{"id":801219,"date":"2023-05-29T14:15:00","date_gmt":"2023-05-29T12:15:00","guid":{"rendered":"https:\/\/www.tomshardware.fr\/?p=801219"},"modified":"2023-07-08T01:09:55","modified_gmt":"2023-07-07T23:09:55","slug":"nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip","status":"publish","type":"post","link":"https:\/\/www.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/","title":{"rendered":"NVIDIA DGX GH200 : 256 GPU Hooper, 18432 c\u0153urs CPU, pour 1 exaFLOPS de performance IA"},"content":{"rendered":"\n<p>\u00c0 l\u2019occasion du Computex 2023, Jensen Huang, PDG de NVIDIA, a fait de nombreuses annonces lors d\u2019une pr\u00e9sentation de plus de deux heures (que vous pouvez voir ou revoir ci-dessous). Les principales sont le lancement de la nouvelle plateforme DGX GH200 AI, le d\u00e9but de la production de masse des Grace CPU Superchip et Grace Hopper Superchip, ainsi qu\u2019un <a href=\"https:\/\/www.tomshardware.fr\/nvidia-souhaite-rendre-les-pnj-des-jeux-video-moins-stupides-grace-a-son-ace-for-games\/\">mod\u00e8le AI baptis\u00e9 ACE for Games<\/a>.<\/p>\n\n\n\n<figure class=\"wp-block-embed aligncenter 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=\"responsive-container embed-youtube-com\"><iframe title=\"COMPUTEX 2023 Keynote NVIDIA Keynote\" width=\"500\" height=\"281\" src=\"https:\/\/www.youtube.com\/embed\/fHwmLOYJU_w?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<p>Nous \u00e9voquerons les deux premiers sujets dans cet article&nbsp;; l\u2019ACE for Games sera trait\u00e9 dans une autre actualit\u00e9.<\/p>\n\n\n\n<p><strong>\u00c0 lire &gt; <a href=\"https:\/\/www.tomshardware.fr\/bientot-des-cpu-et-gpu-consommant-600-w-et-700-w-pour-les-centres-de-donnees\/\">Bient\u00f4t des CPU et GPU consommant 600 W et 700 W pour les centres de donn\u00e9es<\/a><\/strong><\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"h-production-de-masse-des-grace-hopper-superchip-et-grace-cpu-superchip\">Production de masse des Grace Hopper Superchip et Grace CPU Superchip<\/h2>\n\n\n\n<p>Pr\u00e9sent\u00e9s en mars 2022, les <a href=\"https:\/\/www.tomshardware.fr\/nvidia-annonce-les-grace-cpu-superchip-et-grace-hopper-superchip\/\">Grace Hopper Superchip et Grace CPU Superchip<\/a> entrent d\u00e9sormais en production de masse. Le processeur Grace est le CPU Arm de NVIDIA&nbsp;; le Grace Hopper Superchip est une puce \u00e0 200 milliards de transistors combinant un processeur Grace \u00e0 72 c\u0153urs, un GPU Hopper H100, 96 Go de HBM3 et 512 Go de LPPD5X.<\/p>\n\n\n\n<p>Le GPU et le GPU communiquent gr\u00e2ce \u00e0 la technologie d&#8217;interconnexion NVIDIA NVLink-C2C. La bande passante totale s\u2019\u00e9l\u00e8ve \u00e0 900 Go\/s, soit une bande passante sept fois sup\u00e9rieure \u00e0 celle de lignes PCIe Gen5 standards.<\/p>\n\n\n\n<p>Ian Buck, vice pr\u00e9sident de l&#8217;informatique acc\u00e9l\u00e9r\u00e9e chez NVIDIA, a d\u00e9clar\u00e9&nbsp;: \u00ab&nbsp;L&#8217;IA g\u00e9n\u00e9rative transforme rapidement les entreprises, d\u00e9bloque de nouvelles opportunit\u00e9s et acc\u00e9l\u00e8re les d\u00e9couvertes dans les domaines de la sant\u00e9, de la finance, des services aux entreprises et dans bien d&#8217;autres secteurs. Avec les superpuces Grace Hopper en pleine production, les fabricants du monde entier fourniront bient\u00f4t l&#8217;infrastructure acc\u00e9l\u00e9r\u00e9e dont les entreprises ont besoin pour construire et d\u00e9ployer des applications d&#8217;IA g\u00e9n\u00e9rative qui exploitent leurs donn\u00e9es propri\u00e9taires uniques.&nbsp;\u00bb<\/p>\n\n\n\n<div class=\"wp-block-group\"><div class=\"wp-block-group__inner-container is-layout-constrained wp-block-group-is-layout-constrained\">\n<p><strong>A lire aussi, les actualit\u00e9s les plus r\u00e9centes au sujet de NVIDIA et ses technologies :<\/strong><\/p>\n\n\n\n    <ul>\n                    <li><a href=\"https:\/\/www.tomshardware.fr\/rtx-5080-et-5090-nvidia-preparerait-un-monstre-de-performances-et-une-petite-deception\/\">RTX 5080 et 5090 : Nvidia pr\u00e9parerait un monstre de performances mais une d\u00e9ception est \u00e0 pr\u00e9voir<\/a><\/li>\n                    <li><a href=\"https:\/\/www.tomshardware.fr\/entre-progres-douteux-et-promesses-bidons-le-cirque-technologique-continue-edito-recap-semaine\/\">Entre progr\u00e8s douteux et promesses bidons, le cirque technologique continue : c&#8217;est le recap&#8217; critique de la semaine\u00a0!<\/a><\/li>\n                    <li><a href=\"https:\/\/www.tomshardware.fr\/nvidia-lorgne-sur-les-pc-ia-et-compte-bien-avoir-sa-part-du-gateau-des-2025\/\">NVIDIA lorgne sur les PC IA et compte bien avoir sa part du g\u00e2teau d\u00e8s 2025<\/a><\/li>\n                    <li><a href=\"https:\/\/www.tomshardware.fr\/nvidia-fait-planter-les-pc-windows-10-vieux-de-16-ans-mais-est-ce-vraiment-un-probleme\/\">NVIDIA fait planter les PC Windows 10 vieux de 16 ans mais est-ce vraiment un probl\u00e8me\u00a0?<\/a><\/li>\n            <\/ul>\n\n<\/div><\/div>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"h-dgx-gh200-ai\">DGX GH200 AI<\/h2>\n\n\n\n<p>La plateforme DGX GH200 AI succ\u00e8de \u00e0 la <a href=\"https:\/\/www.tomshardware.fr\/nvidia-lance-un-gpu-ampere-a100-avec-80-go-de-memoire-hbm2e\/\">plateforme DGX A100<\/a> lanc\u00e9e en 2020, laquelle succ\u00e9dait d\u00e9j\u00e0 \u00e0 la DGX Volta.<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-large\"><a href=\"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/nvidia-dgx-gh200-1.jpg\"><img alt=\"Image 1 : NVIDIA DGX GH200 : 256 GPU Hooper, 18432 c\u0153urs CPU, pour 1 exaFLOPS de performance IA\" decoding=\"async\" width=\"1024\" height=\"576\" src=\"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/nvidia-dgx-gh200-1-1024x576.jpg\" alt=\"\" class=\"wp-image-801222\" srcset=\"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/nvidia-dgx-gh200-1-1024x576.jpg 1024w, https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/nvidia-dgx-gh200-1-300x169.jpg 300w, https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/nvidia-dgx-gh200-1-768x432.jpg 768w, https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/nvidia-dgx-gh200-1.jpg 1245w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/a><figcaption class=\"wp-element-caption\">\u00a9 NVIDIA<\/figcaption><\/figure><\/div>\n\n\n<p>Cette plateforme DGX GH200 AI de l&#8217;entreprise, con\u00e7ue, comme son nom l\u2019indique, pour les charges de travail d&#8217;IA g\u00e9n\u00e9rative, associe jusqu\u2019\u00e0 256 Grace Hopper Superchips (CPU + GPU) et 144 To de m\u00e9moire partag\u00e9e.<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-large\"><a href=\"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/quantiteGPU.jpg\"><img alt=\"Image 2 : NVIDIA DGX GH200 : 256 GPU Hooper, 18432 c\u0153urs CPU, pour 1 exaFLOPS de performance IA\" decoding=\"async\" width=\"1024\" height=\"653\" src=\"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/quantiteGPU-1024x653.jpg\" alt=\"\" class=\"wp-image-801223\" srcset=\"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/quantiteGPU-1024x653.jpg 1024w, https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/quantiteGPU-300x191.jpg 300w, https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/quantiteGPU-768x490.jpg 768w, https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/quantiteGPU.jpg 1032w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/a><figcaption class=\"wp-element-caption\">\u00a9 NVIDIA<\/figcaption><\/figure><\/div>\n\n\n<p>Pour la comparaison, les plateformes DGX A100 proposent jusqu\u2019\u00e0 huit GPU A100, et 320 Go de m\u00e9moire partag\u00e9e. En outre, le syst\u00e8me d&#8217;interconnexion de la g\u00e9n\u00e9ration pr\u00e9c\u00e9dente ne permettait qu&#8217;\u00e0 huit GPU d&#8217;\u00eatre combin\u00e9s avec NVLink comme un seul GPU sans compromettre les performances. D\u2019apr\u00e8s l\u2019entreprise, le nouveau NVIDIA NVLink Switch System multiplie par dix la bande passante GPU-to-GPU et par sept la bande passante CPU-to-GPU par rapport \u00e0 son syst\u00e8me de pr\u00e9c\u00e9dente g\u00e9n\u00e9ration. La soci\u00e9t\u00e9 mentionne \u00e9galement une efficacit\u00e9 \u00e9nerg\u00e9tique des interconnexions cinq fois sup\u00e9rieure.<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><a href=\"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/NVIDIALink.jpg\"><img alt=\"Image 3 : NVIDIA DGX GH200 : 256 GPU Hooper, 18432 c\u0153urs CPU, pour 1 exaFLOPS de performance IA\" decoding=\"async\" width=\"1012\" height=\"588\" src=\"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/NVIDIALink.jpg\" alt=\"\" class=\"wp-image-801224\" srcset=\"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/NVIDIALink.jpg 1012w, https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/NVIDIALink-300x174.jpg 300w, https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/NVIDIALink-768x446.jpg 768w\" sizes=\"(max-width: 1012px) 100vw, 1012px\" \/><\/a><figcaption class=\"wp-element-caption\">\u00a9 NVIDIA<\/figcaption><\/figure><\/div>\n\n\n<p>\u00ab&nbsp;L&#8217;architecture DGX GH200 offre une bande passante NVLink 48 fois sup\u00e9rieure \u00e0 celle de la g\u00e9n\u00e9ration pr\u00e9c\u00e9dente, ce qui permet d&#8217;obtenir la puissance d&#8217;un supercalculateur d&#8217;IA massif avec la simplicit\u00e9 de programmation d&#8217;un seul GPU&nbsp;\u00bb, r\u00e9sume NVIDIA dans son communiqu\u00e9.<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-large\"><a href=\"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/perfNVIDIAGHG200.jpg\"><img alt=\"Image 4 : NVIDIA DGX GH200 : 256 GPU Hooper, 18432 c\u0153urs CPU, pour 1 exaFLOPS de performance IA\" decoding=\"async\" width=\"1024\" height=\"620\" src=\"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/perfNVIDIAGHG200-1024x620.jpg\" alt=\"\" class=\"wp-image-801226\" srcset=\"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/perfNVIDIAGHG200-1024x620.jpg 1024w, https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/perfNVIDIAGHG200-300x182.jpg 300w, https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/perfNVIDIAGHG200-768x465.jpg 768w, https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/perfNVIDIAGHG200.jpg 1035w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/a><figcaption class=\"wp-element-caption\">\u00a9 NVIDIA<\/figcaption><\/figure><\/div>\n\n\n<p>Les NVIDIA DGX GH200 devraient \u00eatre disponibles d&#8217;ici la fin de l&#8217;ann\u00e9e. Parmi les clients cit\u00e9s : Google, Meta et Microsoft. Les caract\u00e9ristiques principales des DGX H200 sont consign\u00e9es dans le tableau ci-dessous.<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><a href=\"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/DGH200specs.jpg\"><img alt=\"Image 5 : NVIDIA DGX GH200 : 256 GPU Hooper, 18432 c\u0153urs CPU, pour 1 exaFLOPS de performance IA\" decoding=\"async\" width=\"933\" height=\"578\" src=\"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/DGH200specs.jpg\" alt=\"\" class=\"wp-image-801227\" srcset=\"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/DGH200specs.jpg 933w, https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/DGH200specs-300x186.jpg 300w, https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/DGH200specs-768x476.jpg 768w\" sizes=\"(max-width: 933px) 100vw, 933px\" \/><\/a><figcaption class=\"wp-element-caption\">\u00a9 NVIDIA<\/figcaption><\/figure><\/div>\n\n\n<h2 class=\"wp-block-heading\" id=\"h-nouveaux-systemes-mgx\">Nouveaux syst\u00e8mes MGX<\/h2>\n\n\n\n<p>Pour \u00eatre complet, sachez que NVIDIA a \u00e9galement annonc\u00e9 de nouveaux syst\u00e8mes MGX en plus des syst\u00e8mes DGX et HGX \u00e9prouv\u00e9s.<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-large\"><a href=\"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/nvidia-mgx.jpg\"><img alt=\"Image 6 : NVIDIA DGX GH200 : 256 GPU Hooper, 18432 c\u0153urs CPU, pour 1 exaFLOPS de performance IA\" decoding=\"async\" width=\"1024\" height=\"576\" src=\"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/nvidia-mgx-1024x576.jpg\" alt=\"\" class=\"wp-image-801228\" srcset=\"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/nvidia-mgx-1024x576.jpg 1024w, https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/nvidia-mgx-300x169.jpg 300w, https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/nvidia-mgx-768x432.jpg 768w, https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/nvidia-mgx-1536x864.jpg 1536w, https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/nvidia-mgx-1320x743.jpg 1320w, https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/nvidia-mgx.jpg 1920w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/a><figcaption class=\"wp-element-caption\">\u00a9 NVIDIA<\/figcaption><\/figure><\/div>\n\n\n<p>MGX fonctionne avec diff\u00e9rents facteurs de forme \u2013 1U, 2U, 4U (refroidissement par air ou par liquide) \u2013 et est compatible avec les g\u00e9n\u00e9rations actuelles et futures du mat\u00e9riel NVIDIA (GPU, y compris les derniers H100, L40, L4&nbsp;; CPU Grace CPU Superchip, GH200 Grace Hopper Superchip, CPU x86&nbsp;; DPU NVIDIA BlueField-3, adaptateurs r\u00e9seau ConnectX-7).<\/p>\n\n\n\n<p>L\u2019entreprise explique que \u00ab&nbsp;le MGX diff\u00e8re du NVIDIA HGX en ce qu&#8217;il offre une compatibilit\u00e9 flexible et multig\u00e9n\u00e9rationnelle avec les produits NVIDIA afin que les constructeurs de syst\u00e8mes puissent r\u00e9utiliser les conceptions existantes et adopter facilement les produits de la prochaine g\u00e9n\u00e9ration sans avoir \u00e0 proc\u00e9der \u00e0 des reconceptions co\u00fbteuses. HGX repose sur une carte de base multi-GPU connect\u00e9e \u00e0 NVLink et con\u00e7ue pour \u00e9voluer afin de cr\u00e9er le nec plus ultra des syst\u00e8mes d&#8217;IA et de HPC.&nbsp;\u00bb<\/p>\n\n\n\n<p><strong>\u00c0 lire &gt; <a href=\"https:\/\/www.tomshardware.fr\/diapo-24-ans-devolution-des-supercalculateurs-puissance-multipliee-par-15-million\/\">29 ans de supercalculateurs, une puissance multipli\u00e9e par 18 millions !<\/a><\/strong><\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"h-plateforme-de-mise-en-reseau-nvidia-spectrum-x\">Plateforme de mise en r\u00e9seau NVIDIA Spectrum-X<\/h2>\n\n\n\n<p>Terminons par quelques mots au sujet de la plateforme de mise en r\u00e9seau NVIDIA Spectrum-X, \u00e9galement pr\u00e9sent\u00e9e \u00e0 l&#8217;occasion du Computex 2023. Elle combine un commutateur Ethernet NVIDIA Spectrum-4 et le DPU NVIDIA BlueField-3. Spectrum-X \u00ab&nbsp;offre les meilleures performances pour l&#8217;IA, l&#8217;apprentissage automatique et le traitement du langage naturel, ainsi que pour diverses applications industrielles&nbsp;\u00bb \u00e0 en croire NVIDIA, qui pr\u00e9sente d\u2019ailleurs Spectrum-X comme la \u00ab&nbsp;premi\u00e8re plateforme de mise en r\u00e9seau Ethernet pour l&#8217;IA au monde&nbsp;\u00bb.<\/p>\n\n\n\n<figure class=\"wp-block-embed aligncenter 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=\"responsive-container embed-youtube-com\"><iframe title=\"NVIDIA Spectrum-X Platform | World\u2019s first Ethernet fabric built for AI\" width=\"500\" height=\"281\" src=\"https:\/\/www.youtube.com\/embed\/nKqfi3q4S5I?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<p class=\"has-small-font-size\"><em>Sources : <a href=\"https:\/\/investor.nvidia.com\/news\/press-release-details\/2023\/NVIDIA-Grace-Hopper-Superchips-Designed-for-Accelerated-Generative-AI-Enter-Full-Production\/default.aspx\" target=\"_blank\" rel=\"noreferrer noopener\">NVIDIA<\/a>, <a href=\"https:\/\/nvidianews.nvidia.com\/news\/nvidia-announces-dgx-gh200-ai-supercomputer\" target=\"_blank\" rel=\"noreferrer noopener\">NVIDIA DGX H200<\/a>, <a href=\"https:\/\/nvidianews.nvidia.com\/news\/nvidia-mgx-server-specification\" target=\"_blank\" rel=\"noreferrer noopener\">NVIDIA MGX<\/a>, <a href=\"https:\/\/www.nvidia.com\/en-us\/networking\/spectrumx\/\" target=\"_blank\" rel=\"noreferrer noopener\">NVIDIA Spectrum-X<\/a><\/em><\/p>\n","protected":false},"excerpt":{"rendered":"<p>L'entreprise a pr\u00e9sent\u00e9 ses DGX H200 AI : pr\u00e8s de 500 fois plus de m\u00e9moire que la g\u00e9n\u00e9ration pr\u00e9c\u00e9dente (NVIDIA DGX A100), et une architecture qui offre une bande passante NVLink 48 fois sup\u00e9rieure.<\/p>","protected":false},"author":87,"featured_media":801229,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[487,4071],"tags":[538],"hubs":[],"class_list":["post-801219","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-actualite","category-processeurs","tag-nvidia"],"acf":{"display_mode":"medium-img","post_show_excerpt":true,"post_show_in_home":false,"post_show_summary":false,"post_source":"","hide_sidebar":false,"hide_advertising":false},"yoast_head":"<!-- This site is optimized with the Yoast SEO Premium plugin v21.5 (Yoast SEO v23.1) - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>NVIDIA DGX GH200 : 256 GPU Hooper, 18432 c\u0153urs CPU, pour 1 exaFLOPS de performance IA<\/title>\n<meta name=\"description\" content=\"L&#039;entreprise a pr\u00e9sent\u00e9 ses DGX H200 AI : pr\u00e8s de 500 fois plus de m\u00e9moire que la g\u00e9n\u00e9ration pr\u00e9c\u00e9dente (NVIDIA DGX A100), et une architecture qui offre une bande passante NVLink 48 fois sup\u00e9rieure.\" \/>\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.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/\" \/>\n<meta property=\"og:locale\" content=\"fr_FR\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"NVIDIA DGX GH200 : 256 GPU Hooper, 18432 c\u0153urs CPU, pour 1 exaFLOPS de performance IA\" \/>\n<meta property=\"og:description\" content=\"L&#039;entreprise a pr\u00e9sent\u00e9 ses DGX H200 AI : pr\u00e8s de 500 fois plus de m\u00e9moire que la g\u00e9n\u00e9ration pr\u00e9c\u00e9dente (NVIDIA DGX A100), et une architecture qui offre une bande passante NVLink 48 fois sup\u00e9rieure.\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/\" \/>\n<meta property=\"og:site_name\" content=\"Tom\u2019s Hardware\" \/>\n<meta property=\"article:publisher\" content=\"https:\/\/www.facebook.com\/TomsHardwareFrance\/\" \/>\n<meta property=\"article:published_time\" content=\"2023-05-29T12:15:00+00:00\" \/>\n<meta property=\"article:modified_time\" content=\"2023-07-07T23:09:55+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/grace-hopper.jpg\" \/>\n\t<meta property=\"og:image:width\" content=\"2000\" \/>\n\t<meta property=\"og:image:height\" content=\"1127\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/jpeg\" \/>\n<meta name=\"author\" content=\"R\u00e9mi Bouvet\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:creator\" content=\"@tomshardware_fr\" \/>\n<meta name=\"twitter:site\" content=\"@tomshardware_fr\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\/\/www.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/#article\",\"isPartOf\":{\"@id\":\"https:\/\/www.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/\"},\"author\":{\"name\":\"R\u00e9mi Bouvet\",\"@id\":\"https:\/\/www.tomshardware.fr\/#\/schema\/person\/5ccbd8cbd7e2cc86da6a517f62b4d005\"},\"headline\":\"NVIDIA DGX GH200 : 256 GPU Hooper, 18432 c\u0153urs CPU, pour 1 exaFLOPS de performance IA\",\"datePublished\":\"2023-05-29T12:15:00+00:00\",\"dateModified\":\"2023-07-07T23:09:55+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\/\/www.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/\"},\"wordCount\":924,\"commentCount\":0,\"publisher\":{\"@id\":\"https:\/\/www.tomshardware.fr\/#organization\"},\"image\":{\"@id\":\"https:\/\/www.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/#primaryimage\"},\"thumbnailUrl\":\"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/grace-hopper.jpg\",\"keywords\":[\"NVIDIA\"],\"articleSection\":[\"Actualit\u00e9\",\"Processeurs\"],\"inLanguage\":\"fr-FR\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\/\/www.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/\",\"url\":\"https:\/\/www.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/\",\"name\":\"NVIDIA DGX GH200 : 256 GPU Hooper, 18432 c\u0153urs CPU, pour 1 exaFLOPS de performance IA\",\"isPartOf\":{\"@id\":\"https:\/\/www.tomshardware.fr\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/www.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/#primaryimage\"},\"image\":{\"@id\":\"https:\/\/www.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/#primaryimage\"},\"thumbnailUrl\":\"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/grace-hopper.jpg\",\"datePublished\":\"2023-05-29T12:15:00+00:00\",\"dateModified\":\"2023-07-07T23:09:55+00:00\",\"description\":\"L'entreprise a pr\u00e9sent\u00e9 ses DGX H200 AI : pr\u00e8s de 500 fois plus de m\u00e9moire que la g\u00e9n\u00e9ration pr\u00e9c\u00e9dente (NVIDIA DGX A100), et une architecture qui offre une bande passante NVLink 48 fois sup\u00e9rieure.\",\"inLanguage\":\"fr-FR\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"fr-FR\",\"@id\":\"https:\/\/www.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/#primaryimage\",\"url\":\"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/grace-hopper.jpg\",\"contentUrl\":\"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/grace-hopper.jpg\",\"width\":2000,\"height\":1127},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.tomshardware.fr\/#website\",\"url\":\"https:\/\/www.tomshardware.fr\/\",\"name\":\"Tom\u2019s Hardware\",\"description\":\"Toute l&#039;info hardware et gaming !\",\"publisher\":{\"@id\":\"https:\/\/www.tomshardware.fr\/#organization\"},\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.tomshardware.fr\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"fr-FR\"},{\"@type\":\"Organization\",\"@id\":\"https:\/\/www.tomshardware.fr\/#organization\",\"name\":\"Tom\u2019s Hardware\",\"url\":\"https:\/\/www.tomshardware.fr\/\",\"logo\":{\"@type\":\"ImageObject\",\"inLanguage\":\"fr-FR\",\"@id\":\"https:\/\/www.tomshardware.fr\/#\/schema\/logo\/image\/\",\"url\":\"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/06\/th.png\",\"contentUrl\":\"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/06\/th.png\",\"width\":1000,\"height\":1000,\"caption\":\"Tom\u2019s Hardware\"},\"image\":{\"@id\":\"https:\/\/www.tomshardware.fr\/#\/schema\/logo\/image\/\"},\"sameAs\":[\"https:\/\/www.facebook.com\/TomsHardwareFrance\/\",\"https:\/\/x.com\/tomshardware_fr\"]},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.tomshardware.fr\/#\/schema\/person\/5ccbd8cbd7e2cc86da6a517f62b4d005\",\"name\":\"R\u00e9mi Bouvet\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"fr-FR\",\"@id\":\"https:\/\/www.tomshardware.fr\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/secure.gravatar.com\/avatar\/a08bb358c2a9fc36265e8f3225c5de08?s=64&d=mm&r=g\",\"contentUrl\":\"https:\/\/secure.gravatar.com\/avatar\/a08bb358c2a9fc36265e8f3225c5de08?s=64&d=mm&r=g\",\"caption\":\"R\u00e9mi Bouvet\"},\"description\":\"Pas du tout convaincu que \\\"c'\u00e9tait mieux avant\\\", je pr\u00e9f\u00e8re m'int\u00e9resser \u00e0 demain plut\u00f4t que fantasmer hier. Alors devant une carte Voodoo, j'aurais certainement la m\u00eame attitude qu'un singe devant une \u0153uvre d'art. Pour finir, je le confesse, je pr\u00e9f\u00e8re Fallout 4 \u00e0 Fallout, Doom 2016 \u00e0 Doom, Civilization VI \u00e0 Civilization, eFootball \u00e0 PE... non, l\u00e0, faut tout de m\u00eame pas d\u00e9conner.\",\"sameAs\":[\"www.linkedin.com\/in\/r\u00e9mi-bouvet-96049a151\"],\"honorificPrefix\":\"Sieur\",\"gender\":\"Homme\",\"knowsLanguage\":[\"Fran\u00e7ais\",\"Anglais\",\"Burgonde\"],\"jobTitle\":\"R\u00e9dacteur\",\"worksFor\":\"Galaxie Media\",\"url\":\"https:\/\/www.tomshardware.fr\/author\/remi\/\"}]}<\/script>\n<!-- \/ Yoast SEO Premium plugin. -->","yoast_head_json":{"title":"NVIDIA DGX GH200 : 256 GPU Hooper, 18432 c\u0153urs CPU, pour 1 exaFLOPS de performance IA","description":"L'entreprise a pr\u00e9sent\u00e9 ses DGX H200 AI : pr\u00e8s de 500 fois plus de m\u00e9moire que la g\u00e9n\u00e9ration pr\u00e9c\u00e9dente (NVIDIA DGX A100), et une architecture qui offre une bande passante NVLink 48 fois sup\u00e9rieure.","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.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/","og_locale":"fr_FR","og_type":"article","og_title":"NVIDIA DGX GH200 : 256 GPU Hooper, 18432 c\u0153urs CPU, pour 1 exaFLOPS de performance IA","og_description":"L'entreprise a pr\u00e9sent\u00e9 ses DGX H200 AI : pr\u00e8s de 500 fois plus de m\u00e9moire que la g\u00e9n\u00e9ration pr\u00e9c\u00e9dente (NVIDIA DGX A100), et une architecture qui offre une bande passante NVLink 48 fois sup\u00e9rieure.","og_url":"https:\/\/www.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/","og_site_name":"Tom\u2019s Hardware","article_publisher":"https:\/\/www.facebook.com\/TomsHardwareFrance\/","article_published_time":"2023-05-29T12:15:00+00:00","article_modified_time":"2023-07-07T23:09:55+00:00","og_image":[{"width":2000,"height":1127,"url":"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/grace-hopper.jpg","type":"image\/jpeg"}],"author":"R\u00e9mi Bouvet","twitter_card":"summary_large_image","twitter_creator":"@tomshardware_fr","twitter_site":"@tomshardware_fr","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/www.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/#article","isPartOf":{"@id":"https:\/\/www.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/"},"author":{"name":"R\u00e9mi Bouvet","@id":"https:\/\/www.tomshardware.fr\/#\/schema\/person\/5ccbd8cbd7e2cc86da6a517f62b4d005"},"headline":"NVIDIA DGX GH200 : 256 GPU Hooper, 18432 c\u0153urs CPU, pour 1 exaFLOPS de performance IA","datePublished":"2023-05-29T12:15:00+00:00","dateModified":"2023-07-07T23:09:55+00:00","mainEntityOfPage":{"@id":"https:\/\/www.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/"},"wordCount":924,"commentCount":0,"publisher":{"@id":"https:\/\/www.tomshardware.fr\/#organization"},"image":{"@id":"https:\/\/www.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/#primaryimage"},"thumbnailUrl":"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/grace-hopper.jpg","keywords":["NVIDIA"],"articleSection":["Actualit\u00e9","Processeurs"],"inLanguage":"fr-FR","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/www.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/#respond"]}]},{"@type":"WebPage","@id":"https:\/\/www.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/","url":"https:\/\/www.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/","name":"NVIDIA DGX GH200 : 256 GPU Hooper, 18432 c\u0153urs CPU, pour 1 exaFLOPS de performance IA","isPartOf":{"@id":"https:\/\/www.tomshardware.fr\/#website"},"primaryImageOfPage":{"@id":"https:\/\/www.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/#primaryimage"},"image":{"@id":"https:\/\/www.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/#primaryimage"},"thumbnailUrl":"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/grace-hopper.jpg","datePublished":"2023-05-29T12:15:00+00:00","dateModified":"2023-07-07T23:09:55+00:00","description":"L'entreprise a pr\u00e9sent\u00e9 ses DGX H200 AI : pr\u00e8s de 500 fois plus de m\u00e9moire que la g\u00e9n\u00e9ration pr\u00e9c\u00e9dente (NVIDIA DGX A100), et une architecture qui offre une bande passante NVLink 48 fois sup\u00e9rieure.","inLanguage":"fr-FR","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/"]}]},{"@type":"ImageObject","inLanguage":"fr-FR","@id":"https:\/\/www.tomshardware.fr\/nvidia-devoile-ses-solutions-dgx-gh200-et-lance-la-production-des-grace-hopper-superchip\/#primaryimage","url":"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/grace-hopper.jpg","contentUrl":"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/05\/grace-hopper.jpg","width":2000,"height":1127},{"@type":"WebSite","@id":"https:\/\/www.tomshardware.fr\/#website","url":"https:\/\/www.tomshardware.fr\/","name":"Tom\u2019s Hardware","description":"Toute l&#039;info hardware et gaming !","publisher":{"@id":"https:\/\/www.tomshardware.fr\/#organization"},"potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.tomshardware.fr\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"fr-FR"},{"@type":"Organization","@id":"https:\/\/www.tomshardware.fr\/#organization","name":"Tom\u2019s Hardware","url":"https:\/\/www.tomshardware.fr\/","logo":{"@type":"ImageObject","inLanguage":"fr-FR","@id":"https:\/\/www.tomshardware.fr\/#\/schema\/logo\/image\/","url":"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/06\/th.png","contentUrl":"https:\/\/www.tomshardware.fr\/content\/uploads\/sites\/3\/2023\/06\/th.png","width":1000,"height":1000,"caption":"Tom\u2019s Hardware"},"image":{"@id":"https:\/\/www.tomshardware.fr\/#\/schema\/logo\/image\/"},"sameAs":["https:\/\/www.facebook.com\/TomsHardwareFrance\/","https:\/\/x.com\/tomshardware_fr"]},{"@type":"Person","@id":"https:\/\/www.tomshardware.fr\/#\/schema\/person\/5ccbd8cbd7e2cc86da6a517f62b4d005","name":"R\u00e9mi Bouvet","image":{"@type":"ImageObject","inLanguage":"fr-FR","@id":"https:\/\/www.tomshardware.fr\/#\/schema\/person\/image\/","url":"https:\/\/secure.gravatar.com\/avatar\/a08bb358c2a9fc36265e8f3225c5de08?s=64&d=mm&r=g","contentUrl":"https:\/\/secure.gravatar.com\/avatar\/a08bb358c2a9fc36265e8f3225c5de08?s=64&d=mm&r=g","caption":"R\u00e9mi Bouvet"},"description":"Pas du tout convaincu que \"c'\u00e9tait mieux avant\", je pr\u00e9f\u00e8re m'int\u00e9resser \u00e0 demain plut\u00f4t que fantasmer hier. Alors devant une carte Voodoo, j'aurais certainement la m\u00eame attitude qu'un singe devant une \u0153uvre d'art. Pour finir, je le confesse, je pr\u00e9f\u00e8re Fallout 4 \u00e0 Fallout, Doom 2016 \u00e0 Doom, Civilization VI \u00e0 Civilization, eFootball \u00e0 PE... non, l\u00e0, faut tout de m\u00eame pas d\u00e9conner.","sameAs":["www.linkedin.com\/in\/r\u00e9mi-bouvet-96049a151"],"honorificPrefix":"Sieur","gender":"Homme","knowsLanguage":["Fran\u00e7ais","Anglais","Burgonde"],"jobTitle":"R\u00e9dacteur","worksFor":"Galaxie Media","url":"https:\/\/www.tomshardware.fr\/author\/remi\/"}]}},"_links":{"self":[{"href":"https:\/\/www.tomshardware.fr\/wp-json\/wp\/v2\/posts\/801219"}],"collection":[{"href":"https:\/\/www.tomshardware.fr\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.tomshardware.fr\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.tomshardware.fr\/wp-json\/wp\/v2\/users\/87"}],"replies":[{"embeddable":true,"href":"https:\/\/www.tomshardware.fr\/wp-json\/wp\/v2\/comments?post=801219"}],"version-history":[{"count":0,"href":"https:\/\/www.tomshardware.fr\/wp-json\/wp\/v2\/posts\/801219\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.tomshardware.fr\/wp-json\/wp\/v2\/media\/801229"}],"wp:attachment":[{"href":"https:\/\/www.tomshardware.fr\/wp-json\/wp\/v2\/media?parent=801219"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.tomshardware.fr\/wp-json\/wp\/v2\/categories?post=801219"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.tomshardware.fr\/wp-json\/wp\/v2\/tags?post=801219"},{"taxonomy":"hubs","embeddable":true,"href":"https:\/\/www.tomshardware.fr\/wp-json\/wp\/v2\/hubs?post=801219"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}