{"id":124473,"date":"2025-10-12T21:12:40","date_gmt":"2025-10-12T21:12:40","guid":{"rendered":"https:\/\/kamucalisani.net\/?p=124473"},"modified":"2025-10-12T21:12:40","modified_gmt":"2025-10-12T21:12:40","slug":"cinli-sirketten-cigir-acan-basari-11-kat-az-islem-gucuyle-yapay-zeka-modeli-egitti","status":"publish","type":"post","link":"https:\/\/kamucalisani.net\/index.php\/2025\/10\/12\/cinli-sirketten-cigir-acan-basari-11-kat-az-islem-gucuyle-yapay-zeka-modeli-egitti\/","title":{"rendered":"\u00c7inli \u015firketten \u00e7\u0131\u011f\u0131r a\u00e7an ba\u015far\u0131:\u00a011 kat az i\u015flem g\u00fcc\u00fcyle yapay zeka modeli e\u011fitti!"},"content":{"rendered":"

<\/span> \u00c7inli bir yapay zeka giri\u015fimi olan DeepSeek<\/strong>, \u00e7\u0131\u011f\u0131r a\u00e7an bir duyuruda bulunarak, OpenAI, Meta ve Anthropic gibi \u00f6nde gelen yapay zeka \u015firketlerinin modellerine benzer bir yapay modelini, 11 kat d\u00fc\u015f\u00fck GPU hesaplama g\u00fcc\u00fcyle e\u011fitti\u011fini a\u00e7\u0131klad\u0131.\u00a0\u00a0<\/strong> <\/figure>\n

Deepseek makalesinde, DeepSeek-V3 Mixture-of-Experts (MoE)\u00a0isimli dil modelini sadeceiki ayda 2.048 Nvidia H800 GPU<\/strong>‘sunu i\u00e7eren bir k\u00fcme kullanarak 671 milyar parametreyle<\/strong> e\u011fitti, bu da 2,8 milyon GPU saati<\/strong> anlam\u0131na geliyor. Kar\u015f\u0131la\u015ft\u0131rma yapmak gerekirse, Meta’n\u0131n 54 g\u00fcn<\/strong> boyunca 16.384 adet H100 GPU<\/strong> i\u00e7eren bir k\u00fcme kullanarak 405 milyar parametreli<\/strong> Llama 3’\u00fcn\u00fc e\u011fitmesi 11 kat daha fazla i\u015flem g\u00fcc\u00fc (30,8 milyon GPU saati)<\/strong> gerektirdi.<\/p>\n

\u00c7e\u015fitli optimizasyonlar yap\u0131ld\u0131<\/b><\/p>\n

DeepSeek, geli\u015fmi\u015f ileti\u015fim hatt\u0131 (pipeline) algoritmalar\u0131, optimize edilmi\u015f ileti\u015fim \u00e7er\u00e7evesi ve FP8 d\u00fc\u015f\u00fck hassasiyetli hesaplama kullanarak bu \u00f6l\u00e7ekteki modeller i\u00e7in tipik olarak gerekli olan hesaplama ve bellek taleplerini \u00f6nemli \u00f6l\u00e7\u00fcde azaltt\u0131\u011f\u0131n\u0131<\/strong> iddia ediyor.<\/p>\n

DeepSeek, DeepSeek-v3’\u00fcn\u00fcn i\u015flem gereksinimlerini azaltmak i\u00e7in onlarca optimizasyon tekni\u011fi uygularken, birka\u00e7 \u00f6nemli teknoloji etkileyici sonu\u00e7lar\u0131n\u0131 m\u00fcmk\u00fcn k\u0131ld\u0131.<\/p>\n

DeepSeek, hesaplama ve ileti\u015fim a\u015famalar\u0131nnda DualPipe<\/strong> algoritmas\u0131n\u0131 kulland\u0131\u011f\u0131n\u0131 ve bu nedenle iletim hatt\u0131ndaki verimsizlikleri azaltt\u0131\u011f\u0131n\u0131 <\/strong>s\u00f6yl\u00fcyor. DualPipe algoritmas\u0131, \u00f6zellikle MoE mimarisinin gerektirdi\u011fi\u00a0d\u00fc\u011f\u00fcmler aras\u0131 uzman paralelli\u011fi i\u00e7in e\u011fitim darbo\u011fazlar\u0131n\u0131 en aza indirdi\u00a0ve bu optimizasyon, k\u00fcmenin \u00f6n e\u011fitim s\u0131ras\u0131nda s\u0131f\u0131ra yak\u0131n ileti\u015fim y\u00fck\u00fcyle 14,8 trilyon jetonu (token) i\u015flemesine olanak sa\u011flad\u0131,\u00a0<\/p>\n

DeepSeek, DualPipe’\u0131 uygulamaya ek olarak, ileti\u015fime dahil olan d\u00fc\u011f\u00fcm say\u0131s\u0131n\u0131 s\u0131n\u0131rlamak i\u00e7in her tokeni maksimum d\u00f6rt d\u00fc\u011f\u00fcmle s\u0131n\u0131rlad\u0131. Bu da trafi\u011fi azaltt\u0131 ve ileti\u015fimin ve hesaplaman\u0131n etkili bir \u015fekilde \u00f6rt\u00fc\u015fmesini sa\u011flad\u0131.<\/p>\n

DeepSeek-v3 nas\u0131l performans sergiliyor?<\/b><\/p>\n

<\/span> Performansa gelirsek, \u015firket DeepSeek-v3 MoE dil modelinin k\u0131yaslamaya ba\u011fl\u0131 olarak GPT-4x, Claude-3.5-Sonnet ve LLlama-3.1 ile kar\u015f\u0131la\u015ft\u0131r\u0131labilir veya daha iyi performansa<\/strong> sahip oldu\u011funu s\u00f6yl\u00fcyor. Ancak bu iddialar\u0131n \u00fc\u00e7\u00fcnc\u00fc taraflarca ispatlanmas\u0131 gerekiyor. \u015eirket modeli ve a\u011f\u0131rl\u0131klar\u0131 a\u00e7\u0131k kaynakl\u0131 hale getirdi, bu nedenle yak\u0131n zamanda kar\u015f\u0131la\u015ft\u0131rma testleri ortaya \u00e7\u0131kacakt\u0131r. <\/figure>\n
<\/span> DeepSeek-V3, parametre say\u0131s\u0131 veya muhakeme yetenekleri a\u00e7\u0131s\u0131ndan GPT-4o veya o3 gibi \u00f6nc\u00fc modellerin gerisinde kalsa da, DeepSeek’in ba\u015far\u0131lar\u0131, nispeten s\u0131n\u0131rl\u0131 kaynaklar kullanarak geli\u015fmi\u015f bir MoE dil modelinin e\u011fitilmesinin m\u00fcmk\u00fcn oldu\u011funu<\/strong> g\u00f6steriyor. Elbette, bu \u00e7ok fazla optimizasyon ve d\u00fc\u015f\u00fck seviyeli programlama gerektiriyor, ancak sonu\u00e7lar \u015fa\u015f\u0131rt\u0131c\u0131 derecede iyi g\u00f6r\u00fcn\u00fcyor. <\/figure>\n

DeepSeek ekibi, DeepSeek-V3 modelinin uygulanmas\u0131n\u0131n, geli\u015fmi\u015f donan\u0131m\u0131n yan\u0131 s\u0131ra \u00f6n doldurma ve kod \u00e7\u00f6zme a\u015famalar\u0131n\u0131 ay\u0131ran bir da\u011f\u0131t\u0131m stratejisi gerektirdi\u011fini ve bunun kaynak eksikli\u011fi nedeniyle k\u00fc\u00e7\u00fck \u015firketler i\u00e7in eri\u015filemez olabilece\u011fini kabul ediyor.<\/p>\n\n

Kaynak\u00a0 :\u00a0https:\/\/www.donanimhaber.com\/cinli-sirket-11-kat-az-islem-gucuyle-yapay-zeka-modeli-egitti–185854<\/span><\/p>\n","protected":false},"excerpt":{"rendered":"

\u00c7inli bir yapay zeka giri\u015fimi olan DeepSeek, \u00e7\u0131\u011f\u0131r a\u00e7an bir duyuruda bulunarak, OpenAI, Meta ve Anthropic gibi \u00f6nde gelen yapay zeka \u015firketlerinin modellerine benzer bir yapay modelini, 11 kat d\u00fc\u015f\u00fck GPU hesaplama g\u00fcc\u00fcyle e\u011fitti\u011fini a\u00e7\u0131klad\u0131 …<\/p>\n","protected":false},"author":1,"featured_media":124474,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[8278,500,6864,2622,3056],"class_list":["post-124473","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-teknoloji","tag-aza","tag-gpu","tag-hesaplama","tag-iletisim","tag-modelini"],"_links":{"self":[{"href":"https:\/\/kamucalisani.net\/index.php\/wp-json\/wp\/v2\/posts\/124473","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/kamucalisani.net\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/kamucalisani.net\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/kamucalisani.net\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/kamucalisani.net\/index.php\/wp-json\/wp\/v2\/comments?post=124473"}],"version-history":[{"count":1,"href":"https:\/\/kamucalisani.net\/index.php\/wp-json\/wp\/v2\/posts\/124473\/revisions"}],"predecessor-version":[{"id":124478,"href":"https:\/\/kamucalisani.net\/index.php\/wp-json\/wp\/v2\/posts\/124473\/revisions\/124478"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/kamucalisani.net\/index.php\/wp-json\/wp\/v2\/media\/124474"}],"wp:attachment":[{"href":"https:\/\/kamucalisani.net\/index.php\/wp-json\/wp\/v2\/media?parent=124473"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/kamucalisani.net\/index.php\/wp-json\/wp\/v2\/categories?post=124473"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/kamucalisani.net\/index.php\/wp-json\/wp\/v2\/tags?post=124473"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}