPod_140-node\"><\/span><strong>DGX SuperPod 140-node<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>In a 100-node system where all Compute Switch ports are occupied, expanding to more GPUs requires transitioning from 2-layer to 3-layer switches. This involves adding a Core Compute Switch layer, still using QM8790 at 200 Gbps.<\/p>\n\n\n\n<p>Figure shows a 140-node SuperPod with 7 SUs, totaling 56 Leaf Switches. Ideally, 56 Leaf Switches would require 56 Spine Switches and 28 Core Switches. However, the actual design uses 80 Spine Switches, organized into 8 Groups (SG), each with 10 Spine Switches, and each Core Group (CG) with 14 Core Switches. This symmetric Fat-Tree topology simplifies management.<\/p>\n\n\n\n<p>Each Leaf Switch in an SU connects to 10 Spine Switches in the corresponding SG (20 ports per Leaf Switch). Spine Switches alternate connections to Core Switches (odd positions to odd Core Switches, even positions to even Core Switches).<\/p>\n\n\n\n<p>Each Core Switch connects to 40 Spine Switches.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/140-node-GHX-SuperPOD-1024x584.png\" alt=\"140-node GHX SuperPOD\" class=\"wp-image-10292\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/140-node-GHX-SuperPOD-1024x584.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/140-node-GHX-SuperPOD-300x171.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/140-node-GHX-SuperPOD-768x438.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/140-node-GHX-SuperPOD.png 1112w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>This configuration supports a 140*8=1120 GPU cluster, with each GPU having a ConnectX-6 200 Gbps NIC.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-storage-rack\"><span class=\"ez-toc-section\" id=\"Storage_Rack\"><\/span><strong>Storage Rack<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>As shown in the figure below, a Storage Rack contains 4 Storage Switches, also QM8790 200 Gbps IB switches, totaling 160 ports. The corresponding storage units are also present within the rack.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/storage-rack-elevations-1024x898.png\" alt=\"storage rack elevations\" class=\"wp-image-10293\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/storage-rack-elevations-1024x898.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/storage-rack-elevations-300x263.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/storage-rack-elevations-768x674.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/storage-rack-elevations-1536x1347.png 1536w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/storage-rack-elevations.png 1646w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-dgx-superpod-storage-fabric\"><span class=\"ez-toc-section\" id=\"DGX_SuperPod_Storage_Fabric\"><\/span><strong>DGX SuperPod Storage Fabric<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>Figure illustrates the Storage Fabric for the 140-node configuration. It comprises 18 Leaf Switches. Each SuperPod SU (Scalable Unit) contains 2 Leaf Network Racks and 1 Storage Rack. Additionally, there are 8 Spine Switches.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/storage-fabric-topology-for-140-node-system-1024x558.png\" alt=\"storage fabric  topology for 140-node system\" class=\"wp-image-10294\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/storage-fabric-topology-for-140-node-system-1024x558.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/storage-fabric-topology-for-140-node-system-300x164.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/storage-fabric-topology-for-140-node-system-768x419.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/storage-fabric-topology-for-140-node-system.png 1121w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-additional-configurations\"><span class=\"ez-toc-section\" id=\"Additional_Configurations\"><\/span><strong>Additional Configurations<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>Table 3 provides details on the Compute configurations for different nodes.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/compute-fabric-switch-and-cable-counts-1024x385.png\" alt=\"compute fabric switch and cable counts\" class=\"wp-image-10295\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/compute-fabric-switch-and-cable-counts-1024x385.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/compute-fabric-switch-and-cable-counts-300x113.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/compute-fabric-switch-and-cable-counts-768x289.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/compute-fabric-switch-and-cable-counts.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>Table 4 outlines the Storage configurations.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/storage-fabric-counts-1024x501.png\" alt=\"storage fabric counts\" class=\"wp-image-10296\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/storage-fabric-counts-1024x501.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/storage-fabric-counts-300x147.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/storage-fabric-counts-768x375.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/storage-fabric-counts.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"h-nvidia-dgx-superpod-h100\"><span class=\"ez-toc-section\" id=\"NVIDIA_DGX_SuperPod_%E2%80%93_H100\"><\/span><strong>NVIDIA DGX SuperPod &#8211; H100<\/strong><strong><\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-dgx-h100-system\"><span class=\"ez-toc-section\" id=\"DGX_H100_System\"><\/span><strong>DGX H100 System<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>The DGX H100 System (6U), as shown, includes:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>8 H100 GPUs, each with 900 GB\/s NVLink bandwidth.<\/li>\n\n\n\n<li>A total of 7.2 TB\/s NVSwitch bandwidth and 640 GB HBM3 memory (80 GB per GPU).<\/li>\n\n\n\n<li>4 OSFP Ports (IB) corresponding to 8 ConnectX-7 NICs, providing 8*400 Gbps bandwidth.<\/li>\n\n\n\n<li>Slots 1 and 2 with 2 ConnectX-7 NICs, offering 2*400 Gbps bandwidth.<\/li>\n\n\n\n<li>An In-Band Connection (Ethernet).<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/DGX-H100-System-1024x544.png\" alt=\"DGX H100 System\" class=\"wp-image-10297\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/DGX-H100-System-1024x544.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/DGX-H100-System-300x159.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/DGX-H100-System-768x408.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/DGX-H100-System.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/The-DGX-H100-System-6U-1024x516.png\" alt=\"The DGX H100 System (6U)\" class=\"wp-image-10298\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/The-DGX-H100-System-6U-1024x516.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/The-DGX-H100-System-6U-300x151.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/The-DGX-H100-System-6U-768x387.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/The-DGX-H100-System-6U.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>All 8 GPUs are fully interconnected via NVSwitch. The internal bandwidth reaches 7.2 TB\/s, while the overall network bandwidth is 3.2 Tbps, a difference of 22.5 times.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-superpod-su-0\"><span class=\"ez-toc-section\" id=\"SuperPod_SU-2\"><\/span><strong>SuperPod SU<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>Figure 2 depicts the fundamental building block of the DGX-SuperPod-H100, known as the SuperPod SU:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Each SU contains 8 Compute Racks, with each rack providing 40 kW.<\/li>\n\n\n\n<li>Each Compute Rack houses 4 DGX H100 Systems and 3 PDUs (Power Distribution Units), resulting in 32 H100 GPUs per Compute Rack. Thus, an SU accommodates 256 H100 GPUs.<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/complete-single-SU-rack-layout-1024x512.png\" alt=\"complete single SU rack layout\" class=\"wp-image-10299\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/complete-single-SU-rack-layout-1024x512.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/complete-single-SU-rack-layout-300x150.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/complete-single-SU-rack-layout-768x384.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/complete-single-SU-rack-layout.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-management-rack\"><span class=\"ez-toc-section\" id=\"Management_Rack\"><\/span><strong>Management Rack<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>In the DGX SuperPod corresponding to H100 GPUs, NVIDIA offers a Management Rack similar to the A100 series\u2019&nbsp;Spine and Storage Racks. Figure 3 provides an example (specific configurations may vary):<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>32 Leaf Compute Switches (<a class=\"wpil_keyword_link\" href=\"https:\/\/www.fibermall.com\/blog\/qm9700.htm\"   title=\"QM9700\" data-wpil-keyword-link=\"linked\" target=\"_blank\">QM9700<\/a>) offer 64 400 Gbps ports each. Theoretically, there are 1024 400 Gbps ports available to connect to the ConnectX-7 NICs on the nodes. The remaining 1024 ports connect precisely to 16 Spine Compute Switches, achieving a non-blocking network for 1024 GPUs.<\/li>\n\n\n\n<li>16 Spine Compute Switches (also QM9700) connect to half of the ports on 32 Leaf Compute Switches.<\/li>\n\n\n\n<li>8 Leaf Storage Switches (QM9700) are part of the setup.<\/li>\n\n\n\n<li>4 Spine Storage Switches (QM9700) complete the configuration.<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/management-rack-configuration-1-1024x732.png\" alt=\"management rack configuration\" class=\"wp-image-10301\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/management-rack-configuration-1-1024x732.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/management-rack-configuration-1-300x214.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/management-rack-configuration-1-768x549.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/management-rack-configuration-1.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-dgx-superpod-127-node\"><span class=\"ez-toc-section\" id=\"DGX_SuperPod_127-node\"><\/span><strong>DGX SuperPod 127-node<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>Figure 5 illustrates a 127-node DGX SuperPod with 4 Scalable Units (SUs) and an associated Management Rack. In theory, the Management Rack can connect to the 128 nodes across the 4 SUs. However, due to some Leaf Switches being connected to the Unified Fabric Manager (UFM), the actual number of nodes is 127.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/127node-1024x466.png\" alt=\"127node\" class=\"wp-image-10302\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/127node-1024x466.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/127node-300x136.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/127node-768x349.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/127node.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-additional-configurations-0\"><span class=\"ez-toc-section\" id=\"Additional_Configurations-2\"><\/span><strong>Additional Configurations<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>As shown in Table 3, using QM9700 switches, a 2-level Fat-Tree can achieve a non-blocking network for up to 6464\/2=2048 GPUs (corresponding to 8 SUs). A 3-level Fat-Tree can support up to 6464*64\/4=65536 GPUs. In practice, the configuration includes 64 SUs, totaling 16384 GPUs.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/large-superpod-component-couts-1024x410.png\" alt=\"large superpod component couts\" class=\"wp-image-10303\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/large-superpod-component-couts-1024x410.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/large-superpod-component-couts-300x120.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/large-superpod-component-couts-768x307.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/large-superpod-component-couts.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"h-industry-gpu-training-cluster-solutions\"><span class=\"ez-toc-section\" id=\"Industry_GPU_Training_Cluster_Solutions\"><\/span><strong>Industry GPU Training Cluster Solutions<\/strong><strong><\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-two-level-fat-tree-topology\"><span class=\"ez-toc-section\" id=\"Two-Level_Fat-Tree_Topology\"><\/span><strong>Two-Level Fat-Tree Topology<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>The common two-level non-blocking Fat-Tree topology (Spine-Leaf) is prevalent for 8-GPU training machines. Within a single machine, the 8 GPUs are fully interconnected via NVLink + NVSwitch, with communication bandwidth significantly higher than the network bandwidth. Therefore, it\u2019s standard practice to connect each GPU\u2019s NIC to different switches:<\/p>\n\n\n\n<p>Each Group contains 8 Leaf Switches, corresponding to the 8 GPUs in a machine.<\/p>\n\n\n\n<p>Assuming a Leaf Switch has 128 ports, 64 ports connect to the corresponding GPUs\u2019&nbsp;NICs, resulting in 64*8=512 GPUs per Group. Leaf Switch 1 connects all Node 1 GPUs\u2019&nbsp;NICs, and so on.<\/p>\n\n\n\n<p>This feature can be leveraged when designing distributed training strategies.<\/p>\n\n\n\n<p>To achieve Full Mesh between Spine and Leaf Switches, each Leaf Switch connects to one Spine Switch. Thus, there are 64 Spine Switches, and each Spine Switch connects to all 128 Leaf Switches. This requires 16 Groups.<\/p>\n\n\n\n<p>In summary, a maximum of 192 switches with 128 ports each can support 512*16=8192 GPUs.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Two-Level-Fat-Tree-Topology-1024x543.png\" alt=\"Two-Level Fat-Tree Topology\" class=\"wp-image-10304\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Two-Level-Fat-Tree-Topology-1024x543.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Two-Level-Fat-Tree-Topology-300x159.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Two-Level-Fat-Tree-Topology-768x407.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/Two-Level-Fat-Tree-Topology.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-fibermall-two-level-fat-tree-topology\"><span class=\"ez-toc-section\" id=\"FiberMall_Two-Level_Fat-Tree_Topology\"><\/span><strong>FiberMall Two-Level Fat-Tree Topology<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>The FiberMall standard solution for two-level Fat-Tree is similar to the previously described topology. However, it uses 64-port switches.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img decoding=\"async\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/FiberMall-Two-Level-Fat-Tree-Topology-1024x500.png\" alt=\"FiberMall Two-Level Fat-Tree Topology\" class=\"wp-image-10305\" style=\"width:800px\" width=\"800\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/FiberMall-Two-Level-Fat-Tree-Topology-1024x500.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/FiberMall-Two-Level-Fat-Tree-Topology-300x146.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/FiberMall-Two-Level-Fat-Tree-Topology-768x375.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/FiberMall-Two-Level-Fat-Tree-Topology.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>Due to the 64-port 400 Gbps switches:<\/p>\n\n\n\n<p>Leaf and Spine Switches are halved (64 and 32, respectively).<\/p>\n\n\n\n<p>GPU support reduces to 1\/4, resulting in 2*(64\/2)*(64\/2)=2048 GPUs.<\/p>\n\n\n\n<p>The total optical module count includes switch ports and GPU NICs: (64+32)*64+2048=8192.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-three-level-fat-tree-topology\"><span class=\"ez-toc-section\" id=\"Three-Level_Fat-Tree_Topology\"><\/span><strong>Three-Level Fat-Tree Topology<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>The common three-level non-blocking Fat-Tree topology (SuperSpine-Spine-Leaf) treats the two-level Spine-Leaf as a Pod.<\/p>\n\n\n\n<p>Since Spine Switches also connect to SuperSpine Switches, the number of Groups is halved. Each Pod has 64 Spine Switches, corresponding to 4096 GPUs.<\/p>\n\n\n\n<p>Multiple Pods can further build 64 SuperSpine Fabrics, each fully interconnected with Spine Switches from different Pods. For example, with 8 Pods, each Fabric needs only 4 128-port SuperSpine Switches.<\/p>\n\n\n\n<p>The configuration for 8 Pods includes:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Total GPUs: 4096*8=32768<\/li>\n\n\n\n<li>SuperSpine Switches: 64*4=256<\/li>\n\n\n\n<li>Spine Switches: 64*8=512<\/li>\n\n\n\n<li>Leaf Switches: 64*8=512<\/li>\n\n\n\n<li>Total Switches: 256+512+512=1280<\/li>\n\n\n\n<li>Total optical modules: 1280*128+32768=196608<\/li>\n<\/ul>\n\n\n\n<p>The theoretical maximum supports 128 Pods, corresponding to:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>GPUs: 4096128=524288=2(128\/2)^3<\/li>\n\n\n\n<li>SuperSpine Switches: 64*64=4096=(128\/2)^2<\/li>\n\n\n\n<li>Spine Switches: 64128=8192=2(128\/2)^2<\/li>\n\n\n\n<li>Leaf Switches: 64128=8192=2(128\/2)^2<\/li>\n\n\n\n<li>Total Switches: 4096+8192+8192=20480=5*(128\/2)^2<\/li>\n<\/ul>\n<style>\r\n\r\n        .lwrp.link-whisper-related-posts{\r\n            \r\n            margin-top: 40px;\nmargin-bottom: 30px;\r\n        }\r\n        .lwrp .lwrp-title{\r\n            \r\n            \r\n        }\r\n        .lwrp .lwrp-description{\r\n            \r\n            \r\n\r\n        }\r\n        .lwrp .lwrp-list-container{\r\n        }\r\n        .lwrp .lwrp-list-multi-container{\r\n            display: flex;\r\n        }\r\n        .lwrp .lwrp-list-double{\r\n            width: 48%;\r\n        }\r\n        .lwrp .lwrp-list-triple{\r\n            width: 32%;\r\n        }\r\n        .lwrp .lwrp-list-row-container{\r\n            display: flex;\r\n            justify-content: space-between;\r\n        }\r\n        .lwrp .lwrp-list-row-container .lwrp-list-item{\r\n            width: calc(100% - 20px);\r\n        }\r\n        .lwrp .lwrp-list-item:not(.lwrp-no-posts-message-item){\r\n            \r\n            list-style: decimal;\r\n        }\r\n        .lwrp .lwrp-list-item img{\r\n            max-width: 100%;\r\n            height: auto;\r\n        }\r\n        .lwrp .lwrp-list-item.lwrp-empty-list-item{\r\n            background: initial !important;\r\n        }\r\n        .lwrp .lwrp-list-item .lwrp-list-link .lwrp-list-link-title-text,\r\n        .lwrp .lwrp-list-item .lwrp-list-no-posts-message{\r\n            \r\n                \r\n        }\r\n        @media screen and (max-width: 480px) {\r\n            .lwrp.link-whisper-related-posts{\r\n                \r\n                \r\n            }\r\n            .lwrp .lwrp-title{\r\n                \r\n                \r\n            }\r\n            .lwrp .lwrp-description{\r\n                \r\n                \r\n            }\r\n            .lwrp .lwrp-list-multi-container{\r\n                flex-direction: column;\r\n            }\r\n            .lwrp .lwrp-list-multi-container ul.lwrp-list{\r\n                margin-top: 0px;\r\n                margin-bottom: 0px;\r\n                padding-top: 0px;\r\n                padding-bottom: 0px;\r\n            }\r\n            .lwrp .lwrp-list-double,\r\n            .lwrp .lwrp-list-triple{\r\n                width: 100%;\r\n            }\r\n            .lwrp .lwrp-list-row-container{\r\n                justify-content: initial;\r\n                flex-direction: column;\r\n            }\r\n            .lwrp .lwrp-list-row-container .lwrp-list-item{\r\n                width: 100%;\r\n            }\r\n            .lwrp .lwrp-list-item:not(.lwrp-no-posts-message-item){\r\n                \r\n                \r\n            }\r\n            .lwrp .lwrp-list-item .lwrp-list-link .lwrp-list-link-title-text,\r\n            .lwrp .lwrp-list-item .lwrp-list-no-posts-message{\r\n                \r\n                    \r\n            }\r\n        }<\/style>\r\n<div id=\"link-whisper-related-posts-widget\" class=\"link-whisper-related-posts lwrp\">\r\n            <h3 class=\"lwrp-title\">Related Posts<\/h3>    \r\n        <div class=\"lwrp-list-container\">\r\n                                            <ul class=\"lwrp-list lwrp-list-single\">\r\n                    <li class=\"lwrp-list-item\"><a href=\"https:\/\/www.fibermall.com\/blog\/mpo-tester.htm\" class=\"lwrp-list-link\"><span class=\"lwrp-list-link-title-text\">Everything You Need to Know About MPO Tester for Fiber Optic Testing<\/span><\/a><\/li><li class=\"lwrp-list-item\"><a href=\"https:\/\/www.fibermall.com\/blog\/dgx-h200.htm\" class=\"lwrp-list-link\"><span class=\"lwrp-list-link-title-text\">The Game-Changing NVIDIA DGX H200 Delivered to OpenAI<\/span><\/a><\/li><li class=\"lwrp-list-item\"><a href=\"https:\/\/www.fibermall.com\/video\/sfp28-er-optical-transceiver-watch-this-before-using.htm\" class=\"lwrp-list-link\"><span class=\"lwrp-list-link-title-text\">SFP28 ER optical transceiver: watch this before using<\/span><\/a><\/li><li class=\"lwrp-list-item\"><a href=\"https:\/\/www.fibermall.com\/blog\/25g-dwdm-sfp28.htm\" class=\"lwrp-list-link\"><span class=\"lwrp-list-link-title-text\">The Ultimate Guide to 25G DWDM SFP28 Transceivers<\/span><\/a><\/li><li class=\"lwrp-list-item\"><a href=\"https:\/\/www.fibermall.com\/blog\/why-do-25g-optical-transceivers-need-fec.htm\" class=\"lwrp-list-link\"><span class=\"lwrp-list-link-title-text\">Why does 25G SFP28\u00a0Need FEC?<\/span><\/a><\/li>                <\/ul>\r\n                        <\/div>\r\n<\/div>","protected":false},"excerpt":{"rendered":"<p>Background Since OpenAI introduced ChatGPT, large language models (LLMs) have gained significant attention and rapid development. Many enterprises are investing in LLM pre-training to keep up with this trend. However, training a 100B-scale LLM typically requires substantial computational resources, such as clusters equipped with thousands of GPUs. For example, the Falcon series model trains a [&hellip;]<\/p>\n","protected":false},"author":4,"featured_media":10298,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":"","_wpscppro_dont_share_socialmedia":false,"_wpscppro_custom_social_share_image":0,"_facebook_share_type":"default","_twitter_share_type":"default","_linkedin_share_type":"default","_pinterest_share_type":"default","_linkedin_share_type_page":"default","_instagram_share_type":"default","_medium_share_type":"","_threads_share_type":"","_google_business_share_type":"","_selected_social_profile":[{"id":"skM9ewvR8O","platform":"linkedin","platformKey":0,"name":"Jason Xue","type":"person","thumbnail_url":"https:\/\/media.licdn.com\/dms\/image\/C5603AQErPqKD0j6qBg\/profile-displayphoto-shrink_100_100\/0\/1599138392315?e=1723075200&v=beta&t=joEkh1OeKQ0F-QpAPv4xxQyBdGlHyccIQZauRSs6RvU","share_type":"default"}],"_wpsp_enable_custom_social_template":false,"_wpsp_social_scheduling":{"enabled":false,"datetime":null,"platforms":[],"status":"template_only","dateOption":"today","timeOption":"now","customDays":"","customHours":"","customDate":"","customTime":"","schedulingType":"absolute"},"_wpsp_active_default_template":true},"categories":[2,29],"tags":[],"class_list":["post-10270","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-blog","category-networking"],"acf":[],"yoast_head":"<!-- This site is optimized with the Yoast SEO Premium plugin v20.13 (Yoast SEO v25.8) - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>Hardware Configuration and Network Design for Large-Scale GPU Clusters - fibermall.com<\/title>\n<meta name=\"description\" content=\"In this article, we delve into the components and configurations involved in building large-scale GPU clusters.\" \/>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\" \/>\n<meta property=\"og:locale\" content=\"en_US\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Hardware Configuration and Network Design for Large-Scale GPU Clusters\" \/>\n<meta property=\"og:description\" content=\"Background Since OpenAI introduced ChatGPT, large language models (LLMs) have gained significant attention and rapid development. Many enterprises are\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\" \/>\n<meta property=\"og:site_name\" content=\"fibermall.com\" \/>\n<meta property=\"article:published_time\" content=\"2024-07-10T08:56:12+00:00\" \/>\n<meta property=\"article:modified_time\" content=\"2024-07-10T09:08:00+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/The-DGX-H100-System-6U.png\" \/>\n\t<meta property=\"og:image:width\" content=\"1080\" \/>\n\t<meta property=\"og:image:height\" content=\"544\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/png\" \/>\n<meta name=\"author\" content=\"Mia\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"Written by\" \/>\n\t<meta name=\"twitter:data1\" content=\"Mia\" \/>\n\t<meta name=\"twitter:label2\" content=\"Est. reading time\" \/>\n\t<meta name=\"twitter:data2\" content=\"21 minutes\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm#article\",\"isPartOf\":{\"@id\":\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\"},\"author\":{\"name\":\"Mia\",\"@id\":\"https:\/\/www.fibermall.com\/blog.htm\/#\/schema\/person\/56a37c7b0e1e8c61823cd6ef48e8dce3\"},\"headline\":\"Hardware Configuration and Network Design for Large-Scale GPU Clusters\",\"datePublished\":\"2024-07-10T08:56:12+00:00\",\"dateModified\":\"2024-07-10T09:08:00+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\"},\"wordCount\":3183,\"commentCount\":0,\"publisher\":{\"@id\":\"https:\/\/www.fibermall.com\/blog.htm\/#organization\"},\"image\":{\"@id\":\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm#primaryimage\"},\"thumbnailUrl\":\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/The-DGX-H100-System-6U.png\",\"articleSection\":[\"Blog\",\"Networking\"],\"inLanguage\":\"en-US\"},{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\",\"url\":\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\",\"name\":\"Hardware Configuration and Network Design for Large-Scale GPU Clusters - fibermall.com\",\"isPartOf\":{\"@id\":\"https:\/\/www.fibermall.com\/blog.htm\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm#primaryimage\"},\"image\":{\"@id\":\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm#primaryimage\"},\"thumbnailUrl\":\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/The-DGX-H100-System-6U.png\",\"datePublished\":\"2024-07-10T08:56:12+00:00\",\"dateModified\":\"2024-07-10T09:08:00+00:00\",\"description\":\"In this article, we delve into the components and configurations involved in building large-scale GPU clusters.\",\"breadcrumb\":{\"@id\":\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm#breadcrumb\"},\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm#primaryimage\",\"url\":\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/The-DGX-H100-System-6U.png\",\"contentUrl\":\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/The-DGX-H100-System-6U.png\",\"width\":1080,\"height\":544,\"caption\":\"The-DGX-H100-System-6U\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\/\/www.fibermall.com\/blog.htm\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Hardware Configuration and Network Design for Large-Scale GPU Clusters\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.fibermall.com\/blog.htm\/#website\",\"url\":\"https:\/\/www.fibermall.com\/blog.htm\/\",\"name\":\"fibermall.com\",\"description\":\"Optical Communication Expert\",\"publisher\":{\"@id\":\"https:\/\/www.fibermall.com\/blog.htm\/#organization\"},\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.fibermall.com\/blog.htm\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"en-US\"},{\"@type\":\"Organization\",\"@id\":\"https:\/\/www.fibermall.com\/blog.htm\/#organization\",\"name\":\"fibermall.com\",\"url\":\"https:\/\/www.fibermall.com\/blog.htm\/\",\"logo\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\/\/www.fibermall.com\/blog.htm\/#\/schema\/logo\/image\/\",\"url\":\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/12\/cropped-Fiber-Mall-Logo.jpg\",\"contentUrl\":\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/12\/cropped-Fiber-Mall-Logo.jpg\",\"width\":200,\"height\":67,\"caption\":\"fibermall.com\"},\"image\":{\"@id\":\"https:\/\/www.fibermall.com\/blog.htm\/#\/schema\/logo\/image\/\"}},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.fibermall.com\/blog.htm\/#\/schema\/person\/56a37c7b0e1e8c61823cd6ef48e8dce3\",\"name\":\"Mia\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\/\/www.fibermall.com\/blog.htm\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/secure.gravatar.com\/avatar\/2e83831e57346d5d451c45e43aa96d14fc758d8f72f7304810e123b36585382b?s=96&d=mm&r=g\",\"contentUrl\":\"https:\/\/secure.gravatar.com\/avatar\/2e83831e57346d5d451c45e43aa96d14fc758d8f72f7304810e123b36585382b?s=96&d=mm&r=g\",\"caption\":\"Mia\"},\"description\":\"Optical Transmission Consultant\",\"sameAs\":[\"https:\/\/www.fibermall.com\/\"],\"url\":\"https:\/\/www.fibermall.com\/blog.htm?author=4\"}]}<\/script>\n<!-- \/ Yoast SEO Premium plugin. -->","yoast_head_json":{"title":"Hardware Configuration and Network Design for Large-Scale GPU Clusters - fibermall.com","description":"In this article, we delve into the components and configurations involved in building large-scale GPU clusters.","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm","og_locale":"en_US","og_type":"article","og_title":"Hardware Configuration and Network Design for Large-Scale GPU Clusters","og_description":"Background Since OpenAI introduced ChatGPT, large language models (LLMs) have gained significant attention and rapid development. Many enterprises are","og_url":"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm","og_site_name":"fibermall.com","article_published_time":"2024-07-10T08:56:12+00:00","article_modified_time":"2024-07-10T09:08:00+00:00","og_image":[{"width":1080,"height":544,"url":"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/The-DGX-H100-System-6U.png","type":"image\/png"}],"author":"Mia","twitter_card":"summary_large_image","twitter_misc":{"Written by":"Mia","Est. reading time":"21 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm#article","isPartOf":{"@id":"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm"},"author":{"name":"Mia","@id":"https:\/\/www.fibermall.com\/blog.htm\/#\/schema\/person\/56a37c7b0e1e8c61823cd6ef48e8dce3"},"headline":"Hardware Configuration and Network Design for Large-Scale GPU Clusters","datePublished":"2024-07-10T08:56:12+00:00","dateModified":"2024-07-10T09:08:00+00:00","mainEntityOfPage":{"@id":"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm"},"wordCount":3183,"commentCount":0,"publisher":{"@id":"https:\/\/www.fibermall.com\/blog.htm\/#organization"},"image":{"@id":"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm#primaryimage"},"thumbnailUrl":"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/The-DGX-H100-System-6U.png","articleSection":["Blog","Networking"],"inLanguage":"en-US"},{"@type":"WebPage","@id":"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm","url":"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm","name":"Hardware Configuration and Network Design for Large-Scale GPU Clusters - fibermall.com","isPartOf":{"@id":"https:\/\/www.fibermall.com\/blog.htm\/#website"},"primaryImageOfPage":{"@id":"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm#primaryimage"},"image":{"@id":"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm#primaryimage"},"thumbnailUrl":"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/The-DGX-H100-System-6U.png","datePublished":"2024-07-10T08:56:12+00:00","dateModified":"2024-07-10T09:08:00+00:00","description":"In this article, we delve into the components and configurations involved in building large-scale GPU clusters.","breadcrumb":{"@id":"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm#breadcrumb"},"inLanguage":"en-US","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm"]}]},{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm#primaryimage","url":"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/The-DGX-H100-System-6U.png","contentUrl":"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/07\/The-DGX-H100-System-6U.png","width":1080,"height":544,"caption":"The-DGX-H100-System-6U"},{"@type":"BreadcrumbList","@id":"https:\/\/www.fibermall.com\/blog\/large-scale-gpu-clusters.htm#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/www.fibermall.com\/blog.htm"},{"@type":"ListItem","position":2,"name":"Hardware Configuration and Network Design for Large-Scale GPU Clusters"}]},{"@type":"WebSite","@id":"https:\/\/www.fibermall.com\/blog.htm\/#website","url":"https:\/\/www.fibermall.com\/blog.htm\/","name":"fibermall.com","description":"Optical Communication Expert","publisher":{"@id":"https:\/\/www.fibermall.com\/blog.htm\/#organization"},"potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.fibermall.com\/blog.htm\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"en-US"},{"@type":"Organization","@id":"https:\/\/www.fibermall.com\/blog.htm\/#organization","name":"fibermall.com","url":"https:\/\/www.fibermall.com\/blog.htm\/","logo":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/www.fibermall.com\/blog.htm\/#\/schema\/logo\/image\/","url":"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/12\/cropped-Fiber-Mall-Logo.jpg","contentUrl":"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/12\/cropped-Fiber-Mall-Logo.jpg","width":200,"height":67,"caption":"fibermall.com"},"image":{"@id":"https:\/\/www.fibermall.com\/blog.htm\/#\/schema\/logo\/image\/"}},{"@type":"Person","@id":"https:\/\/www.fibermall.com\/blog.htm\/#\/schema\/person\/56a37c7b0e1e8c61823cd6ef48e8dce3","name":"Mia","image":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/www.fibermall.com\/blog.htm\/#\/schema\/person\/image\/","url":"https:\/\/secure.gravatar.com\/avatar\/2e83831e57346d5d451c45e43aa96d14fc758d8f72f7304810e123b36585382b?s=96&d=mm&r=g","contentUrl":"https:\/\/secure.gravatar.com\/avatar\/2e83831e57346d5d451c45e43aa96d14fc758d8f72f7304810e123b36585382b?s=96&d=mm&r=g","caption":"Mia"},"description":"Optical Transmission Consultant","sameAs":["https:\/\/www.fibermall.com\/"],"url":"https:\/\/www.fibermall.com\/blog.htm?author=4"}]}},"_links":{"self":[{"href":"https:\/\/www.fibermall.com\/blog.htm\/index.php?rest_route=\/wp\/v2\/posts\/10270","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.fibermall.com\/blog.htm\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.fibermall.com\/blog.htm\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.fibermall.com\/blog.htm\/index.php?rest_route=\/wp\/v2\/users\/4"}],"replies":[{"embeddable":true,"href":"https:\/\/www.fibermall.com\/blog.htm\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=10270"}],"version-history":[{"count":8,"href":"https:\/\/www.fibermall.com\/blog.htm\/index.php?rest_route=\/wp\/v2\/posts\/10270\/revisions"}],"predecessor-version":[{"id":10335,"href":"https:\/\/www.fibermall.com\/blog.htm\/index.php?rest_route=\/wp\/v2\/posts\/10270\/revisions\/10335"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.fibermall.com\/blog.htm\/index.php?rest_route=\/wp\/v2\/media\/10298"}],"wp:attachment":[{"href":"https:\/\/www.fibermall.com\/blog.htm\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=10270"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.fibermall.com\/blog.htm\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=10270"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.fibermall.com\/blog.htm\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=10270"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}