{"id":8834,"date":"2025-12-21T23:36:16","date_gmt":"2025-12-21T15:36:16","guid":{"rendered":"https:\/\/www.lxbchip.com\/?p=8834"},"modified":"2025-12-21T23:36:16","modified_gmt":"2025-12-21T15:36:16","slug":"powering-sustainable-data-centers","status":"publish","type":"post","link":"https:\/\/www.lxbchip.com\/tr\/powering-sustainable-data-centers\/","title":{"rendered":"S\u00fcrd\u00fcr\u00fclebilir Veri Merkezlerine G\u00fc\u00e7 Sa\u011flama"},"content":{"rendered":"

Yapay zeka (AI) h\u0131zla geni\u015fliyor ve i\u00e7erik \u00fcretiminden m\u00fc\u015fteri hizmetleri sa\u011flayan \u00e7evrimi\u00e7i sohbet robotlar\u0131na kadar g\u00fcnl\u00fck hayat\u0131m\u0131z\u0131n pek \u00e7ok alan\u0131na n\u00fcfuz ediyor. Bunun arkas\u0131nda, \u00e7o\u011fu g\u00fc\u00e7l\u00fc bilgi i\u015flem altyap\u0131s\u0131 gerektiren veri i\u015flemede b\u00fcy\u00fck bir b\u00fcy\u00fcme var. Yapay zeka modelleri kullan\u0131ma haz\u0131r hale gelmeden \u00f6nce, normalde geli\u015fmi\u015f veri merkezlerinde ger\u00e7ekle\u015ftirilen e\u011fitim ve \u00e7\u0131kar\u0131m i\u015flemlerine ihtiya\u00e7 duyuyor.<\/p>\n

Modern bir veri merkezinin i\u00e7inde genellikle binlerce y\u00fcksek performansl\u0131 sunucu bulunur. Bunlar\u0131n her biri hem i\u015fletim hem de so\u011futma i\u00e7in \u00f6nemli miktarda enerji gerektirir. Yayg\u0131n h\u0131zl\u0131 b\u00fcy\u00fcmeyle birlikte, yapay zeka veri merkezi sekt\u00f6r\u00fcndeki enerji t\u00fcketimi balon gibi artmakta, bu da s\u00fcrd\u00fcr\u00fclebilirlik ve bu teknolojik devrimin \u00e7evresel etkileri konusunda endi\u015felere yol a\u00e7maktad\u0131r.<\/p>\n

Uluslararas\u0131 Enerji Ajans\u0131 (IEA), veri merkezlerinin 2024 y\u0131l\u0131nda toplam elektrik talebinin 1,5%'sini (yakla\u015f\u0131k 415 terawatt-saat (TWh)) olu\u015fturdu\u011funu tahmin etmektedir. T\u00fcketimin 2030 y\u0131l\u0131na kadar iki kattan fazla artarak yakla\u015f\u0131k 945 TWh'ye ula\u015fmas\u0131 ve yakla\u015f\u0131k 3% paya y\u00fckselmesi bekleniyor. Veri merkezleri genellikle bir arada k\u00fcmelendi\u011finden, elektrik \u015febekesi alanlar\u0131 \u00fczerinde \u00f6nemli bir bask\u0131 olu\u015fabilir.<\/p>\n

Gelece\u011fin M\u00fchendisli\u011fi: Veri Merkezi G\u00fc\u00e7 Mimarisinin Yapay Zeka i\u00e7in Uyarlanmas\u0131<\/strong><\/p>\n

Arama gibi tipik web kullan\u0131m\u0131yla kar\u015f\u0131la\u015ft\u0131r\u0131ld\u0131\u011f\u0131nda, yapay zeka i\u00e7in gereken enerji \u00e7ok daha y\u00fcksektir, genellikle on kat daha fazlad\u0131r. Bunun ba\u015fl\u0131ca nedeni, her biri y\u00fczlerce Watt t\u00fcketebilen g\u00fc\u00e7l\u00fc grafik i\u015flem birimleridir (GPU'lar). OpenAI'ye g\u00f6re, e\u011fitim modellerinin g\u00fc\u00e7 t\u00fcketimi \u00f6zellikle y\u00fcksektir - \u00f6rnek olarak GPT-4'\u00fcn e\u011fitimi 3 ay boyunca 25.000 NVIDIA A100 GPU gerektirir, 50 gigawatt-saat (GWh) enerji t\u00fcketir ve $100 milyona mal olur.<\/p>\n

Yava\u015flamak bir yana, yapay zeka her alt\u0131 ayda bir g\u00fc\u00e7 t\u00fcketimini ikiye katl\u0131yor ve sekt\u00f6r k\u00fc\u00e7\u00fck bir ulus kadar enerji t\u00fcketiyor. Bu \u00f6l\u00e7ekte, kay\u0131plar ger\u00e7ek bir endi\u015fe kayna\u011f\u0131d\u0131r. Elektrik iletim ve da\u011f\u0131t\u0131m s\u0131ras\u0131nda hareket ederken, kablolardaki diren\u00e7 nedeniyle 6%'ye kadar enerji bo\u015fa harcan\u0131yor. \u015eebekeden GPU'ya giden g\u00fc\u00e7 d\u00f6rt kereden fazla d\u00f6n\u00fc\u015ft\u00fcr\u00fclmekte ve bu da ortalama 12% enerji kayb\u0131na yol a\u00e7maktad\u0131r.<\/p>\n

Binlerce sunucunun her biri 40 kilowatt (kW) t\u00fcketebilir, bu nedenle g\u00fcc\u00fc raflara ta\u015f\u0131mak i\u00e7in a\u011f\u0131r hizmet tipi bir veri yolu kullan\u0131l\u0131r. Standart 12 volt do\u011fru ak\u0131m (VDC<\/sub>) veri yolu 48 VDC<\/sub>\u00a0ak\u0131mlar\u0131 azaltmak i\u00e7in. Ancak yapay zeka i\u00e7in enerji talebini kar\u015f\u0131lamak \u00fczere, daha y\u00fcksek bir +\/-400 VDC<\/sub>\u00a0otob\u00fcs mimarisi muhtemelen gerekli olacakt\u0131r.<\/p>\n

\"\"<\/p>\n

\u015eekil 1: Veri merkezleri birden fazla g\u00fc\u00e7 d\u00f6n\u00fc\u015ft\u00fcrme a\u015famas\u0131 gerektirir<\/p>\n

G\u00fc\u00e7 yar\u0131 iletkenleri, yapay zeka i\u015flemcilerinin ve GPU'lar\u0131n ihtiya\u00e7lar\u0131n\u0131 kar\u015f\u0131lamak \u00fczere g\u00fcc\u00fcn verimli bir \u015fekilde d\u00f6n\u00fc\u015ft\u00fcr\u00fclmesi i\u00e7in gereklidir. Silisyum karb\u00fcr (SiC) ve galyum nitr\u00fcr (GaN), son derece kompakt ve enerji tasarruflu g\u00fc\u00e7 d\u00f6n\u00fc\u015ft\u00fcr\u00fcc\u00fcleri sa\u011flayarak veri merkezlerinin toplam sahip olma maliyetini (TCO) \u00f6nemli \u00f6l\u00e7\u00fcde art\u0131rd\u0131\u011f\u0131 i\u00e7in silikonun yerini almaktad\u0131r.<\/p>\n

Verimlilik ve S\u00fcrd\u00fcr\u00fclebilirlik i\u00e7in Yenilik\u00e7i \u00c7\u00f6z\u00fcmler<\/strong><\/p>\n

\u015eebekeden GPU raf\u0131na kadar veri merkezi g\u00fc\u00e7 da\u011f\u0131t\u0131m\u0131 bir\u00e7ok g\u00fc\u00e7 d\u00f6n\u00fc\u015f\u00fcm\u00fcnden ge\u00e7er. Ak\u0131ll\u0131 SiC ve silikon (Si) g\u00fc\u00e7 \u00e7\u00f6z\u00fcmleri, g\u00fc\u00e7 a\u011fac\u0131n\u0131n her bir dal\u0131nda etkili olmaktad\u0131r. G\u00fc\u00e7 ilk olarak bir kat\u0131 hal transformat\u00f6r\u00fcnden (SST) ve bir dizel jenerat\u00f6r taraf\u0131ndan desteklenen otomatik transfer anahtar\u0131 (ATS) kontrol\u00fcnden ge\u00e7er. 20k VAC<\/sub>\u00a0hatt\u0131 \u00fc\u00e7 fazl\u0131 400 V'a d\u00f6n\u00fc\u015ft\u00fcr\u00fcl\u00fcrAC<\/sub>, ve daha sonra kesintisiz bir g\u00fc\u00e7 kayna\u011f\u0131ndan (UPS) ge\u00e7er. EliteSiC ayr\u0131k ve g\u00fc\u00e7 mod\u00fclleri, veri merkezine bu giri\u015f noktas\u0131nda daha y\u00fcksek verimlilik ve g\u00fc\u00e7 yo\u011funlu\u011fu sa\u011flamak i\u00e7in kullan\u0131labilir. G\u00fc\u00e7 da\u011f\u0131t\u0131m \u00fcnitesi daha sonra \u00fc\u00e7 fazl\u0131 400 VAC<\/sub>\u00a0tek fazl\u0131 230 V'aAC<\/sub>\u00a0raf seviyesinde hat.<\/p>\n

GPU sunucular\u0131n\u0131n bulundu\u011fu rafta, g\u00fc\u00e7 d\u00f6n\u00fc\u015f\u00fcm\u00fcn\u00fcn geri kalan\u0131 ger\u00e7ekle\u015fir. G\u00fc\u00e7 kayna\u011f\u0131 \u00fcnitesi (PSU) ve batarya yedekleme \u00fcnitesinde, SiC cascode JFET'ler ve PowerTrench T10 Si MOSFET'lerin kombinasyonu, y\u00fcksek g\u00fc\u00e7l\u00fc AC-DC \u00e7\u00f6z\u00fcmleri i\u00e7in idealdir. Y\u00fcksek ak\u0131ml\u0131 SiC kaskod JFET'ler, yeni nesil hiper \u00f6l\u00e7ekli mimaride gerekli olan 3 kW'tan 5 kW PSU'ya ge\u00e7i\u015f i\u00e7in gereklidir.<\/p>\n

onsemi'nin EliteSiC 650 V MOSFET'leri ve T10 MOSFET'leri 230 V'un d\u00f6n\u00fc\u015ft\u00fcr\u00fclmesinde kullan\u0131l\u0131rAC<\/sub>\u00a0hat gerilimi 48 V'aDC<\/sub>\u00a0\u00f6nce ve sonra 12 V'aDC\u00a0<\/sub>g\u00fc\u00e7 ak\u0131\u015f\u0131 boyunca. Open Rack V3 (ORV3) spesifikasyonu olan 97,5% tepe verimlili\u011fini korumak i\u00e7in d\u00f6n\u00fc\u015ft\u00fcrme verimlili\u011fi burada anahtard\u0131r. Bu y\u00fcksek verimlilik bo\u015fa harcanan enerjiyi azalt\u0131r ve i\u015fletme giderlerini ve so\u011futma taleplerini d\u00fc\u015f\u00fcrmeye yard\u0131mc\u0131 olur. T10 Si MOSFET'ler ve g\u00fc\u00e7 y\u00f6netimi IC'leri, g\u00fc\u00e7 a\u011fac\u0131n\u0131n Vcore (CPU \u00e7ekirdek voltaj\u0131) dal\u0131na g\u00fc\u00e7 sa\u011flamak i\u00e7in 48 V'u 12 V'luk bir ara veri yolu d\u00f6n\u00fc\u015ft\u00fcr\u00fcc\u00fc (IBC) voltaj\u0131na d\u00f6n\u00fc\u015ft\u00fcrmek i\u00e7in de kullan\u0131l\u0131r. Ayr\u0131ca, 400\/800 V veri yolu mimarileri i\u00e7in SiC JFET'ler ve SiC Combo JFET'ler, IBC a\u015famas\u0131ndan \u00f6nce hot swap\/e-Fuse i\u00e7in g\u00fcvenilir a\u015f\u0131r\u0131 ak\u0131m korumas\u0131 sunar.<\/p>\n

Yapay Zeka Veri Merkezlerinde G\u00fc\u00e7 Y\u00f6netiminin Gelece\u011fi\u00a0<\/strong><\/p>\n

Verimlilik, yapay zeka veri merkezlerindeki en kritik g\u00fc\u00e7 parametresidir. Bu, kay\u0131plar\u0131n m\u00fcmk\u00fcn olan her yerde en aza indirilmesi gerekti\u011fi anlam\u0131na gelir, \u00e7\u00fcnk\u00fc so\u011futma, veri merkezlerinde kullan\u0131lan g\u00fcc\u00fcn 50%'sine kadar\u0131n\u0131 t\u00fcketebilirken, di\u011fer yar\u0131s\u0131 sunucular, depolama sistemleri ve g\u00fc\u00e7 altyap\u0131s\u0131 gibi BT ekipman\u0131 taraf\u0131ndan t\u00fcketilir.<\/p>\n

onsemi, yapay zeka veri merkezi \u00e7\u00f6z\u00fcmlerinde liderdir ve \u015febekeden GPU'ya kadar t\u00fcm g\u00fc\u00e7 a\u011fac\u0131n\u0131n ihtiya\u00e7lar\u0131n\u0131 kar\u015f\u0131layabilen birka\u00e7 tedarik\u00e7iden biridir. Gelecek, daha y\u00fcksek frekanslarda sa\u011flam g\u00fc\u00e7 d\u00f6n\u00fc\u015f\u00fcm\u00fc ve daha kompakt tasar\u0131mlara izin veren daha y\u00fcksek verimlilikleri i\u00e7in onsemi'nin EliteSiC ve dikey GaN gibi geli\u015fmi\u015f geni\u015f bant aral\u0131\u011f\u0131 teknolojilerini gerektirecektir. Bu cihazlar daha y\u00fcksek s\u0131cakl\u0131klarda g\u00fcvenilir bir \u015fekilde \u00e7al\u0131\u015fabilir, daha az so\u011futma gerektirir ve i\u015fletme maliyetlerini d\u00fc\u015f\u00fcrmenin yan\u0131 s\u0131ra daha kompakt \u00e7\u00f6z\u00fcmler sa\u011flar.<\/p>","protected":false},"excerpt":{"rendered":"

Yapay zeka (AI) h\u0131zla geli\u015fiyor ve i\u00e7erik \u00fcretiminden \u00e7evrimi\u00e7i ortama kadar g\u00fcnl\u00fck hayat\u0131m\u0131z\u0131n pek \u00e7ok alan\u0131na n\u00fcfuz ediyor.<\/p>","protected":false},"author":2,"featured_media":8835,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":"","jetpack_publicize_message":"","jetpack_publicize_feature_enabled":true,"jetpack_social_post_already_shared":false,"jetpack_social_options":{"image_generator_settings":{"template":"highway","default_image_id":0,"font":"","enabled":false},"version":2}},"categories":[1],"tags":[],"class_list":["post-8834","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news"],"acf":[],"aioseo_notices":[],"jetpack_publicize_connections":[],"_links":{"self":[{"href":"https:\/\/www.lxbchip.com\/tr\/wp-json\/wp\/v2\/posts\/8834","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.lxbchip.com\/tr\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.lxbchip.com\/tr\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.lxbchip.com\/tr\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.lxbchip.com\/tr\/wp-json\/wp\/v2\/comments?post=8834"}],"version-history":[{"count":1,"href":"https:\/\/www.lxbchip.com\/tr\/wp-json\/wp\/v2\/posts\/8834\/revisions"}],"predecessor-version":[{"id":8837,"href":"https:\/\/www.lxbchip.com\/tr\/wp-json\/wp\/v2\/posts\/8834\/revisions\/8837"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.lxbchip.com\/tr\/wp-json\/wp\/v2\/media\/8835"}],"wp:attachment":[{"href":"https:\/\/www.lxbchip.com\/tr\/wp-json\/wp\/v2\/media?parent=8834"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.lxbchip.com\/tr\/wp-json\/wp\/v2\/categories?post=8834"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.lxbchip.com\/tr\/wp-json\/wp\/v2\/tags?post=8834"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}