{"id":16880,"date":"2025-06-17T07:59:54","date_gmt":"2025-06-17T06:59:54","guid":{"rendered":"https:\/\/ibertronica.es\/blog\/?p=16880"},"modified":"2025-06-17T15:51:17","modified_gmt":"2025-06-17T14:51:17","slug":"la-tecnologia-detras-del-cloud-para-ia","status":"publish","type":"post","link":"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/","title":{"rendered":"La tecnolog\u00eda detr\u00e1s del Cloud para IA"},"content":{"rendered":"<p><span data-contrast=\"auto\">Los modelos de IA generativa actuales \u2014capaces de manejar cientos de miles de millones de par\u00e1metros\u2014 han multiplicado exponencialmente las exigencias de computaci\u00f3n en los centros de datos y en las nubes corporativas que los sustentan (Cloud para IA). Para entrenar e inferir estos modelos sin cuellos de botella es necesario mover terabytes de datos por segundo, disipar m\u00e1s de 100\u202fkW por rack y escalar la capacidad de c\u00e1lculo de forma eficiente. Este reto solo puede afrontarse apoy\u00e1ndose en cuatro pilares tecnol\u00f3gicos bien definidos:<\/span><span data-ccp-props=\"{&quot;335559738&quot;:180,&quot;335559739&quot;:180}\">\u00a0<\/span><\/p>\n<ul>\n<li data-leveltext=\"\uf0b7\" data-font=\"Symbol\" data-listid=\"1001\" data-list-defn-props=\"{&quot;335551671&quot;:0,&quot;335552541&quot;:1,&quot;335559685&quot;:720,&quot;335559991&quot;:360,&quot;469769226&quot;:&quot;Symbol&quot;,&quot;469769242&quot;:[8226],&quot;469777803&quot;:&quot;left&quot;,&quot;469777804&quot;:&quot;\uf0b7&quot;,&quot;469777815&quot;:&quot;multilevel&quot;}\" aria-setsize=\"-1\" data-aria-posinset=\"0\" data-aria-level=\"1\"><b><span data-contrast=\"auto\">GPUs de \u00faltima generaci\u00f3n<\/span><\/b><span data-contrast=\"auto\">, con memoria HBM y enlaces NVLink\/NVSwitch.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:36,&quot;335559739&quot;:36}\">\u00a0<\/span><\/li>\n<li data-leveltext=\"\uf0b7\" data-font=\"Symbol\" data-listid=\"1001\" data-list-defn-props=\"{&quot;335551671&quot;:0,&quot;335552541&quot;:1,&quot;335559685&quot;:720,&quot;335559991&quot;:360,&quot;469769226&quot;:&quot;Symbol&quot;,&quot;469769242&quot;:[8226],&quot;469777803&quot;:&quot;left&quot;,&quot;469777804&quot;:&quot;\uf0b7&quot;,&quot;469777815&quot;:&quot;multilevel&quot;}\" aria-setsize=\"-1\" data-aria-posinset=\"0\" data-aria-level=\"1\"><b><span data-contrast=\"auto\">Servidores ultradensos<\/span><\/b><span data-contrast=\"auto\">, que maximizan el uso de cada unidad de rack.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:36,&quot;335559739&quot;:36}\">\u00a0<\/span><\/li>\n<li data-leveltext=\"\uf0b7\" data-font=\"Symbol\" data-listid=\"1001\" data-list-defn-props=\"{&quot;335551671&quot;:0,&quot;335552541&quot;:1,&quot;335559685&quot;:720,&quot;335559991&quot;:360,&quot;469769226&quot;:&quot;Symbol&quot;,&quot;469769242&quot;:[8226],&quot;469777803&quot;:&quot;left&quot;,&quot;469777804&quot;:&quot;\uf0b7&quot;,&quot;469777815&quot;:&quot;multilevel&quot;}\" aria-setsize=\"-1\" data-aria-posinset=\"0\" data-aria-level=\"1\"><b><span data-contrast=\"auto\">Refrigeraci\u00f3n l\u00edquida avanzada<\/span><\/b><span data-contrast=\"auto\">, capaz de operar con agua tibia para extraer grandes cargas t\u00e9rmicas.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:36,&quot;335559739&quot;:36}\">\u00a0<\/span><\/li>\n<li data-leveltext=\"\uf0b7\" data-font=\"Symbol\" data-listid=\"1001\" data-list-defn-props=\"{&quot;335551671&quot;:0,&quot;335552541&quot;:1,&quot;335559685&quot;:720,&quot;335559991&quot;:360,&quot;469769226&quot;:&quot;Symbol&quot;,&quot;469769242&quot;:[8226],&quot;469777803&quot;:&quot;left&quot;,&quot;469777804&quot;:&quot;\uf0b7&quot;,&quot;469777815&quot;:&quot;multilevel&quot;}\" aria-setsize=\"-1\" data-aria-posinset=\"0\" data-aria-level=\"1\"><b><span data-contrast=\"auto\">Redes de interconexi\u00f3n de alta velocidad<\/span><\/b><span data-contrast=\"auto\">, a partir de 200\u202fGb\/s por nodo.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:36,&quot;335559739&quot;:36}\">\u00a0<\/span><\/li>\n<\/ul>\n<p>Este tipo de <strong data-start=\"621\" data-end=\"638\">cloud para IA<\/strong> requiere una integraci\u00f3n completa entre hardware especializado, refrigeraci\u00f3n eficiente y conectividad de alta velocidad para soportar modelos de nueva generaci\u00f3n.<\/p>\n<p><a onclick=\"javascript:pageTracker._trackPageview('\/downloads\/blog\/wp-content\/uploads\/2025\/06\/Tecnologia-cloud-para-IA-Diagrama.jpg');\"  href=\"https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Tecnologia-cloud-para-IA-Diagrama.jpg\"><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-16883 size-full\" title=\"Tecnolog\u00eda Cloud Para Ia Diagrama\" src=\"https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Tecnologia-cloud-para-IA-Diagrama.jpg\" alt=\"Tecnolog\u00eda Cloud Para Ia Diagrama\" width=\"1024\" height=\"1024\" srcset=\"https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Tecnologia-cloud-para-IA-Diagrama.jpg 1024w, https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Tecnologia-cloud-para-IA-Diagrama-300x300.jpg 300w, https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Tecnologia-cloud-para-IA-Diagrama-150x150.jpg 150w, https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Tecnologia-cloud-para-IA-Diagrama-116x116.jpg 116w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/a><\/p>\n<p><span data-contrast=\"auto\">A lo largo del art\u00edculo desgranamos cada uno de estos pilares, explicando c\u00f3mo se combinan en la pr\u00e1ctica y qu\u00e9 tecnolog\u00edas reales los hacen posibles.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:180,&quot;335559739&quot;:180}\">\u00a0<\/span><\/p>\n<h2 aria-level=\"3\"><span data-contrast=\"none\">GPUs: el motor de la IA<\/span><span data-ccp-props=\"{&quot;134245418&quot;:true,&quot;134245529&quot;:true,&quot;335559738&quot;:160,&quot;335559739&quot;:80}\">\u00a0<\/span><\/h2>\n<p><span data-contrast=\"auto\">Los procesadores gr\u00e1ficos lideran hoy el rendimiento computacional gracias a su capacidad masiva para realizar operaciones en coma flotante (<\/span><b><span data-contrast=\"auto\">FLOPS<\/span><\/b><span data-contrast=\"auto\">) y mover datos a gran velocidad. Una sola GPU moderna ofrece hasta 288\u202fGB de <\/span><b><span data-contrast=\"auto\">memoria HBM3e<\/span><\/b><span data-contrast=\"auto\"> y m\u00e1s de 8\u202fTB\/s de ancho de banda interno, cifras imposibles de alcanzar para cualquier CPU convencional.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:180,&quot;335559739&quot;:180}\">\u00a0<\/span><\/p>\n<p><span data-contrast=\"auto\">Tarjetas como la <\/span><b><span data-contrast=\"auto\">NVIDIA\u202fH100\/H200<\/span><\/b><span data-contrast=\"auto\"> o la nueva <\/span><b><span data-contrast=\"auto\">AMD\u202fInstinct\u202fMI350<\/span><\/b><span data-contrast=\"auto\"> combinan esta memoria con decenas de miles de n\u00facleos y enlaces de alta velocidad como <\/span><b><span data-contrast=\"auto\">NVLink<\/span><\/b><span data-contrast=\"auto\"> y <\/span><b><span data-contrast=\"auto\">NVSwitch<\/span><\/b><span data-contrast=\"auto\">, que permiten que varias GPUs funcionen como si fueran una sola. Esto libera a las CPUs tradicionales para tareas de coordinaci\u00f3n, mientras las GPUs se enfocan en el c\u00e1lculo intensivo.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:180,&quot;335559739&quot;:180}\">\u00a0<\/span><\/p>\n<p><span data-contrast=\"auto\">Fabricantes como ASUS y Gigabyte ya han adaptado sus servidores para aprovechar estas capacidades. Por ejemplo, nodos con hasta ocho GPUs interconectadas por NVSwitch pueden ofrecer m\u00e1s de <\/span><b><span data-contrast=\"auto\">4\u202fPFLOPS<\/span><\/b><span data-contrast=\"auto\"> en precisi\u00f3n FP16, lo que multiplica por 30 la potencia de un rack de servidores CPU convencionales.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:180,&quot;335559739&quot;:180}\">\u00a0<\/span><\/p>\n<p><span data-contrast=\"auto\">Para aprovechar todo este rendimiento, el software debe acompa\u00f1ar: bibliotecas como <\/span><b><span data-contrast=\"auto\">NCCL<\/span><\/b><span data-contrast=\"auto\">, <\/span><b><span data-contrast=\"auto\">GPUDirect\u202fRDMA<\/span><\/b><span data-contrast=\"auto\"> y <\/span><b><span data-contrast=\"auto\">GPUDirect\u202fStorage<\/span><\/b><span data-contrast=\"auto\"> permiten que los datos fluyan entre GPUs, red y almacenamiento sin pasar por la RAM del sistema, reduciendo latencias a nivel de microsegundos.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:180,&quot;335559739&quot;:180}\">\u00a0<\/span><\/p>\n<h2 aria-level=\"3\">Servidores densos para cloud para IA: m\u00e1s potencia en menos espacio<\/h2>\n<p><span data-contrast=\"auto\">La m\u00e9trica clave en un cl\u00faster de IA no es solo la potencia bruta, sino cu\u00e1nta de esa potencia cabe en cada rack (<\/span><b><span data-contrast=\"auto\">FLOPS\/U<\/span><\/b><span data-contrast=\"auto\">). Para lograr densidades extremas, se agrupan servidores con m\u00faltiples GPUs de alto rendimiento, colocados con precisi\u00f3n para maximizar el uso del espacio, refrigeraci\u00f3n y conectividad interna.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:180,&quot;335559739&quot;:180}\">\u00a0<\/span><\/p>\n<p><span data-contrast=\"auto\">Un rack de 42\u202fU bien optimizado puede incluir:<\/span><span data-ccp-props=\"{&quot;335559738&quot;:180,&quot;335559739&quot;:180}\">\u00a0<\/span><\/p>\n<ul>\n<li data-leveltext=\"\uf0b7\" data-font=\"Symbol\" data-listid=\"1002\" data-list-defn-props=\"{&quot;335551671&quot;:0,&quot;335552541&quot;:1,&quot;335559685&quot;:720,&quot;335559991&quot;:360,&quot;469769226&quot;:&quot;Symbol&quot;,&quot;469769242&quot;:[8226],&quot;469777803&quot;:&quot;left&quot;,&quot;469777804&quot;:&quot;\uf0b7&quot;,&quot;469777815&quot;:&quot;multilevel&quot;}\" aria-setsize=\"-1\" data-aria-posinset=\"4\" data-aria-level=\"1\"><b><span data-contrast=\"auto\">4\u20136 servidores de 8\u202fGPUs<\/span><\/b><span data-contrast=\"auto\">, interconectados mediante buses PCIe Gen5 o enlaces NVSwitch.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:36,&quot;335559739&quot;:36}\">\u00a0<\/span><\/li>\n<li data-leveltext=\"\uf0b7\" data-font=\"Symbol\" data-listid=\"1002\" data-list-defn-props=\"{&quot;335551671&quot;:0,&quot;335552541&quot;:1,&quot;335559685&quot;:720,&quot;335559991&quot;:360,&quot;469769226&quot;:&quot;Symbol&quot;,&quot;469769242&quot;:[8226],&quot;469777803&quot;:&quot;left&quot;,&quot;469777804&quot;:&quot;\uf0b7&quot;,&quot;469777815&quot;:&quot;multilevel&quot;}\" aria-setsize=\"-1\" data-aria-posinset=\"4\" data-aria-level=\"1\"><b><span data-contrast=\"auto\">Switches de red y gesti\u00f3n<\/span><\/b><span data-contrast=\"auto\"> (200\u202fGb\/s), montados en la parte superior.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:36,&quot;335559739&quot;:36}\">\u00a0<\/span><\/li>\n<li data-leveltext=\"\uf0b7\" data-font=\"Symbol\" data-listid=\"1002\" data-list-defn-props=\"{&quot;335551671&quot;:0,&quot;335552541&quot;:1,&quot;335559685&quot;:720,&quot;335559991&quot;:360,&quot;469769226&quot;:&quot;Symbol&quot;,&quot;469769242&quot;:[8226],&quot;469777803&quot;:&quot;left&quot;,&quot;469777804&quot;:&quot;\uf0b7&quot;,&quot;469777815&quot;:&quot;multilevel&quot;}\" aria-setsize=\"-1\" data-aria-posinset=\"4\" data-aria-level=\"1\"><b><span data-contrast=\"auto\">M\u00f3dulos de potencia distribuidos<\/span><\/b><span data-contrast=\"auto\"> (busbar a\u202f48\u201354\u202fV\u202fDC, exclusivos de racks\u202fOCPV3) y, seg\u00fan el dise\u00f1o global, tambi\u00e9n m\u00f3dulos de control t\u00e9rmico o racks externos dedicados a refrigeraci\u00f3n compartida.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:36,&quot;335559739&quot;:36}\">\u00a0<\/span><\/li>\n<\/ul>\n<p><span data-contrast=\"auto\">Este tipo de arquitectura puede alcanzar <\/span><b><span data-contrast=\"auto\">20\u201325\u202fPFLOPS FP16 por rack<\/span><\/b><span data-contrast=\"auto\">. Pero el rendimiento final depender\u00e1 tambi\u00e9n del <\/span><i><span data-contrast=\"auto\">software<\/span><\/i><span data-contrast=\"auto\"> de comunicaciones, que debe adaptarse al fabricante de las GPUs:<\/span><span data-ccp-props=\"{&quot;335559738&quot;:180,&quot;335559739&quot;:180}\">\u00a0<\/span><\/p>\n<ul>\n<li data-leveltext=\"\uf0b7\" data-font=\"Symbol\" data-listid=\"1003\" data-list-defn-props=\"{&quot;335551671&quot;:0,&quot;335552541&quot;:1,&quot;335559685&quot;:720,&quot;335559991&quot;:360,&quot;469769226&quot;:&quot;Symbol&quot;,&quot;469769242&quot;:[8226],&quot;469777803&quot;:&quot;left&quot;,&quot;469777804&quot;:&quot;\uf0b7&quot;,&quot;469777815&quot;:&quot;multilevel&quot;}\" aria-setsize=\"-1\" data-aria-posinset=\"7\" data-aria-level=\"1\"><span data-contrast=\"auto\">Con <\/span><b><span data-contrast=\"auto\">NVIDIA<\/span><\/b><span data-contrast=\"auto\">: bibliotecas <\/span><b><span data-contrast=\"auto\">NCCL<\/span><\/b><span data-contrast=\"auto\"> y <\/span><b><span data-contrast=\"auto\">GPUDirect\u202fRDMA<\/span><\/b><span data-contrast=\"auto\">.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:36,&quot;335559739&quot;:36}\">\u00a0<\/span><\/li>\n<li data-leveltext=\"\uf0b7\" data-font=\"Symbol\" data-listid=\"1003\" data-list-defn-props=\"{&quot;335551671&quot;:0,&quot;335552541&quot;:1,&quot;335559685&quot;:720,&quot;335559991&quot;:360,&quot;469769226&quot;:&quot;Symbol&quot;,&quot;469769242&quot;:[8226],&quot;469777803&quot;:&quot;left&quot;,&quot;469777804&quot;:&quot;\uf0b7&quot;,&quot;469777815&quot;:&quot;multilevel&quot;}\" aria-setsize=\"-1\" data-aria-posinset=\"7\" data-aria-level=\"1\"><span data-contrast=\"auto\">Con <\/span><b><span data-contrast=\"auto\">AMD<\/span><\/b><span data-contrast=\"auto\">: se usa <\/span><b><span data-contrast=\"auto\">RCCL<\/span><\/b><span data-contrast=\"auto\"> y <\/span><b><span data-contrast=\"auto\">ROCm\u202fDirect\u202fRDMA<\/span><\/b><span data-contrast=\"auto\">.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:36,&quot;335559739&quot;:36}\">\u00a0<\/span><\/li>\n<li data-leveltext=\"\uf0b7\" data-font=\"Symbol\" data-listid=\"1003\" data-list-defn-props=\"{&quot;335551671&quot;:0,&quot;335552541&quot;:1,&quot;335559685&quot;:720,&quot;335559991&quot;:360,&quot;469769226&quot;:&quot;Symbol&quot;,&quot;469769242&quot;:[8226],&quot;469777803&quot;:&quot;left&quot;,&quot;469777804&quot;:&quot;\uf0b7&quot;,&quot;469777815&quot;:&quot;multilevel&quot;}\" aria-setsize=\"-1\" data-aria-posinset=\"7\" data-aria-level=\"1\"><span data-contrast=\"auto\">Con <\/span><b><span data-contrast=\"auto\">Intel<\/span><\/b><span data-contrast=\"auto\">: bibliotecas <\/span><b><span data-contrast=\"auto\">oneCCL<\/span><\/b><span data-contrast=\"auto\"> con soporte en <\/span><b><span data-contrast=\"auto\">oneAPI\/Level\u202fZero<\/span><\/b><span data-contrast=\"auto\">.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:36,&quot;335559739&quot;:36}\">\u00a0<\/span><\/li>\n<\/ul>\n<p><span data-contrast=\"auto\">Cada ecosistema tiene su propio stack optimizado, por lo que conviene definir la plataforma de hardware antes de elegir el entorno software.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:180,&quot;335559739&quot;:180}\">\u00a0<\/span><\/p>\n<h2 aria-level=\"3\">Refrigeraci\u00f3n en un cloud para IA: \u00bfaire o l\u00edquido?<\/h2>\n<p><span data-contrast=\"auto\">Antes de dise\u00f1ar un cl\u00faster hay que decidir c\u00f3mo disipar los <\/span><b><span data-contrast=\"auto\">20\u201340\u202fkW<\/span><\/b><span data-contrast=\"auto\"> que puede generar un rack de IA. Existen dos enfoques principales:<\/span><span data-ccp-props=\"{&quot;335559738&quot;:180,&quot;335559739&quot;:180}\">\u00a0<\/span><\/p>\n<ul>\n<li data-leveltext=\"\uf0b7\" data-font=\"Symbol\" data-listid=\"1004\" data-list-defn-props=\"{&quot;335551671&quot;:0,&quot;335552541&quot;:1,&quot;335559685&quot;:720,&quot;335559991&quot;:360,&quot;469769226&quot;:&quot;Symbol&quot;,&quot;469769242&quot;:[8226],&quot;469777803&quot;:&quot;left&quot;,&quot;469777804&quot;:&quot;\uf0b7&quot;,&quot;469777815&quot;:&quot;multilevel&quot;}\" aria-setsize=\"-1\" data-aria-posinset=\"10\" data-aria-level=\"1\"><b><span data-contrast=\"auto\">Refrigeraci\u00f3n por aire<\/span><\/b><span data-contrast=\"auto\">: es la m\u00e1s sencilla y se basa en ventiladores de alto caudal. Suele ser suficiente en instalaciones peque\u00f1as (hasta tres racks).<\/span><span data-ccp-props=\"{&quot;335559738&quot;:36,&quot;335559739&quot;:36}\">\u00a0<\/span><\/li>\n<li data-leveltext=\"\uf0b7\" data-font=\"Symbol\" data-listid=\"1004\" data-list-defn-props=\"{&quot;335551671&quot;:0,&quot;335552541&quot;:1,&quot;335559685&quot;:720,&quot;335559991&quot;:360,&quot;469769226&quot;:&quot;Symbol&quot;,&quot;469769242&quot;:[8226],&quot;469777803&quot;:&quot;left&quot;,&quot;469777804&quot;:&quot;\uf0b7&quot;,&quot;469777815&quot;:&quot;multilevel&quot;}\" aria-setsize=\"-1\" data-aria-posinset=\"10\" data-aria-level=\"1\"><b><span data-contrast=\"auto\">Refrigeraci\u00f3n l\u00edquida (LC)<\/span><\/b><span data-contrast=\"auto\">: m\u00e1s eficiente a partir de los 20\u202fkW\/rack. Utiliza agua glicolada para extraer calor directamente desde las GPUs y CPUs mediante <\/span><i><span data-contrast=\"auto\">cold-plates<\/span><\/i><span data-contrast=\"auto\">.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:36,&quot;335559739&quot;:36}\">\u00a0<\/span><\/li>\n<\/ul>\n<p><span data-contrast=\"auto\">Dentro de LC hay dos modalidades:<\/span><span data-ccp-props=\"{&quot;335559738&quot;:180,&quot;335559739&quot;:180}\">\u00a0<\/span><\/p>\n<ul>\n<li data-leveltext=\"\uf0b7\" data-font=\"Symbol\" data-listid=\"1005\" data-list-defn-props=\"{&quot;335551671&quot;:0,&quot;335552541&quot;:1,&quot;335559685&quot;:720,&quot;335559991&quot;:360,&quot;469769226&quot;:&quot;Symbol&quot;,&quot;469769242&quot;:[8226],&quot;469777803&quot;:&quot;left&quot;,&quot;469777804&quot;:&quot;\uf0b7&quot;,&quot;469777815&quot;:&quot;multilevel&quot;}\" aria-setsize=\"-1\" data-aria-posinset=\"12\" data-aria-level=\"1\"><b><span data-contrast=\"auto\">Liquid-to-Air (L2A)<\/span><\/b><span data-contrast=\"auto\">: el agua caliente se enfr\u00eda en un intercambiador dentro del rack, disipando el calor como aire templado.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:36,&quot;335559739&quot;:36}\">\u00a0<\/span><\/li>\n<li data-leveltext=\"\uf0b7\" data-font=\"Symbol\" data-listid=\"1005\" data-list-defn-props=\"{&quot;335551671&quot;:0,&quot;335552541&quot;:1,&quot;335559685&quot;:720,&quot;335559991&quot;:360,&quot;469769226&quot;:&quot;Symbol&quot;,&quot;469769242&quot;:[8226],&quot;469777803&quot;:&quot;left&quot;,&quot;469777804&quot;:&quot;\uf0b7&quot;,&quot;469777815&quot;:&quot;multilevel&quot;}\" aria-setsize=\"-1\" data-aria-posinset=\"12\" data-aria-level=\"1\"><b><span data-contrast=\"auto\">Liquid-to-Liquid (L2L)<\/span><\/b><span data-contrast=\"auto\">: el agua transfiere su calor a un circuito externo, eliminando por completo el calor de la sala t\u00e9cnica.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:36,&quot;335559739&quot;:36}\">\u00a0<\/span><\/li>\n<\/ul>\n<p><span data-contrast=\"auto\">Estas soluciones no son excluyentes. En muchos casos se combinan: por ejemplo, refrigeraci\u00f3n l\u00edquida para GPUs y refrigeraci\u00f3n por aire para discos y componentes secundarios.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:180,&quot;335559739&quot;:180}\">\u00a0<\/span><\/p>\n<p><span data-contrast=\"auto\">La LC puede estar integrada en el propio rack (<\/span><b><span data-contrast=\"auto\">CDU in-rack<\/span><\/b><span data-contrast=\"auto\">) o gestionarse desde un armario externo que da servicio a varios racks (<\/span><b><span data-contrast=\"auto\">CDU remota<\/span><\/b><span data-contrast=\"auto\">).<\/span><span data-ccp-props=\"{&quot;335559738&quot;:180,&quot;335559739&quot;:180}\">\u00a0<\/span><\/p>\n<p><b><span data-contrast=\"auto\">Nota importante:<\/span><\/b><span data-contrast=\"auto\"> la generaci\u00f3n <\/span><b><span data-contrast=\"auto\">Blackwell de NVIDIA (B100\/B200)<\/span><\/b><span data-contrast=\"auto\"> requiere refrigeraci\u00f3n l\u00edquida de forma obligatoria, al superar 1\u202fkW por GPU. Las generaciones anteriores (como H100\/H200) a\u00fan pueden funcionar por aire si se dispone de una sala bien acondicionada.<\/span><\/p>\n<h2 aria-level=\"3\"><span data-contrast=\"none\">Redes de 200\u202fGb\/s: mover datos sin cuellos de botella<\/span><span data-ccp-props=\"{&quot;134245418&quot;:true,&quot;134245529&quot;:true,&quot;335559738&quot;:160,&quot;335559739&quot;:80}\">\u00a0<\/span><\/h2>\n<p><span data-contrast=\"auto\">Para entrenar modelos grandes, cada GPU puede enviar o recibir hasta <\/span><b><span data-contrast=\"auto\">50\u2013100\u202fGB por iteraci\u00f3n<\/span><\/b><span data-contrast=\"auto\">. En un cl\u00faster de 32 nodos, esto genera varios <\/span><b><span data-contrast=\"auto\">terabits por segundo<\/span><\/b><span data-contrast=\"auto\"> de tr\u00e1fico. Para evitar cuellos de botella, cada GPU deber\u00eda tener su <\/span><b><span data-contrast=\"auto\">propia tarjeta de red (NIC)<\/span><\/b><span data-contrast=\"auto\"> a 200\u202fGb\/s, conectada a la red y alineada con su dominio NUMA.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:180,&quot;335559739&quot;:180}\">\u00a0<\/span><\/p>\n<h3><b><span data-contrast=\"auto\">C\u00f3mo se estructura la red del rack en un cloud para IA<\/span><\/b><span data-ccp-props=\"{&quot;335559738&quot;:180,&quot;335559739&quot;:180}\">\u00a0<\/span><\/h3>\n<ol>\n<li data-leveltext=\"%1.\" data-font=\"\" data-listid=\"1006\" data-list-defn-props=\"{&quot;335552541&quot;:0,&quot;335559685&quot;:720,&quot;335559991&quot;:360,&quot;469769242&quot;:[65533,0],&quot;469777803&quot;:&quot;left&quot;,&quot;469777804&quot;:&quot;%1.&quot;,&quot;469777815&quot;:&quot;multilevel&quot;}\" aria-setsize=\"-1\" data-aria-posinset=\"1\" data-aria-level=\"1\"><b><span data-contrast=\"auto\">Conexi\u00f3n GPU\u2013NIC<\/span><\/b><span data-contrast=\"auto\">: cada GPU se conecta a su propia NIC de 200\u202fGb\/s (InfiniBand o Ethernet).<\/span><span data-ccp-props=\"{&quot;335559738&quot;:36,&quot;335559739&quot;:36}\">\u00a0<\/span><\/li>\n<li data-leveltext=\"%1.\" data-font=\"\" data-listid=\"1006\" data-list-defn-props=\"{&quot;335552541&quot;:0,&quot;335559685&quot;:720,&quot;335559991&quot;:360,&quot;469769242&quot;:[65533,0],&quot;469777803&quot;:&quot;left&quot;,&quot;469777804&quot;:&quot;%1.&quot;,&quot;469777815&quot;:&quot;multilevel&quot;}\" aria-setsize=\"-1\" data-aria-posinset=\"1\" data-aria-level=\"1\"><b><span data-contrast=\"auto\">Switch leaf<\/span><\/b><span data-contrast=\"auto\">: las NICs se conectan a switches situados en la parte superior del rack. Estos switches <\/span><b><span data-contrast=\"auto\">leaf<\/span><\/b><span data-contrast=\"auto\"> act\u00faan como punto de entrada a la red del cl\u00faster.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:36,&quot;335559739&quot;:36}\">\u00a0<\/span><\/li>\n<li data-leveltext=\"%1.\" data-font=\"\" data-listid=\"1006\" data-list-defn-props=\"{&quot;335552541&quot;:0,&quot;335559685&quot;:720,&quot;335559991&quot;:360,&quot;469769242&quot;:[65533,0],&quot;469777803&quot;:&quot;left&quot;,&quot;469777804&quot;:&quot;%1.&quot;,&quot;469777815&quot;:&quot;multilevel&quot;}\" aria-setsize=\"-1\" data-aria-posinset=\"1\" data-aria-level=\"1\"><b><span data-contrast=\"auto\">Switch spine<\/span><\/b><span data-contrast=\"auto\">: cada leaf se conecta a varios <\/span><b><span data-contrast=\"auto\">spines<\/span><\/b><span data-contrast=\"auto\">, switches troncales que interconectan los distintos leafs. As\u00ed, cualquier GPU puede comunicarse con otra en m\u00e1ximo dos saltos: GPU \u2192 leaf \u2192 spine \u2192 leaf \u2192 GPU.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:36,&quot;335559739&quot;:36}\">\u00a0<\/span><\/li>\n<li data-leveltext=\"%1.\" data-font=\"\" data-listid=\"1006\" data-list-defn-props=\"{&quot;335552541&quot;:0,&quot;335559685&quot;:720,&quot;335559991&quot;:360,&quot;469769242&quot;:[65533,0],&quot;469777803&quot;:&quot;left&quot;,&quot;469777804&quot;:&quot;%1.&quot;,&quot;469777815&quot;:&quot;multilevel&quot;}\" aria-setsize=\"-1\" data-aria-posinset=\"1\" data-aria-level=\"1\"><b><span data-contrast=\"auto\">Escalado<\/span><\/b><span data-contrast=\"auto\">: esta topolog\u00eda hoja-espina funciona bien hasta 256 nodos. Para despliegues m\u00e1s grandes se pasa a estructuras m\u00e1s complejas como <\/span><b><span data-contrast=\"auto\">dragonfly+<\/span><\/b><span data-contrast=\"auto\">.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:36,&quot;335559739&quot;:36}\">\u00a0<\/span><\/li>\n<\/ol>\n<h3><b><span data-contrast=\"auto\">Regla pr\u00e1ctica<\/span><\/b><span data-contrast=\"auto\">:<\/span><span data-ccp-props=\"{&quot;335559738&quot;:180,&quot;335559739&quot;:180}\">\u00a0<\/span><\/h3>\n<p><span data-contrast=\"auto\">El ancho de banda total de red del rack debe ser similar al ancho de banda de memoria agregado de sus GPUs, para no desaprovechar potencia de c\u00e1lculo.<\/span><span data-ccp-props=\"{&quot;335559685&quot;:480,&quot;335559731&quot;:0,&quot;335559737&quot;:480,&quot;335559738&quot;:100,&quot;335559739&quot;:100}\">\u00a0<\/span><\/p>\n<h4><b><span data-contrast=\"auto\">Ejemplos de hardware (200\u202fGb\/s)<\/span><\/b><span data-contrast=\"auto\">:<\/span><span data-ccp-props=\"{&quot;335559738&quot;:180,&quot;335559739&quot;:180}\">\u00a0<\/span><\/h4>\n<ul>\n<li data-leveltext=\"\uf0b7\" data-font=\"Symbol\" data-listid=\"1007\" data-list-defn-props=\"{&quot;335551671&quot;:0,&quot;335552541&quot;:1,&quot;335559685&quot;:720,&quot;335559991&quot;:360,&quot;469769226&quot;:&quot;Symbol&quot;,&quot;469769242&quot;:[8226],&quot;469777803&quot;:&quot;left&quot;,&quot;469777804&quot;:&quot;\uf0b7&quot;,&quot;469777815&quot;:&quot;multilevel&quot;}\" aria-setsize=\"-1\" data-aria-posinset=\"14\" data-aria-level=\"1\"><span data-contrast=\"auto\">NICs: NVIDIA ConnectX\u20117, AMD Pensando DPU, Intel E810\u2011CQDA2<\/span><span data-ccp-props=\"{&quot;335559738&quot;:36,&quot;335559739&quot;:36}\">\u00a0<\/span><\/li>\n<li data-leveltext=\"\uf0b7\" data-font=\"Symbol\" data-listid=\"1007\" data-list-defn-props=\"{&quot;335551671&quot;:0,&quot;335552541&quot;:1,&quot;335559685&quot;:720,&quot;335559991&quot;:360,&quot;469769226&quot;:&quot;Symbol&quot;,&quot;469769242&quot;:[8226],&quot;469777803&quot;:&quot;left&quot;,&quot;469777804&quot;:&quot;\uf0b7&quot;,&quot;469777815&quot;:&quot;multilevel&quot;}\" aria-setsize=\"-1\" data-aria-posinset=\"14\" data-aria-level=\"1\"><span data-contrast=\"auto\">Switches: NVIDIA InfiniBand QM9700, Broadcom Tomahawk\u202f6 (Ethernet)<\/span><span data-ccp-props=\"{&quot;335559738&quot;:36,&quot;335559739&quot;:36}\">\u00a0<\/span><\/li>\n<\/ul>\n<p><span data-contrast=\"auto\">Estos componentes soportan <\/span><b><span data-contrast=\"auto\">RDMA<\/span><\/b><span data-contrast=\"auto\">, tecnolog\u00eda que permite a las GPUs intercambiar datos directamente sin pasar por la CPU, lo que mejora la eficiencia y reduce la latencia.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:180,&quot;335559739&quot;:180}\">\u00a0<\/span><\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"size-full wp-image-16884 aligncenter\" title=\"Tecnolog\u00eda Cloud Para Ia Red\" src=\"https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Tecnologia-cloud-para-IA-Red.jpg\" alt=\"Tecnolog\u00eda Cloud Para Ia Red\" width=\"800\" height=\"800\" srcset=\"https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Tecnologia-cloud-para-IA-Red.jpg 800w, https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Tecnologia-cloud-para-IA-Red-300x300.jpg 300w, https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Tecnologia-cloud-para-IA-Red-150x150.jpg 150w, https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Tecnologia-cloud-para-IA-Red-116x116.jpg 116w\" sizes=\"auto, (max-width: 800px) 100vw, 800px\" \/><\/p>\n<h3 aria-level=\"3\"><span data-contrast=\"none\">Conclusi\u00f3n: acelerar la IA sin improvisar<\/span><span data-ccp-props=\"{&quot;134245418&quot;:true,&quot;134245529&quot;:true,&quot;335559738&quot;:160,&quot;335559739&quot;:80}\">\u00a0<\/span><\/h3>\n<p>Adoptar IA avanzada implica transformar la infraestructura. No basta con comprar servidores potentes: hay que dise\u00f1ar un cloud para IA coherente, donde c\u00e1lculo, refrigeraci\u00f3n y red trabajen en armon\u00eda.<\/p>\n<p><span data-contrast=\"auto\">Los servidores GPU de ASUS, Gigabyte, ASRock Rack y Supermicro muestran c\u00f3mo el ecosistema ya est\u00e1 preparado para entregar soluciones listas para entrenar modelos de gran escala. Pero desplegarlas con \u00e9xito exige experiencia, precisi\u00f3n y conocimiento pr\u00e1ctico.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:180,&quot;335559739&quot;:180}\">\u00a0<\/span><\/p>\n<p><b><span data-contrast=\"auto\">ViveRack<\/span><\/b><span data-contrast=\"auto\">, la l\u00ednea de soluciones de alto rendimiento comercializada por <\/span><b><span data-contrast=\"auto\">ibertr\u00f3nica<\/span><\/b><span data-contrast=\"auto\">, re\u00fane todos estos elementos en configuraciones probadas: nodos GPU certificados, distribuci\u00f3n de refrigeraci\u00f3n, switching a 200\u202fGb\/s y una arquitectura lista para crecer.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:180,&quot;335559739&quot;:180}\">\u00a0<\/span><\/p>\n<p><span data-contrast=\"auto\">\ud83d\udc49 Si su empresa est\u00e1 evaluando un salto cualitativo en IA, le invitamos a conocer en detalle las configuraciones ViveRack y descubrir c\u00f3mo podemos ayudarle a convertir su CPD en una verdadera f\u00e1brica de modelos.<\/span><span data-ccp-props=\"{&quot;335559738&quot;:180,&quot;335559739&quot;:180}\">\u00a0<\/span><\/p>\n<p><b><span data-contrast=\"auto\">Puede ponerse en contacto con nosotros a trav\u00e9s del <a href=\"https:\/\/ibertronica.es\/contact\">formulario<\/a> disponible en nuestra web para recibir asesoramiento personalizado.<\/span><\/b><span data-ccp-props=\"{&quot;335559738&quot;:180,&quot;335559739&quot;:180}\">\u00a0<\/span><\/p>\n<table style=\"border-collapse: collapse; width: 100%;\">\n<tbody>\n<tr>\n<td style=\"width: 33.3333%; text-align: center;\"><strong>C\u00f3mo Desplegar Tu Cloud Para IA Con Racks VibeRack<\/strong><\/td>\n<td style=\"width: 33.3333%; text-align: center;\"><strong>Nuevos Armarios Ocp V3<\/strong><\/td>\n<td style=\"width: 33.3333%; text-align: center;\"><strong>NVIDIA HGX: Plataforma Abierta que Impulsa la IA y HPC a Gran Escala<\/strong><\/td>\n<\/tr>\n<tr>\n<td style=\"width: 33.3333%;\"><a href=\"https:\/\/ibertronica.es\/blog\/actualidad\/como-desplegar-tu-cloud-para-ia-con-racks-viberack-de-ibertronica\/\"><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-16889 size-full\" title=\"Cabecera Guia Cloud Para Ia\" src=\"https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Cabecera-Guia-Cloud-para-IA.jpg\" alt=\"Cabecera Guia Cloud Para Ia\" width=\"1000\" height=\"571\" srcset=\"https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Cabecera-Guia-Cloud-para-IA.jpg 1000w, https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Cabecera-Guia-Cloud-para-IA-300x171.jpg 300w, https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Cabecera-Guia-Cloud-para-IA-150x86.jpg 150w, https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Cabecera-Guia-Cloud-para-IA-203x116.jpg 203w\" sizes=\"auto, (max-width: 1000px) 100vw, 1000px\" \/><\/a><\/td>\n<td style=\"width: 33.3333%;\"><a href=\"https:\/\/ibertronica.es\/blog\/actualidad\/nuevos-armarios-ocp-v3-cloud-para-ia\/\"><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-16900 size-full\" title=\"Cabecera Nuevos Armarios Ocp V3 2\" src=\"https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Cabecera-Nuevos-Armarios-Ocp-V3-2.jpg\" alt=\"Cabecera Nuevos Armarios Ocp V3 2\" width=\"1000\" height=\"571\" srcset=\"https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Cabecera-Nuevos-Armarios-Ocp-V3-2.jpg 1000w, https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Cabecera-Nuevos-Armarios-Ocp-V3-2-300x171.jpg 300w, https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Cabecera-Nuevos-Armarios-Ocp-V3-2-150x86.jpg 150w, https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Cabecera-Nuevos-Armarios-Ocp-V3-2-203x116.jpg 203w\" sizes=\"auto, (max-width: 1000px) 100vw, 1000px\" \/><\/a><\/td>\n<td style=\"width: 33.3333%;\"><a href=\"https:\/\/ibertronica.es\/blog\/actualidad\/nvidia-hgx-plataforma-abierta-que-impulsa-la-ia-y-hpc-a-gran-escala\/\"><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-16906 size-full\" title=\"Cabecera Nvidia Hgx\" src=\"https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Cabecera-nvidia-hgx.jpg\" alt=\"Cabecera Nvidia Hgx\" width=\"1000\" height=\"571\" srcset=\"https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Cabecera-nvidia-hgx.jpg 1000w, https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Cabecera-nvidia-hgx-300x171.jpg 300w, https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Cabecera-nvidia-hgx-150x86.jpg 150w, https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Cabecera-nvidia-hgx-203x116.jpg 203w\" sizes=\"auto, (max-width: 1000px) 100vw, 1000px\" \/><\/a><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n","protected":false},"excerpt":{"rendered":"<p>Los modelos de IA generativa actuales \u2014capaces de manejar cientos de miles de millones de par\u00e1metros\u2014 han multiplicado exponencialmente las exigencias de computaci\u00f3n en los centros de datos y en las nubes corporativas que los sustentan (Cloud para IA). Para entrenar e inferir estos modelos sin cuellos de botella es necesario mover terabytes de datos por segundo, disipar m\u00e1s de 100\u202fkW por rack y escalar la capacidad de c\u00e1lculo de forma eficiente. Este reto solo puede afrontarse apoy\u00e1ndose en cuatro&hellip;<\/p>\n","protected":false},"author":2,"featured_media":16882,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[341],"tags":[4811],"class_list":["post-16880","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-actualidad","tag-cloud-para-ia","post-has-thumbnail"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v27.1 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>Cloud para IA: GPUs, Red y Refrigeraci\u00f3n<\/title>\n<meta name=\"description\" content=\"Infraestructura cloud para IA con GPUs de alto rendimiento, refrigeraci\u00f3n l\u00edquida y redes de 200 Gb\/s para entrenar modelos avanzados.\" \/>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/\" \/>\n<meta property=\"og:locale\" content=\"es_ES\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Cloud para IA: GPUs, Red y Refrigeraci\u00f3n\" \/>\n<meta property=\"og:description\" content=\"Infraestructura cloud para IA con GPUs de alto rendimiento, refrigeraci\u00f3n l\u00edquida y redes de 200 Gb\/s para entrenar modelos avanzados.\" \/>\n<meta property=\"og:url\" content=\"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/\" \/>\n<meta property=\"og:site_name\" content=\"Blog de tecnolog\u00eda\" \/>\n<meta property=\"article:publisher\" content=\"https:\/\/www.facebook.com\/IbertronicaES\/\" \/>\n<meta property=\"article:published_time\" content=\"2025-06-17T06:59:54+00:00\" \/>\n<meta property=\"article:modified_time\" content=\"2025-06-17T14:51:17+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Cabecera-Tecnologia-cloud-para-IA.jpg\" \/>\n\t<meta property=\"og:image:width\" content=\"1000\" \/>\n\t<meta property=\"og:image:height\" content=\"571\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/jpeg\" \/>\n<meta name=\"author\" content=\"magazine\" \/>\n<meta name=\"twitter:label1\" content=\"Escrito por\" \/>\n\t<meta name=\"twitter:data1\" content=\"magazine\" \/>\n\t<meta name=\"twitter:label2\" content=\"Tiempo de lectura\" \/>\n\t<meta name=\"twitter:data2\" content=\"7 minutos\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/#article\",\"isPartOf\":{\"@id\":\"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/\"},\"author\":{\"name\":\"magazine\",\"@id\":\"https:\/\/ibertronica.es\/blog\/#\/schema\/person\/0375a85dbad874ed61652f20a07ea33f\"},\"headline\":\"La tecnolog\u00eda detr\u00e1s del Cloud para IA\",\"datePublished\":\"2025-06-17T06:59:54+00:00\",\"dateModified\":\"2025-06-17T14:51:17+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/\"},\"wordCount\":1344,\"commentCount\":0,\"publisher\":{\"@id\":\"https:\/\/ibertronica.es\/blog\/#organization\"},\"image\":{\"@id\":\"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/#primaryimage\"},\"thumbnailUrl\":\"https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Cabecera-Tecnologia-cloud-para-IA.jpg\",\"keywords\":[\"cloud para IA\"],\"articleSection\":[\"Actualidad\"],\"inLanguage\":\"es\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/\",\"url\":\"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/\",\"name\":\"Cloud para IA: GPUs, Red y Refrigeraci\u00f3n\",\"isPartOf\":{\"@id\":\"https:\/\/ibertronica.es\/blog\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/#primaryimage\"},\"image\":{\"@id\":\"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/#primaryimage\"},\"thumbnailUrl\":\"https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Cabecera-Tecnologia-cloud-para-IA.jpg\",\"datePublished\":\"2025-06-17T06:59:54+00:00\",\"dateModified\":\"2025-06-17T14:51:17+00:00\",\"description\":\"Infraestructura cloud para IA con GPUs de alto rendimiento, refrigeraci\u00f3n l\u00edquida y redes de 200 Gb\/s para entrenar modelos avanzados.\",\"breadcrumb\":{\"@id\":\"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/#breadcrumb\"},\"inLanguage\":\"es\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"es\",\"@id\":\"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/#primaryimage\",\"url\":\"https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Cabecera-Tecnologia-cloud-para-IA.jpg\",\"contentUrl\":\"https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Cabecera-Tecnologia-cloud-para-IA.jpg\",\"width\":1000,\"height\":571,\"caption\":\"Tecnolog\u00eda Cloud Para Ia\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Portada\",\"item\":\"https:\/\/ibertronica.es\/blog\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"La tecnolog\u00eda detr\u00e1s del Cloud para IA\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/ibertronica.es\/blog\/#website\",\"url\":\"https:\/\/ibertronica.es\/blog\/\",\"name\":\"Blog de tecnolog\u00eda\",\"description\":\"Ibertr\u00f3nica, un blog sobre hardware inform\u00e1tico y servidores para todo tipo de prestaciones\",\"publisher\":{\"@id\":\"https:\/\/ibertronica.es\/blog\/#organization\"},\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/ibertronica.es\/blog\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"es\"},{\"@type\":\"Organization\",\"@id\":\"https:\/\/ibertronica.es\/blog\/#organization\",\"name\":\"Sistemas Ibertr\u00f3nica\",\"url\":\"https:\/\/ibertronica.es\/blog\/\",\"logo\":{\"@type\":\"ImageObject\",\"inLanguage\":\"es\",\"@id\":\"https:\/\/ibertronica.es\/blog\/#\/schema\/logo\/image\/\",\"url\":\"https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2020\/03\/logotipo_web_2019-1.png\",\"contentUrl\":\"https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2020\/03\/logotipo_web_2019-1.png\",\"width\":417,\"height\":45,\"caption\":\"Sistemas Ibertr\u00f3nica\"},\"image\":{\"@id\":\"https:\/\/ibertronica.es\/blog\/#\/schema\/logo\/image\/\"},\"sameAs\":[\"https:\/\/www.facebook.com\/IbertronicaES\/\",\"https:\/\/x.com\/Ibertronica_Es\"]},{\"@type\":\"Person\",\"@id\":\"https:\/\/ibertronica.es\/blog\/#\/schema\/person\/0375a85dbad874ed61652f20a07ea33f\",\"name\":\"magazine\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"es\",\"@id\":\"https:\/\/ibertronica.es\/blog\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/secure.gravatar.com\/avatar\/7d1ae06745040f270e4f310dafae96f9bc6d960c810d653ac06972e979569589?s=96&d=https%3A%2F%2Fwww.ibertronica.es%2Fimages%2Fperfil-azul.jpg&r=g\",\"contentUrl\":\"https:\/\/secure.gravatar.com\/avatar\/7d1ae06745040f270e4f310dafae96f9bc6d960c810d653ac06972e979569589?s=96&d=https%3A%2F%2Fwww.ibertronica.es%2Fimages%2Fperfil-azul.jpg&r=g\",\"caption\":\"magazine\"}}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Cloud para IA: GPUs, Red y Refrigeraci\u00f3n","description":"Infraestructura cloud para IA con GPUs de alto rendimiento, refrigeraci\u00f3n l\u00edquida y redes de 200 Gb\/s para entrenar modelos avanzados.","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:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/","og_locale":"es_ES","og_type":"article","og_title":"Cloud para IA: GPUs, Red y Refrigeraci\u00f3n","og_description":"Infraestructura cloud para IA con GPUs de alto rendimiento, refrigeraci\u00f3n l\u00edquida y redes de 200 Gb\/s para entrenar modelos avanzados.","og_url":"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/","og_site_name":"Blog de tecnolog\u00eda","article_publisher":"https:\/\/www.facebook.com\/IbertronicaES\/","article_published_time":"2025-06-17T06:59:54+00:00","article_modified_time":"2025-06-17T14:51:17+00:00","og_image":[{"width":1000,"height":571,"url":"https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Cabecera-Tecnologia-cloud-para-IA.jpg","type":"image\/jpeg"}],"author":"magazine","twitter_misc":{"Escrito por":"magazine","Tiempo de lectura":"7 minutos"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/#article","isPartOf":{"@id":"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/"},"author":{"name":"magazine","@id":"https:\/\/ibertronica.es\/blog\/#\/schema\/person\/0375a85dbad874ed61652f20a07ea33f"},"headline":"La tecnolog\u00eda detr\u00e1s del Cloud para IA","datePublished":"2025-06-17T06:59:54+00:00","dateModified":"2025-06-17T14:51:17+00:00","mainEntityOfPage":{"@id":"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/"},"wordCount":1344,"commentCount":0,"publisher":{"@id":"https:\/\/ibertronica.es\/blog\/#organization"},"image":{"@id":"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/#primaryimage"},"thumbnailUrl":"https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Cabecera-Tecnologia-cloud-para-IA.jpg","keywords":["cloud para IA"],"articleSection":["Actualidad"],"inLanguage":"es","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/#respond"]}]},{"@type":"WebPage","@id":"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/","url":"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/","name":"Cloud para IA: GPUs, Red y Refrigeraci\u00f3n","isPartOf":{"@id":"https:\/\/ibertronica.es\/blog\/#website"},"primaryImageOfPage":{"@id":"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/#primaryimage"},"image":{"@id":"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/#primaryimage"},"thumbnailUrl":"https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Cabecera-Tecnologia-cloud-para-IA.jpg","datePublished":"2025-06-17T06:59:54+00:00","dateModified":"2025-06-17T14:51:17+00:00","description":"Infraestructura cloud para IA con GPUs de alto rendimiento, refrigeraci\u00f3n l\u00edquida y redes de 200 Gb\/s para entrenar modelos avanzados.","breadcrumb":{"@id":"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/#breadcrumb"},"inLanguage":"es","potentialAction":[{"@type":"ReadAction","target":["https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/"]}]},{"@type":"ImageObject","inLanguage":"es","@id":"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/#primaryimage","url":"https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Cabecera-Tecnologia-cloud-para-IA.jpg","contentUrl":"https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2025\/06\/Cabecera-Tecnologia-cloud-para-IA.jpg","width":1000,"height":571,"caption":"Tecnolog\u00eda Cloud Para Ia"},{"@type":"BreadcrumbList","@id":"https:\/\/ibertronica.es\/blog\/actualidad\/la-tecnologia-detras-del-cloud-para-ia\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Portada","item":"https:\/\/ibertronica.es\/blog\/"},{"@type":"ListItem","position":2,"name":"La tecnolog\u00eda detr\u00e1s del Cloud para IA"}]},{"@type":"WebSite","@id":"https:\/\/ibertronica.es\/blog\/#website","url":"https:\/\/ibertronica.es\/blog\/","name":"Blog de tecnolog\u00eda","description":"Ibertr\u00f3nica, un blog sobre hardware inform\u00e1tico y servidores para todo tipo de prestaciones","publisher":{"@id":"https:\/\/ibertronica.es\/blog\/#organization"},"potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/ibertronica.es\/blog\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"es"},{"@type":"Organization","@id":"https:\/\/ibertronica.es\/blog\/#organization","name":"Sistemas Ibertr\u00f3nica","url":"https:\/\/ibertronica.es\/blog\/","logo":{"@type":"ImageObject","inLanguage":"es","@id":"https:\/\/ibertronica.es\/blog\/#\/schema\/logo\/image\/","url":"https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2020\/03\/logotipo_web_2019-1.png","contentUrl":"https:\/\/ibertronica.es\/blog\/wp-content\/uploads\/2020\/03\/logotipo_web_2019-1.png","width":417,"height":45,"caption":"Sistemas Ibertr\u00f3nica"},"image":{"@id":"https:\/\/ibertronica.es\/blog\/#\/schema\/logo\/image\/"},"sameAs":["https:\/\/www.facebook.com\/IbertronicaES\/","https:\/\/x.com\/Ibertronica_Es"]},{"@type":"Person","@id":"https:\/\/ibertronica.es\/blog\/#\/schema\/person\/0375a85dbad874ed61652f20a07ea33f","name":"magazine","image":{"@type":"ImageObject","inLanguage":"es","@id":"https:\/\/ibertronica.es\/blog\/#\/schema\/person\/image\/","url":"https:\/\/secure.gravatar.com\/avatar\/7d1ae06745040f270e4f310dafae96f9bc6d960c810d653ac06972e979569589?s=96&d=https%3A%2F%2Fwww.ibertronica.es%2Fimages%2Fperfil-azul.jpg&r=g","contentUrl":"https:\/\/secure.gravatar.com\/avatar\/7d1ae06745040f270e4f310dafae96f9bc6d960c810d653ac06972e979569589?s=96&d=https%3A%2F%2Fwww.ibertronica.es%2Fimages%2Fperfil-azul.jpg&r=g","caption":"magazine"}}]}},"_links":{"self":[{"href":"https:\/\/ibertronica.es\/blog\/wp-json\/wp\/v2\/posts\/16880","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ibertronica.es\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ibertronica.es\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ibertronica.es\/blog\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/ibertronica.es\/blog\/wp-json\/wp\/v2\/comments?post=16880"}],"version-history":[{"count":8,"href":"https:\/\/ibertronica.es\/blog\/wp-json\/wp\/v2\/posts\/16880\/revisions"}],"predecessor-version":[{"id":16909,"href":"https:\/\/ibertronica.es\/blog\/wp-json\/wp\/v2\/posts\/16880\/revisions\/16909"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/ibertronica.es\/blog\/wp-json\/wp\/v2\/media\/16882"}],"wp:attachment":[{"href":"https:\/\/ibertronica.es\/blog\/wp-json\/wp\/v2\/media?parent=16880"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ibertronica.es\/blog\/wp-json\/wp\/v2\/categories?post=16880"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ibertronica.es\/blog\/wp-json\/wp\/v2\/tags?post=16880"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}