{"id":7773,"date":"2024-04-08T09:03:21","date_gmt":"2024-04-08T09:03:21","guid":{"rendered":"https:\/\/www.fibermall.com\/blog\/?p=7773"},"modified":"2026-08-13T08:31:24","modified_gmt":"2026-08-13T08:31:24","slug":"nvidia-gb200-analysis-interconnect-architecture-and-future-evolution","status":"publish","type":"post","link":"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm","title":{"rendered":"NVIDIA GB200 Analysis: Interconnect Architecture and Future Evolution"},"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\/nvidia-gb200-interconnect-architecture.htm\/#GB200_Interconnect_Architecture_Analysis\" >GB200 Interconnect Architecture Analysis<\/a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-2\" href=\"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm\/#NVLink_Bandwidth_Calculation\" >NVLink Bandwidth Calculation<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-3\" href=\"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm\/#GB200_NVL72\" >GB200 NVL72<\/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\/nvidia-gb200-interconnect-architecture.htm\/#NVL576\" >NVL576<\/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\/nvidia-gb200-interconnect-architecture.htm\/#Examining_NVL576_from_a_Business_Perspective\" >Examining NVL576 from a Business Perspective<\/a><\/li><\/ul><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-6\" href=\"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm\/#Evolution_of_Interconnect_Systems_The_Story_of_Cisco\" >Evolution of Interconnect Systems: The Story of Cisco<\/a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-7\" href=\"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm\/#Distributed_Architecture_Due_to_ComputeMemory_Bottleneck\" >Distributed Architecture Due to Compute\/Memory Bottleneck<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-8\" href=\"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm\/#The_Emergence_of_Switch_Fabric\" >The Emergence of Switch Fabric<\/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\/nvidia-gb200-interconnect-architecture.htm\/#Multi-Stage_Architecture_and_the_Evolution_of_Optical_Interconnects\" >Multi-Stage Architecture and the Evolution of Optical Interconnects<\/a><\/li><\/ul><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-10\" href=\"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm\/#NVIDIAs_Future_Evolution\" >NVIDIA&#8217;s Future Evolution<\/a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-11\" href=\"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm\/#Optical_Interconnects\" >Optical Interconnects<\/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\/nvidia-gb200-interconnect-architecture.htm\/#Algorithms_and_Special_Hardware_Integration\" >Algorithms and Special Hardware Integration<\/a><\/li><\/ul><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-13\" href=\"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm\/#Conclusion\" >Conclusion<\/a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-14\" href=\"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm\/#Related_Posts\" >Related Posts<\/a><\/li><\/ul><\/li><\/ul><\/nav><\/div>\n<h2 class=\"wp-block-heading\" id=\"h-gb200-interconnect-architecture-analysis\"><span class=\"ez-toc-section\" id=\"GB200_Interconnect_Architecture_Analysis\"><\/span><strong>GB200 Interconnect Architecture Analysis<\/strong><strong><\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-nvlink-bandwidth-calculation\"><span class=\"ez-toc-section\" id=\"NVLink_Bandwidth_Calculation\"><\/span><strong>NVLink Bandwidth Calculation<\/strong><strong><\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>NVIDIA has a lot of confusion in the calculation of NVLink transmission bandwidth and the concepts of SubLink\/Port\/Lane. Typically, the NVLink bandwidth of a single B200 chip is 1.8TB\/s. This is usually calculated using the memory bandwidth algorithm, with the unit being bytes per second (B\/s). However, on the NVLink Switch or IB\/Ethernet switches and network cards, the perspective is from Mellanox, who calculates the network bandwidth in bits per second (b\/s). Let&#8217;s explain the NVLink calculation method in detail. Starting from NVLink 3.0, four differential pairs form a &#8220;sub-link&#8221; (NVIDIA often uses the terms Port\/Link, with a somewhat vague definition). These 4 pairs of differential signal lines contain both the receive and transmit direction signals. When calculating network bandwidth, a 400Gbps interface typically refers to the ability to transmit and receive 400Gbps of data simultaneously.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img fetchpriority=\"high\" decoding=\"async\" width=\"761\" height=\"551\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/400Gbps-interface.png\" alt=\"400Gbps interface\" class=\"wp-image-7780\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/400Gbps-interface.png 761w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/400Gbps-interface-300x217.png 300w\" sizes=\"(max-width: 761px) 100vw, 761px\" \/><\/figure>\n\n\n\n<p>It is composed of a total of 4 pairs of differential signal lines, with 2 pairs each for RX and TX. From the network&#8217;s perspective, it is a unidirectional 400Gbps link, while from the memory bandwidth perspective, it supports a memory access bandwidth of 100GB\/s.<\/p>\n\n\n\n<p><strong>NVLINK 5.0 Interconnect Bandwidth<\/strong><strong><\/strong><\/p>\n\n\n\n<p>The Blackwell generation uses 224G Serdes, with a sub-link transmission rate of 200Gbps * 4 (4 differential pairs) \/ 8 = 100GB\/s, and a single-direction network bandwidth of 400Gbps. The B200 has 18 sub-links, resulting in a bandwidth of 100GB\/s * 18 = 1.8TB\/s, which is equivalent to 9 single-direction 400Gbps interfaces from a network perspective. Similarly, the NVSwitch introduction mentions that Dual 200Gb\/sec SerDes make up a 400Gbp<\/p>\n\n\n\n<p><\/p>\n\n\n\n<p>s Port.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img decoding=\"async\" width=\"1024\" height=\"560\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/nvlink-Switch-chip-2-1024x560.png\" alt=\"nvlink Switch chip\" class=\"wp-image-7781\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/nvlink-Switch-chip-2-1024x560.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/nvlink-Switch-chip-2-300x164.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/nvlink-Switch-chip-2-768x420.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/nvlink-Switch-chip-2.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>For the sake of clarity, we define the following terms:<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img decoding=\"async\" width=\"847\" height=\"392\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/hbm-2.png\" alt=\"hbm\" class=\"wp-image-7782\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/hbm-2.png 847w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/hbm-2-300x139.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/hbm-2-768x355.png 768w\" sizes=\"(max-width: 847px) 100vw, 847px\" \/><\/figure>\n\n\n\n<p>B200 NVLINK bandwidth is 1.8TB\/s, composed of 18 Ports, each at 100GB\/s, made up of four differential pairs, with each Port containing two 224Gbps Serdes (2x224G PAM4 equates to <a href=\"https:\/\/www.fibermall.com\/store-21995-400g-ndr-infiniband.htm\" target=\"_blank\" rel=\"noreferrer noopener\">400Gbps<\/a> single-direction bandwidth per Port).<\/p>\n\n\n\n<p><strong>NVLINK 4.0 Interconnect<\/strong><strong><\/strong><\/p>\n\n\n\n<p>Regarding Hopper, NVLINK 4.0 uses 112G Serdes, with a single differential signal line capable of 100Gbps, resulting in a cumulative single NVLINK Sub-link of 4x100Gbps = 50GB\/s. Hopper products supporting NVLINK 4.0 have 18 sub-links (ports), so a single H100 supports 50GB\/s * 18 = 900GB\/s. A single system with 8 cards can utilize 4 NVSwitches for connectivity, as shown in the image.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"472\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/DGX-H100-1024x472.png\" alt=\"DGX H100\" class=\"wp-image-7783\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/DGX-H100-1024x472.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/DGX-H100-300x138.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/DGX-H100-768x354.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/DGX-H100.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>It is also possible to add a second-level switch to create a 256-card cluster.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"502\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/scale-up-with-nvlink-network-1024x502.png\" alt=\"scale-up with nvlink network\" class=\"wp-image-7784\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/scale-up-with-nvlink-network-1024x502.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/scale-up-with-nvlink-network-300x147.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/scale-up-with-nvlink-network-768x376.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/scale-up-with-nvlink-network.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>The expansion interface uses <a href=\"https:\/\/www.fibermall.com\/store-21951-400g-osfp.htm\" target=\"_blank\" rel=\"noreferrer noopener\">OSFP optical modules<\/a>, which can support 16 differential signal lines, allowing a single OSFP to support 4 NVLINK Ports.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"546\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/a-single-OSFP-to-support-4-NVLINK-Ports-1024x546.png\" alt=\"a single OSFP to support 4 NVLINK Ports\" class=\"wp-image-7785\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/a-single-OSFP-to-support-4-NVLINK-Ports-1024x546.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/a-single-OSFP-to-support-4-NVLINK-Ports-300x160.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/a-single-OSFP-to-support-4-NVLINK-Ports-768x410.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/a-single-OSFP-to-support-4-NVLINK-Ports.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"804\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/osfp-module-pinout-1024x804.png\" alt=\"osfp module pinout\" class=\"wp-image-7786\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/osfp-module-pinout-1024x804.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/osfp-module-pinout-300x235.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/osfp-module-pinout-768x603.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/osfp-module-pinout.png 1073w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>The NVLink Switch in the image contains 32 OSFP optical module connectors, supporting a total of 32 * 4 = 128 NVLINK 4 Ports.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"420\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/128nvlink-4-ports-1024x420.png\" alt=\"128nvlink 4 ports\" class=\"wp-image-7787\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/128nvlink-4-ports-1024x420.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/128nvlink-4-ports-300x123.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/128nvlink-4-ports-768x315.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/128nvlink-4-ports.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-gb200-nvl72\"><span class=\"ez-toc-section\" id=\"GB200_NVL72\"><\/span><strong>GB200 NVL72<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>The GB200 NVL72 system has the following specifications, with the main focus being on the NVLINK interconnect:<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"863\" height=\"663\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/NVLINK-interconnect.png\" alt=\"NVLINK interconnect\" class=\"wp-image-7788\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/NVLINK-interconnect.png 863w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/NVLINK-interconnect-300x230.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/NVLINK-interconnect-768x590.png 768w\" sizes=\"(max-width: 863px) 100vw, 863px\" \/><\/figure>\n\n\n\n<p>Each GB200 contains one 72-core Grace ARM CPU and two Blackwell GPUs.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"828\" height=\"563\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Each-GB200-contains-one-72-core-Grace-ARM-CPU-and-two-Blackwell-GPUs.png\" alt=\"Each GB200 contains one 72-core Grace ARM CPU and two Blackwell GPUs.\" class=\"wp-image-7789\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Each-GB200-contains-one-72-core-Grace-ARM-CPU-and-two-Blackwell-GPUs.png 828w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Each-GB200-contains-one-72-core-Grace-ARM-CPU-and-two-Blackwell-GPUs-300x204.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Each-GB200-contains-one-72-core-Grace-ARM-CPU-and-two-Blackwell-GPUs-768x522.png 768w\" sizes=\"(max-width: 828px) 100vw, 828px\" \/><\/figure>\n\n\n\n<p>The entire system is composed of Compute Trays and Switch Trays. Each Compute Tray contains two GB200 subsystems, totaling 4 Blackwell GPUs.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"897\" height=\"465\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GB200-NVL72.png\" alt=\"GB200 NVL72\" class=\"wp-image-7790\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GB200-NVL72.png 897w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GB200-NVL72-300x156.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GB200-NVL72-768x398.png 768w\" sizes=\"(max-width: 897px) 100vw, 897px\" \/><\/figure>\n\n\n\n<p>Each Switch Tray contains two NVLINK Switch chips, providing a total of 72 * 2 = 144 NVLINK Ports. The internal structure of a single switch chip is shown, with 36 ports on the top and bottom, providing 7.2TB\/s of bandwidth. Using network calculations, this equates to 28.8Tbps of switching capacity, slightly less than the current leading 51.2Tbps switch chip, but this is due to the implementation of SHARP (NVLS) functionality.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"533\" height=\"489\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/SHARP-NVLS-functionality.png\" alt=\"SHARP (NVLS) functionality\" class=\"wp-image-7791\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/SHARP-NVLS-functionality.png 533w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/SHARP-NVLS-functionality-300x275.png 300w\" sizes=\"(max-width: 533px) 100vw, 533px\" \/><\/figure>\n\n\n\n<p>The entire rack supports 18 Compute Trays and 9 Switch Trays, forming the NVL72 architecture with 72 fully interconnected Blackwell chips.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"530\" height=\"839\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/The-entire-rack-supports-18-Compute-Trays-and-9-Switch-Trays.png\" alt=\"The entire rack supports 18 Compute Trays and 9 Switch Trays\" class=\"wp-image-7792\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/The-entire-rack-supports-18-Compute-Trays-and-9-Switch-Trays.png 530w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/The-entire-rack-supports-18-Compute-Trays-and-9-Switch-Trays-190x300.png 190w\" sizes=\"(max-width: 530px) 100vw, 530px\" \/><\/figure>\n\n\n\n<p>Each GB200 subsystem has 2 * 18 = 36 NVLink5 Ports. The external interconnect of the system does not use OSFP optical modules, but rather a direct copper backplane connection, as shown in the diagram.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1000\" height=\"563\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GB200-subsystem.png\" alt=\"GB200 subsystem\" class=\"wp-image-7793\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GB200-subsystem.png 1000w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GB200-subsystem-300x169.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GB200-subsystem-768x432.png 768w\" sizes=\"(max-width: 1000px) 100vw, 1000px\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"576\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/The-external-interconnect-of-the-system-1024x576.png\" alt=\"The external interconnect of the system\" class=\"wp-image-7794\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/The-external-interconnect-of-the-system-1024x576.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/The-external-interconnect-of-the-system-300x169.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/The-external-interconnect-of-the-system-768x432.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/The-external-interconnect-of-the-system.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>The overall NVL72 interconnect topology is as follows:<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"938\" height=\"341\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/overall-NVL72-interconnect-topology.png\" alt=\"overall NVL72 interconnect topology\" class=\"wp-image-7795\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/overall-NVL72-interconnect-topology.png 938w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/overall-NVL72-interconnect-topology-300x109.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/overall-NVL72-interconnect-topology-768x279.png 768w\" sizes=\"(max-width: 938px) 100vw, 938px\" \/><\/figure>\n\n\n\n<p>Each B200 has 18 NVLINK Ports, and there are 18 NVLINK Switch chips across the 9 Switch Trays. Therefore, each B200&#8217;s ports connect to one NVSwitch chip, resulting in a total of 72 Ports per NVSwitch, which is how the NVL72 system fully connects all 72 B200 chips.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-nvl576\"><span class=\"ez-toc-section\" id=\"NVL576\"><\/span><strong>NVL576<\/strong><strong><\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>We notice that in the NVL72 cabinet, all the switches no longer have additional interfaces to form a larger two-tier switch cluster. From the official NVIDIA images, 16 cabinets are arranged in two rows, and although the total is exactly 72 * 8 = 576 cards liquid-cooled cluster, the connection cables between the cards seem to be more through the Scale-Out RDMA network interconnect, rather than the Scale-Up NVLINK network interconnect.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"768\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GB200-NVL72-compute-racks.png\" alt=\"GB200 NVL72 compute racks\" class=\"wp-image-7796\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GB200-NVL72-compute-racks.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GB200-NVL72-compute-racks-300x225.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GB200-NVL72-compute-racks-768x576.png 768w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>For a 32,000-card cluster, it is also through such NVL72 cabinets, one row of 9 cabinets, 4 NVL72 and 5 network cabinets, two rows of 18 cabinets forming a Sub-Pod, and connected through the RDMA Scale-Out network.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"768\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/full-data-center-with-32000-GPUs.png\" alt=\"full data center with 32000 GPUs\" class=\"wp-image-7797\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/full-data-center-with-32000-GPUs.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/full-data-center-with-32000-GPUs-300x225.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/full-data-center-with-32000-GPUs-768x576.png 768w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>Of course, this is not the so-called NVL576. If NVL576 is required, each 72 GB200 needs to be configured with 18 NVSwitches, which would not fit in a single cabinet. We notice that the official has mentioned that there is a single-cabinet version of NVL72, as well as a dual-cabinet version, and in the dual-cabinet version, each Compute Tray only has one GB200 subsystem.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"970\" height=\"272\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/NVIDIA-GB200-NVL36.png\" alt=\"NVIDIA GB200 NVL36\" class=\"wp-image-7798\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/NVIDIA-GB200-NVL36.png 970w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/NVIDIA-GB200-NVL36-300x84.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/NVIDIA-GB200-NVL36-768x215.png 768w\" sizes=\"(max-width: 970px) 100vw, 970px\" \/><\/figure>\n\n\n\n<p>On the other hand, we notice that there are spare copper cable connectors on the NVSwitch, which are likely customized for different copper backplane connections.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"732\" height=\"438\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/different-copper-backplane-connections.png\" alt=\"different copper backplane connections\" class=\"wp-image-7799\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/different-copper-backplane-connections.png 732w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/different-copper-backplane-connections-300x180.png 300w\" sizes=\"(max-width: 732px) 100vw, 732px\" \/><\/figure>\n\n\n\n<p>It is unknown whether these interfaces will have additional OSFP Cages above the copper interconnect backplane for second-tier NVSwitch interconnect, but this method has one advantage: the single-cabinet version is non-scalable, while the dual-cabinet version is scalable, as shown in the image.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"286\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/second-tier-NVSwitch-interconnect-1024x286.png\" alt=\"second-tier NVSwitch interconnect\" class=\"wp-image-7800\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/second-tier-NVSwitch-interconnect-1024x286.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/second-tier-NVSwitch-interconnect-300x84.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/second-tier-NVSwitch-interconnect-768x215.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/second-tier-NVSwitch-interconnect.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>The dual-cabinet version has 18 NVSwitch Trays, which can be interconnected back-to-back to form NVL72. Although the number of switches has doubled, each switch provides 36 uplink ports for future expansion to the 576-card cluster. A single cabinet has a total of 36*2*9 = 648 uplink ports, and 16 cabinets are required to form NVL576, resulting in a total of 648*16 = 10,368 uplink ports, which can be constructed by 9 second-tier switch planes, each with 36 sub-planes, formed by 18 Switch Trays. The interconnect structure of NVL576 is shown below.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"342\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/The-interconnect-structure-of-NVL576-1024x342.png\" alt=\"The interconnect structure of NVL576\" class=\"wp-image-7801\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/The-interconnect-structure-of-NVL576-1024x342.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/The-interconnect-structure-of-NVL576-300x100.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/The-interconnect-structure-of-NVL576-768x257.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/The-interconnect-structure-of-NVL576.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-examining-nvl576-from-a-business-perspective\"><span class=\"ez-toc-section\" id=\"Examining_NVL576_from_a_Business_Perspective\"><\/span><strong>Examining NVL576 from a Business Perspective<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>I am skeptical about whether there are truly customers for a massive single NVLink Scale-Up network like NVL576. Even AWS has only chosen to offer the NVL72 in their cloud services. The main issues are the reliability and scalability challenges of a two-tier network architecture, making NVL576 not an ideal solution due to its high system complexity.<\/p>\n\n\n\n<p>On the other hand, when considering the compute requirements of next-generation large models, the Meta paper &#8220;How to Build Low-cost Networks for Large Language Models (without Sacrificing Performance)?&#8221; discusses this. The paper refers to the NVLink-based Scale-Up network as a &#8220;High Bandwidth Domain (HBD),&#8221; and analyzes the optimal number of cards within the HBD:<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"991\" height=\"447\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/High-Bandwidth-Domain-HBD.png\" alt=\"High Bandwidth Domain (HBD)\" class=\"wp-image-7802\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/High-Bandwidth-Domain-HBD.png 991w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/High-Bandwidth-Domain-HBD-300x135.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/High-Bandwidth-Domain-HBD-768x346.png 768w\" sizes=\"(max-width: 991px) 100vw, 991px\" \/><\/figure>\n\n\n\n<p>For a GPT-1T model, the performance gains are still quite significant when K&gt;36 compared to K=8, but the marginal benefits of scaling from K&gt;72 to K=576 do not justify the increased system complexity. Additionally, as the Scale-Up NVLINK network size grows, the performance benefits of the RDMA bandwidth between HBDs start to diminish. The ultimate balance is to use NVL72 combined with RDMA Scale-Out to construct a 32,000-card cluster.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"253\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Scale-Up-NVLINK-network-size-grows-1024x253.png\" alt=\"Scale-Up NVLINK network size grows\" class=\"wp-image-7803\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Scale-Up-NVLINK-network-size-grows-1024x253.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Scale-Up-NVLINK-network-size-grows-300x74.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Scale-Up-NVLINK-network-size-grows-768x190.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Scale-Up-NVLINK-network-size-grows.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-embed is-type-video is-provider-youtube wp-block-embed-youtube wp-embed-aspect-16-9 wp-has-aspect-ratio\"><div class=\"wp-block-embed__wrapper\">\n<div class=\"ast-oembed-container \" style=\"height: 100%;\"><iframe title=\"Unboxing the NVIDIA ConnectX-8 SuperNIC: 1.6T\/800G InfiniBand Interconnection Domo | FiberMall\" width=\"500\" height=\"281\" src=\"https:\/\/www.youtube.com\/embed\/kktuzospV0c?feature=oembed\" frameborder=\"0\" allow=\"accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share\" referrerpolicy=\"strict-origin-when-cross-origin\" allowfullscreen><\/iframe><\/div>\n<\/div><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"h-evolution-of-interconnect-systems-the-story-of-cisco\"><span class=\"ez-toc-section\" id=\"Evolution_of_Interconnect_Systems_The_Story_of_Cisco\"><\/span><strong>Evolution of Interconnect Systems: The Story of Cisco<\/strong><strong><\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-distributed-architecture-due-to-compute-memory-bottleneck\"><span class=\"ez-toc-section\" id=\"Distributed_Architecture_Due_to_ComputeMemory_Bottleneck\"><\/span><strong>Distributed Architecture Due to Compute\/Memory Bottleneck<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>In the early days, Cisco&#8217;s routers used a single PowerPC processor to perform forwarding. As the internet exploded, the performance bottleneck was caused by memory-intensive operations like routing table lookups. This led to the gradual emergence of approaches like process switching\/CEF, which connected multiple processors via a data bus:<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"500\" height=\"341\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Distributed-Architecture.png\" alt=\"Distributed Architecture\" class=\"wp-image-7804\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Distributed-Architecture.png 500w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Distributed-Architecture-300x205.png 300w\" sizes=\"(max-width: 500px) 100vw, 500px\" \/><\/figure>\n\n\n\n<p>These methods are similar to early NVLINK 1.0 \/ NVLINK 2.0, where chips were directly interconnected via a bus, like the Pascal generation.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"626\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/any-to-any-1024x626.png\" alt=\"any-to-any\" class=\"wp-image-7805\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/any-to-any-1024x626.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/any-to-any-300x183.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/any-to-any-768x469.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/any-to-any.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-the-emergence-of-switch-fabric\"><span class=\"ez-toc-section\" id=\"The_Emergence_of_Switch_Fabric\"><\/span><strong>The Emergence of Switch Fabric<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>In 1995, Nick Mckeown proposed using a CrossBar switch fabric to support higher-scale Gigabit routers in his paper &#8220;Fast Switched Backplane for a Gigabit Switched Router&#8221; &#8211; which later became Cisco&#8217;s high-end 12000 series routers.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"583\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/line-card-1024x583.png\" alt=\"line card\" class=\"wp-image-7806\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/line-card-1024x583.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/line-card-300x171.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/line-card-768x437.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/line-card.png 1042w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>The switch fabrics in these systems are conceptually identical to today&#8217;s NVSwitch and NVSwitch Tray that build NVL8~NVL72 systems. They are all about interconnecting multiple chips to build a larger-scale system when a single chip hits the memory wall.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"690\" height=\"502\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/NVSwitch-Tray-1.png\" alt=\"NVSwitch Tray\" class=\"wp-image-7808\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/NVSwitch-Tray-1.png 690w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/NVSwitch-Tray-1-300x218.png 300w\" sizes=\"(max-width: 690px) 100vw, 690px\" \/><\/figure>\n\n\n\n<p>The single-chassis design of Cisco 12000, with the Switch Fabric in the middle and 9 Switch Trays, is similar to the GB200, where the top and bottom have 8 line card slots each, corresponding to the Compute Trays in GB200.<\/p>\n\n\n\n<p>The core technology here is the VOQ (Virtual Output Queuing) design and iSLIP scheduling algorithm. When the model executes All-to-All, multiple B200s may write to the same B200 simultaneously, causing Head-Of-Line Blocking (HOLB). Humans cleverly add buffers before and after intersections, which are the Input Queue and Output Queue:<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"512\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Input-Queue-and-Output-Queue-1024x512.png\" alt=\"Input Queue and Output Queue\" class=\"wp-image-7809\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Input-Queue-and-Output-Queue-1024x512.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Input-Queue-and-Output-Queue-300x150.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Input-Queue-and-Output-Queue-768x384.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Input-Queue-and-Output-Queue.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>Unfortunately, the Output Queue can maximize bandwidth utilization but requires N*R speed, while the Input Queue can process at R speed but suffers from HOLB. The maximum throughput of an IQ switch limited by HOLB is calculated to be 58.6%.<\/p>\n\n\n\n<p>A simple solution to the IQ HOLB problem is to use Virtual Output Queue (VOQ), where each input port has a queue for each output, eliminating HOLB while maintaining R-speed buffering.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"337\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Virtual-Output-Queue-1024x337.png\" alt=\"Virtual Output Queue\" class=\"wp-image-7810\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Virtual-Output-Queue-1024x337.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Virtual-Output-Queue-300x99.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Virtual-Output-Queue-768x252.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Virtual-Output-Queue.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>Of course, NVIDIA&#8217;s NVLINK uses a credit-based design, and credit distribution arbitration is an area worth in-depth research for domestic GPU startups.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-multi-stage-architecture-and-the-evolution-of-optical-interconnects\"><span class=\"ez-toc-section\" id=\"Multi-Stage_Architecture_and_the_Evolution_of_Optical_Interconnects\"><\/span><strong>Multi-Stage Architecture and the Evolution of Optical Interconnects<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>The NVL576 is similar to Cisco&#8217;s Carrier Routing System (CRS-1) introduced in 2003.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"620\" height=\"505\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Carrier-Routing-System.png\" alt=\"Carrier Routing System\" class=\"wp-image-7811\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Carrier-Routing-System.png 620w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Carrier-Routing-System-300x244.png 300w\" sizes=\"(max-width: 620px) 100vw, 620px\" \/><\/figure>\n\n\n\n<p>At that time, Cisco built a multi-stage switching network system facing the huge bandwidth demand during the internet bubble.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"501\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/multi-stage-switching-network-system-1024x501.png\" alt=\"multi-stage switching network system\" class=\"wp-image-7812\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/multi-stage-switching-network-system-1024x501.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/multi-stage-switching-network-system-300x147.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/multi-stage-switching-network-system-768x375.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/multi-stage-switching-network-system.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>The 3-stage switching network within a single cabinet, constructed using Switch Trays, is equivalent to the current non-scalable GB200 NVL72. The multi-cabinet structure then corresponds to NVL576. Back then, Cisco could expand from a single cabinet with 16 Line Cards to a system with 8 Fabric cabinets + 72 Line Card cabinets, building a large-scale 1152 Line Card cluster. Cisco&#8217;s internal connections also use optical interconnects.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"431\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/a-large-scale-1152-Line-Card-cluster-1024x431.png\" alt=\"a large-scale 1152 Line Card cluster\" class=\"wp-image-7813\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/a-large-scale-1152-Line-Card-cluster-1024x431.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/a-large-scale-1152-Line-Card-cluster-300x126.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/a-large-scale-1152-Line-Card-cluster-768x324.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/a-large-scale-1152-Line-Card-cluster.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>The inter-chassis optical connectors are shown in the image.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"548\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/inter-chassis-optical-connectors-1024x548.png\" alt=\"inter-chassis optical connectors\" class=\"wp-image-7814\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/inter-chassis-optical-connectors-1024x548.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/inter-chassis-optical-connectors-300x161.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/inter-chassis-optical-connectors-768x411.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/inter-chassis-optical-connectors.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>It&#8217;s worth noting that during this time, Bill Dally, now NVIDIA&#8217;s Chief Scientist, founded Avici and used 3D-Torus interconnects to build Terabit-scale routers.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"334\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Terabit-scale-routers-1024x334.png\" alt=\"Terabit-scale routers\" class=\"wp-image-7815\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Terabit-scale-routers-1024x334.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Terabit-scale-routers-300x98.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Terabit-scale-routers-768x250.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Terabit-scale-routers.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>The 3D-Torus interconnect reminds one of Google&#8217;s TPU. Later, Huawei OEM&#8217;d Avici&#8217;s system and branded it as NE5000, before developing their core router product NE5000E. Concurrently, the emergence of Juniper also put significant pressure on Cisco in the core router domain. Perhaps NVIDIA&#8217;s dominance will also face more challenges going forward.<\/p>\n\n\n\n<p>On the other hand, MEMS-based optical switches were also introduced around that era, which seems to have some similarities to Google&#8217;s current use of optical switches.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"516\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/MEMS-based-optical-switches-1024x516.png\" alt=\"MEMS-based optical switches\" class=\"wp-image-7816\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/MEMS-based-optical-switches-1024x516.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/MEMS-based-optical-switches-300x151.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/MEMS-based-optical-switches-768x387.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/MEMS-based-optical-switches.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"h-nvidia-s-future-evolution\"><span class=\"ez-toc-section\" id=\"NVIDIAs_Future_Evolution\"><\/span><strong>NVIDIA&#8217;s Future Evolution<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p>At the 2023 HOTI conference on interconnect systems, Bill Dally gave a keynote titled &#8220;Accelerator Clusters, The New Supercomputer&#8221;, discussing three main topics from the perspective of on-chip networks and interconnect systems:<\/p>\n\n\n\n<p>Topology:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>CLOS\/3D-Torus\/Dragonfly<\/li>\n\n\n\n<li>Routing:<\/li>\n\n\n\n<li>Flow control<\/li>\n<\/ul>\n\n\n\n<p>Different device connections have varying bandwidths and power consumption.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"583\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/The-New-Supercomputer-1024x583.png\" alt=\"The New Supercomputer\" class=\"wp-image-7817\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/The-New-Supercomputer-1024x583.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/The-New-Supercomputer-300x171.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/The-New-Supercomputer-768x437.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/The-New-Supercomputer.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>The challenge is how to organically combine them, considering factors like power, cost, density, and connection distance.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"312\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/The-challenge-is-how-to-organically-combine-them-1024x312.png\" alt=\"The challenge is how to organically combine them\" class=\"wp-image-7818\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/The-challenge-is-how-to-organically-combine-them-1024x312.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/The-challenge-is-how-to-organically-combine-them-300x91.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/The-challenge-is-how-to-organically-combine-them-768x234.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/The-challenge-is-how-to-organically-combine-them.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-optical-interconnects\"><span class=\"ez-toc-section\" id=\"Optical_Interconnects\"><\/span><strong>Optical Interconnects<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>Through these dimensional measurements, Co-Package Optic DWDM becomes a viable choice:<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"540\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Co-Package-Optic-DWDM-1024x540.png\" alt=\"Co-Package Optic DWDM\" class=\"wp-image-7819\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Co-Package-Optic-DWDM-1024x540.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Co-Package-Optic-DWDM-300x158.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Co-Package-Optic-DWDM-768x405.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Co-Package-Optic-DWDM.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>The system concept diagram for building optical interconnects is as follows:<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"481\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/system-concept-1024x481.png\" alt=\"system concept\" class=\"wp-image-7820\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/system-concept-1024x481.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/system-concept-300x141.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/system-concept-768x361.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/system-concept.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>The ultimate goal is to construct a massive-scale optical interconnect system.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"528\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GPU-rack-1024x528.png\" alt=\"GPU rack\" class=\"wp-image-7821\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GPU-rack-1024x528.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GPU-rack-300x155.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GPU-rack-768x396.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GPU-rack.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>In this aspect, you will see that it is almost identical to the multi-chassis CRS-1 system that Cisco built, where the GPU Rack is equivalent to the Cisco LineCard Chassis, and the Switch Rack is equivalent to Cisco&#8217;s Fabric Chassis. They both use optical interconnects and DWDM technology to reduce connection complexity and increase bandwidth.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"445\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Cisco-LineCard-Chassis-1024x445.png\" alt=\"Cisco LineCard Chassis\" class=\"wp-image-7822\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Cisco-LineCard-Chassis-1024x445.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Cisco-LineCard-Chassis-300x130.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Cisco-LineCard-Chassis-768x334.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/Cisco-LineCard-Chassis.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>At the chip architecture level, an Optical Engine is used as a chiplet for interconnection.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"492\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/photonic-connected-gpu-1024x492.png\" alt=\"photonic connected gpu\" class=\"wp-image-7824\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/photonic-connected-gpu-1024x492.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/photonic-connected-gpu-300x144.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/photonic-connected-gpu-768x369.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/photonic-connected-gpu.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>For the interconnect structure, there is a greater inclination towards adopting the Dragonfly topology and utilizing OCS optical switches.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"561\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/dragonfly-1-1024x561.png\" alt=\"dragonfly\" class=\"wp-image-7825\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/dragonfly-1-1024x561.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/dragonfly-1-300x164.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/dragonfly-1-768x421.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/dragonfly-1.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>As for flow control algorithms and congestion control, Bill discussed mechanisms similar to HOMA\/NDP and Adaptive Routing. It is not necessary to be that complex, as we have better MultiPath CC algorithms that do not require any new switch features.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"h-algorithms-and-special-hardware-integration\"><span class=\"ez-toc-section\" id=\"Algorithms_and_Special_Hardware_Integration\"><\/span><strong>Algorithms and Special Hardware Integration<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>On the other hand, Transformer has been around for 7 years, and it is an excellent algorithm that balances compute-bound and memory-bound operators. However, is there an even more sophisticated algorithm in the industry?<\/p>\n\n\n\n<p>Sparse Attention models like Monarch Mixer, and models that do not require Attention like Mamba\/RMKV, as well as algorithms based on category theory, algebraic geometry, and algebraic topology, are being researched. There are also different numerical formats like FP4\/FP6 supported by Blackwell, and potentially Log8 in the future.<\/p>\n\n\n\n<p>Historically, Cisco also relied on algorithms and special hardware to gradually improve single-chip performance and overcome the complexity of interconnect structures. They used algorithms like TreeBitMap for large-scale routing table lookups on regular DRAM.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"442\" height=\"266\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/TreeBitMap.png\" alt=\"TreeBitMap\" class=\"wp-image-7826\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/TreeBitMap.png 442w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/TreeBitMap-300x181.png 300w\" sizes=\"(max-width: 442px) 100vw, 442px\" \/><\/figure>\n\n\n\n<p>With the development of multi-core and on-chip networks, they built high-performance SPP\/QFP\/QFA network processors, and these technologies have resurfaced in AWS Nitro, NVIDIA BlueField, and Intel IPU DPU processors.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"h-conclusion\"><span class=\"ez-toc-section\" id=\"Conclusion\"><\/span><strong>Conclusion<\/strong><strong><\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p>FibeMall analyzed the latest Blackwell GPU&#8217;s interconnect architecture and explored the distributed system construction and interconnect architecture that Cisco and NVIDIA faced when single-chip performance could not keep up with the explosive demand during two technology waves, as discussed in &#8220;NVIDIA&#8217;s Cisco Moment&#8221;. It also analyzed Bill Dally&#8217;s 2023 HOTI keynote, providing a clear view of NVIDIA&#8217;s future development path.<\/p>\n\n\n\n<p>However, we also note that during the internet bubble peak, companies like Juniper and Avici emerged as challengers to Cisco, and NVIDIA also defeated 3Dfx as a challenger in that era before dominating the professional market. Every era has its opportunities, and the winners are not simply those who pile on more resources, but those who innovate through the combination of algorithms and computing power with hardware.<\/p>\n\n\n\n<p>From the challenger&#8217;s perspective, the difficulty of the computing core itself, apart from the CUDA ecosystem, is not that great. Recently, Jim Keller and some HBM players in South Korea and Japan have been active, and whether BUDA+RISC-V+HBM will become a new emerging force is worth watching.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"517\" src=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/new-trend-towards-disaggregated-computing-1024x517.png\" alt=\"new trend towards disaggregated computing\" class=\"wp-image-7827\" style=\"width:800px;height:undefinedpx\" srcset=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/new-trend-towards-disaggregated-computing-1024x517.png 1024w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/new-trend-towards-disaggregated-computing-300x151.png 300w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/new-trend-towards-disaggregated-computing-768x388.png 768w, https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/new-trend-towards-disaggregated-computing.png 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>From the perspective of replacing IB\/NVLINK interconnect systems, Ethernet already has 51.2Tbps switch chips, and communication protocols based on high-speed Ethernet connections to HBM, supporting in-network computing like SHARP, were already designed in NetDAM three years ago.<\/p>\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\/nvidia-gb200-superchip.htm\" class=\"lwrp-list-link\"><span class=\"lwrp-list-link-title-text\">Introduction to NVIDIA GB200 Superchip and Liquid-Cooled Servers and Cabinets<\/span><\/a><\/li><li class=\"lwrp-list-item\"><a href=\"https:\/\/www.fibermall.com\/questions\/same-module-with-different-cables.htm\" class=\"lwrp-list-link\"><span class=\"lwrp-list-link-title-text\">Can the Same Module on the NDR Switch Plug an NDR Cable into One Port and an NDR 200 Splitter Cable into Another Port?<\/span><\/a><\/li><li class=\"lwrp-list-item\"><a href=\"https:\/\/www.fibermall.com\/blog\/edfa-in-dwdm.htm\" class=\"lwrp-list-link\"><span class=\"lwrp-list-link-title-text\">How to use EDFA in the DWDM system?<\/span><\/a><\/li><li class=\"lwrp-list-item\"><a href=\"https:\/\/www.fibermall.com\/blog\/osfp-xd-connector.htm\" class=\"lwrp-list-link\"><span class=\"lwrp-list-link-title-text\">Understanding the OSFP-XD Connector: The Ultimate Guide to Optical Transceiver Modules<\/span><\/a><\/li><li class=\"lwrp-list-item\"><a href=\"https:\/\/www.fibermall.com\/blog\/silicon-photonics-bridging-the-gap.htm\" class=\"lwrp-list-link\"><span class=\"lwrp-list-link-title-text\">Silicon Photonics: Bridging the Gap in Chip Interconnects<\/span><\/a><\/li>                <\/ul>\r\n                        <\/div>\r\n<\/div>","protected":false},"excerpt":{"rendered":"<p>GB200 Interconnect Architecture Analysis NVLink Bandwidth Calculation NVIDIA has a lot of confusion in the calculation of NVLink transmission bandwidth and the concepts of SubLink\/Port\/Lane. Typically, the NVLink bandwidth of a single B200 chip is 1.8TB\/s. This is usually calculated using the memory bandwidth algorithm, with the unit being bytes per second (B\/s). However, on [&hellip;]<\/p>\n","protected":false},"author":5,"featured_media":7829,"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":"default","_threads_share_type":"default","_google_business_share_type":"default","_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-7773","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>NVIDIA GB200: Interconnect Architecture and Evolution | FiberMall<\/title>\n<meta name=\"description\" content=\"The entire system is composed of Compute Trays and Switch Trays. Each Compute Tray contains two GB200 subsystems, totaling 4 Blackwell GPUs.\" \/>\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\/nvidia-gb200-interconnect-architecture.htm\" \/>\n<meta property=\"og:locale\" content=\"en_US\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"NVIDIA GB200 Analysis: Interconnect Architecture and Future Evolution\" \/>\n<meta property=\"og:description\" content=\"GB200 Interconnect Architecture Analysis NVLink Bandwidth Calculation NVIDIA has a lot of confusion in the calculation of NVLink transmission bandwidth\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm\" \/>\n<meta property=\"og:site_name\" content=\"fibermall.com\" \/>\n<meta property=\"article:published_time\" content=\"2024-04-08T09:03:21+00:00\" \/>\n<meta property=\"article:modified_time\" content=\"2026-08-13T08:31:24+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GB200-NVL72-1.png\" \/>\n\t<meta property=\"og:image:width\" content=\"897\" \/>\n\t<meta property=\"og:image:height\" content=\"598\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/png\" \/>\n<meta name=\"author\" content=\"Brian\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"Written by\" \/>\n\t<meta name=\"twitter:data1\" content=\"Brian\" \/>\n\t<meta name=\"twitter:label2\" content=\"Est. reading time\" \/>\n\t<meta name=\"twitter:data2\" content=\"20 minutes\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm#article\",\"isPartOf\":{\"@id\":\"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm\"},\"author\":{\"name\":\"Brian\",\"@id\":\"https:\/\/www.fibermall.com\/blog.htm\/#\/schema\/person\/9421292f8817d668e9e64f94dfe9b901\"},\"headline\":\"NVIDIA GB200 Analysis: Interconnect Architecture and Future Evolution\",\"datePublished\":\"2024-04-08T09:03:21+00:00\",\"dateModified\":\"2026-08-13T08:31:24+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm\"},\"wordCount\":2296,\"commentCount\":0,\"publisher\":{\"@id\":\"https:\/\/www.fibermall.com\/blog.htm\/#organization\"},\"image\":{\"@id\":\"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm#primaryimage\"},\"thumbnailUrl\":\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GB200-NVL72-1.png\",\"articleSection\":[\"Blog\",\"Networking\"],\"inLanguage\":\"en-US\"},{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm\",\"url\":\"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm\",\"name\":\"NVIDIA GB200: Interconnect Architecture and Evolution | FiberMall\",\"isPartOf\":{\"@id\":\"https:\/\/www.fibermall.com\/blog.htm\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm#primaryimage\"},\"image\":{\"@id\":\"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm#primaryimage\"},\"thumbnailUrl\":\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GB200-NVL72-1.png\",\"datePublished\":\"2024-04-08T09:03:21+00:00\",\"dateModified\":\"2026-08-13T08:31:24+00:00\",\"description\":\"The entire system is composed of Compute Trays and Switch Trays. Each Compute Tray contains two GB200 subsystems, totaling 4 Blackwell GPUs.\",\"breadcrumb\":{\"@id\":\"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm#breadcrumb\"},\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm#primaryimage\",\"url\":\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GB200-NVL72-1.png\",\"contentUrl\":\"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GB200-NVL72-1.png\",\"width\":897,\"height\":598,\"caption\":\"GB200 NVL72\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\/\/www.fibermall.com\/blog.htm\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"NVIDIA GB200 Analysis: Interconnect Architecture and Future Evolution\"}]},{\"@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\/9421292f8817d668e9e64f94dfe9b901\",\"name\":\"Brian\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\/\/www.fibermall.com\/blog.htm\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/secure.gravatar.com\/avatar\/c196835c5ddcc99559fcfaa0046ffbd895891b78d40ab46bf34f7a5ecd73dfd9?s=96&d=mm&r=g\",\"contentUrl\":\"https:\/\/secure.gravatar.com\/avatar\/c196835c5ddcc99559fcfaa0046ffbd895891b78d40ab46bf34f7a5ecd73dfd9?s=96&d=mm&r=g\",\"caption\":\"Brian\"},\"description\":\"Optical Network Engineer\",\"sameAs\":[\"https:\/\/www.fibermall.com\"],\"url\":\"https:\/\/www.fibermall.com\/blog.htm?author=5\"}]}<\/script>\n<!-- \/ Yoast SEO Premium plugin. -->","yoast_head_json":{"title":"NVIDIA GB200: Interconnect Architecture and Evolution | FiberMall","description":"The entire system is composed of Compute Trays and Switch Trays. Each Compute Tray contains two GB200 subsystems, totaling 4 Blackwell GPUs.","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\/nvidia-gb200-interconnect-architecture.htm","og_locale":"en_US","og_type":"article","og_title":"NVIDIA GB200 Analysis: Interconnect Architecture and Future Evolution","og_description":"GB200 Interconnect Architecture Analysis NVLink Bandwidth Calculation NVIDIA has a lot of confusion in the calculation of NVLink transmission bandwidth","og_url":"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm","og_site_name":"fibermall.com","article_published_time":"2024-04-08T09:03:21+00:00","article_modified_time":"2026-08-13T08:31:24+00:00","og_image":[{"width":897,"height":598,"url":"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GB200-NVL72-1.png","type":"image\/png"}],"author":"Brian","twitter_card":"summary_large_image","twitter_misc":{"Written by":"Brian","Est. reading time":"20 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm#article","isPartOf":{"@id":"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm"},"author":{"name":"Brian","@id":"https:\/\/www.fibermall.com\/blog.htm\/#\/schema\/person\/9421292f8817d668e9e64f94dfe9b901"},"headline":"NVIDIA GB200 Analysis: Interconnect Architecture and Future Evolution","datePublished":"2024-04-08T09:03:21+00:00","dateModified":"2026-08-13T08:31:24+00:00","mainEntityOfPage":{"@id":"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm"},"wordCount":2296,"commentCount":0,"publisher":{"@id":"https:\/\/www.fibermall.com\/blog.htm\/#organization"},"image":{"@id":"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm#primaryimage"},"thumbnailUrl":"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GB200-NVL72-1.png","articleSection":["Blog","Networking"],"inLanguage":"en-US"},{"@type":"WebPage","@id":"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm","url":"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm","name":"NVIDIA GB200: Interconnect Architecture and Evolution | FiberMall","isPartOf":{"@id":"https:\/\/www.fibermall.com\/blog.htm\/#website"},"primaryImageOfPage":{"@id":"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm#primaryimage"},"image":{"@id":"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm#primaryimage"},"thumbnailUrl":"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GB200-NVL72-1.png","datePublished":"2024-04-08T09:03:21+00:00","dateModified":"2026-08-13T08:31:24+00:00","description":"The entire system is composed of Compute Trays and Switch Trays. Each Compute Tray contains two GB200 subsystems, totaling 4 Blackwell GPUs.","breadcrumb":{"@id":"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm#breadcrumb"},"inLanguage":"en-US","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm"]}]},{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm#primaryimage","url":"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GB200-NVL72-1.png","contentUrl":"https:\/\/www.fibermall.com\/blog\/wp-content\/uploads\/2024\/04\/GB200-NVL72-1.png","width":897,"height":598,"caption":"GB200 NVL72"},{"@type":"BreadcrumbList","@id":"https:\/\/www.fibermall.com\/blog\/nvidia-gb200-interconnect-architecture.htm#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/www.fibermall.com\/blog.htm"},{"@type":"ListItem","position":2,"name":"NVIDIA GB200 Analysis: Interconnect Architecture and Future Evolution"}]},{"@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\/9421292f8817d668e9e64f94dfe9b901","name":"Brian","image":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/www.fibermall.com\/blog.htm\/#\/schema\/person\/image\/","url":"https:\/\/secure.gravatar.com\/avatar\/c196835c5ddcc99559fcfaa0046ffbd895891b78d40ab46bf34f7a5ecd73dfd9?s=96&d=mm&r=g","contentUrl":"https:\/\/secure.gravatar.com\/avatar\/c196835c5ddcc99559fcfaa0046ffbd895891b78d40ab46bf34f7a5ecd73dfd9?s=96&d=mm&r=g","caption":"Brian"},"description":"Optical Network Engineer","sameAs":["https:\/\/www.fibermall.com"],"url":"https:\/\/www.fibermall.com\/blog.htm?author=5"}]}},"_links":{"self":[{"href":"https:\/\/www.fibermall.com\/blog.htm\/index.php?rest_route=\/wp\/v2\/posts\/7773","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\/5"}],"replies":[{"embeddable":true,"href":"https:\/\/www.fibermall.com\/blog.htm\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=7773"}],"version-history":[{"count":6,"href":"https:\/\/www.fibermall.com\/blog.htm\/index.php?rest_route=\/wp\/v2\/posts\/7773\/revisions"}],"predecessor-version":[{"id":20052,"href":"https:\/\/www.fibermall.com\/blog.htm\/index.php?rest_route=\/wp\/v2\/posts\/7773\/revisions\/20052"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.fibermall.com\/blog.htm\/index.php?rest_route=\/wp\/v2\/media\/7829"}],"wp:attachment":[{"href":"https:\/\/www.fibermall.com\/blog.htm\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=7773"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.fibermall.com\/blog.htm\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=7773"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.fibermall.com\/blog.htm\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=7773"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}