{"id":10079,"date":"2025-06-26T14:17:50","date_gmt":"2025-06-26T14:17:50","guid":{"rendered":"https:\/\/dev-ai.cloudsway.ai\/?page_id=10079"},"modified":"2025-07-11T01:49:42","modified_gmt":"2025-07-11T01:49:42","slug":"networking-service","status":"publish","type":"page","link":"https:\/\/dev-ai.cloudsway.ai\/fr\/networking-service\/","title":{"rendered":"Service de mise en r\u00e9seau"},"content":{"rendered":"
\n\t\t\t\t
\n\t\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t

Service de mise en r\u00e9seau<\/h1>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t

Dites adieu aux goulets d'\u00e9tranglement en mati\u00e8re de communication<\/p>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

D\u00e9veloppement d'un r\u00e9seau de calcul AI d'une dur\u00e9e inf\u00e9rieure \u00e0 la nanoseconde<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\n\t\t\t\t\t\t\n\t\t\t\t\t\t\t\t\tNous contacter<\/span>\n\t\t\t\t\t<\/span>\n\t\t\t\t\t<\/a>\n\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t
\n\t\t\t\t\t
\n\t\t
\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t

Avantages technologiques<\/h2>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

Architecture \u00e0 tr\u00e8s faible latence<\/strong><\/p>

Temps de latence de bout en bout aussi bas que 0,8 \u00b5s (20 fois plus rapide que le TCP\/IP traditionnel)<\/p>

Prise en charge des normes InfiniBand \u00e0 200\/400\/800 Gbps<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\t\t\t\t
\n\t\t\t\t\t\t\t
\n\t\t\t\n\t\t\t\t\t\t<\/span>\n\t\t<\/div>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

Ordonnancement tenant compte de la topologie<\/strong><\/p>

Optimisation dynamique du routage et \u00e9quilibrage automatique du trafic entre les racks L'optimisation en profondeur de la biblioth\u00e8que NCCL se traduit par une acc\u00e9l\u00e9ration de 40% des op\u00e9rations AllReduce.<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\"\"\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t
\n\t\t\t\t\t
\n\t\t
\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t

Prestations de service<\/h2>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

Topologie \u00e0 un seul rack<\/strong><\/p>

32 H100 GPU enti\u00e8rement interconnect\u00e9s via NVSwitch\u00a0
Bande passante de bissection de 900GB\/s

<\/p>

Prise en charge de la mise \u00e0 l'\u00e9chelle de plusieurs clusters<\/strong><\/p>

Routage distribu\u00e9 bas\u00e9 sur SHARP
permettant une extension \u00e0 1 000-10 000 n\u0153uds<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t

\n\t\t\t\t\t
\n\t\t
\n\t\t
\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t

Caract\u00e9ristiques de la solution pour la formation acc\u00e9l\u00e9r\u00e9e de grands mod\u00e8les\u00a0<\/h2>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t
GPUDirect RDMA et optimisation de la topologie NCCL<\/h6>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

R\u00e9sultats :<\/strong> T\u00e2che d'apprentissage du mod\u00e8le de param\u00e8tres 175B
le surd\u00e9bit de communication a \u00e9t\u00e9 r\u00e9duit de 35% \u00e0 8%, le trafic inter-racks a \u00e9t\u00e9 r\u00e9duit de 60%<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t

\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\"\"\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t
\n\t\t\t\t\t
\n\t\t
\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t

Convergence \u00e9quilibr\u00e9e entre le stockage et l'informatique : Les trois piliers\npiliers de la performance optimale<\/h2>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t
\n\t\t
\n\t\t
\n\t\t
\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t

Dimensions<\/p>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t

\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t
    \n\t\t\t\t\t\t\t
  • \n\t\t\t\t\t\t\t\t\t\tLargeur de bande<\/span>\n\t\t\t\t\t\t\t\t\t<\/li>\n\t\t\t\t\t\t\t\t
  • \n\t\t\t\t\t\t\t\t\t\tTemps de latence<\/span>\n\t\t\t\t\t\t\t\t\t<\/li>\n\t\t\t\t\t\t\t\t
  • \n\t\t\t\t\t\t\t\t\t\tUnit\u00e9 d'\u00e9volutivit\u00e9<\/span>\n\t\t\t\t\t\t\t\t\t<\/li>\n\t\t\t\t\t\t<\/ul>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t
    \n\t\t
    \n\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t

    Couche r\u00e9seau<\/p>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t

    \n\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t\t\t
      \n\t\t\t\t\t\t\t
    • \n\t\t\t\t\t\t\t\t\t\t400Gbps<\/span>\n\t\t\t\t\t\t\t\t\t<\/li>\n\t\t\t\t\t\t\t\t
    • \n\t\t\t\t\t\t\t\t\t\t0,6\u03bcs<\/span>\n\t\t\t\t\t\t\t\t\t<\/li>\n\t\t\t\t\t\t\t\t
    • \n\t\t\t\t\t\t\t\t\t\t36-Port Lossless Fabric<\/span>\n\t\t\t\t\t\t\t\t\t<\/li>\n\t\t\t\t\t\t<\/ul>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t
      \n\t\t
      \n\t\t\t\t
      \n\t\t\t\t
      \n\t\t\t\t\t

      Couche de calcul<\/p>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t

      \n\t\t\t\t
      \n\t\t\t\t
      \n\t\t\t\t\t\t\t
        \n\t\t\t\t\t\t\t
      • \n\t\t\t\t\t\t\t\t\t\t1,8 To\/s NVLink<\/span>\n\t\t\t\t\t\t\t\t\t<\/li>\n\t\t\t\t\t\t\t\t
      • \n\t\t\t\t\t\t\t\t\t\t0,3\u03bcs GPU Interconnect<\/span>\n\t\t\t\t\t\t\t\t\t<\/li>\n\t\t\t\t\t\t\t\t
      • \n\t\t\t\t\t\t\t\t\t\t8-GPU Node\u00a0<\/span>\n\t\t\t\t\t\t\t\t\t<\/li>\n\t\t\t\t\t\t<\/ul>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t
        \n\t\t
        \n\t\t\t\t
        \n\t\t\t\t
        \n\t\t\t\t\t

        Couche de stockage<\/p>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t

        \n\t\t\t\t
        \n\t\t\t\t
        \n\t\t\t\t\t\t\t
          \n\t\t\t\t\t\t\t
        • \n\t\t\t\t\t\t\t\t\t\t320GB\/s All-Flash<\/span>\n\t\t\t\t\t\t\t\t\t<\/li>\n\t\t\t\t\t\t\t\t
        • \n\t\t\t\t\t\t\t\t\t\t200\u03bcs Lecture al\u00e9atoire<\/span>\n\t\t\t\t\t\t\t\t\t<\/li>\n\t\t\t\t\t\t\t\t
        • \n\t\t\t\t\t\t\t\t\t\tEspace de noms unifi\u00e9 100PB<\/span>\n\t\t\t\t\t\t\t\t\t<\/li>\n\t\t\t\t\t\t<\/ul>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>","protected":false},"excerpt":{"rendered":"

          Networking Service Say goodbye to communication bottlenecks Developing a sub-nanosecond AI compute network Contact Us Technological Advantages Ultra-Low Latency Architecture End-to-end latency as low as 0.8 \u00b5s (20x faster than traditional TCP\/IP) Supports 200\/400\/800 Gbps InfiniBand standards Topology-aware Scheduling Dynamic routing optimization and automatic inter-rack traffic balancing Deep optimization of the NCCL library results in […]<\/p>","protected":false},"author":2,"featured_media":0,"parent":0,"menu_order":0,"comment_status":"closed","ping_status":"closed","template":"","meta":{"_acf_changed":false,"inline_featured_image":false,"footnotes":""},"class_list":["post-10079","page","type-page","status-publish","hentry"],"acf":[],"_links":{"self":[{"href":"https:\/\/dev-ai.cloudsway.ai\/fr\/wp-json\/wp\/v2\/pages\/10079","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/dev-ai.cloudsway.ai\/fr\/wp-json\/wp\/v2\/pages"}],"about":[{"href":"https:\/\/dev-ai.cloudsway.ai\/fr\/wp-json\/wp\/v2\/types\/page"}],"author":[{"embeddable":true,"href":"https:\/\/dev-ai.cloudsway.ai\/fr\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/dev-ai.cloudsway.ai\/fr\/wp-json\/wp\/v2\/comments?post=10079"}],"version-history":[{"count":0,"href":"https:\/\/dev-ai.cloudsway.ai\/fr\/wp-json\/wp\/v2\/pages\/10079\/revisions"}],"wp:attachment":[{"href":"https:\/\/dev-ai.cloudsway.ai\/fr\/wp-json\/wp\/v2\/media?parent=10079"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}