{"id":2784,"date":"2026-10-07T15:00:00","date_gmt":"2026-10-07T12:00:00","guid":{"rendered":"https:\/\/narhost.com\/blog\/?p=2784"},"modified":"2026-09-26T23:19:02","modified_gmt":"2026-09-26T20:19:02","slug":"yapay-zeka-sunucu-kiralama","status":"publish","type":"post","link":"https:\/\/narhost.com\/blog\/yapay-zeka-sunucu-kiralama\/","title":{"rendered":"Yapay Zek\u00e2 Sunucusu Kiralarken GPU Gerekir mi"},"content":{"rendered":"<div id=\"ai-ozet\" class=\"ai-box\">\n<p class=\"ai-box-title\">K\u0131saca \u00f6zet<\/p>\n<ul>\n<li>Yapay zeka sunucu kiralama karar\u0131nda ilk soru GPU de\u011fil, modeli kimin \u00e7al\u0131\u015ft\u0131raca\u011f\u0131d\u0131r: bir yapay zek\u00e2 servisini API ile kullan\u0131yorsan\u0131z sunucunuzda GPU gerekmez, s\u0131radan bir VPS ya da VDS yeter.<\/li>\n<li>Modeli kendi sunucunuzda bar\u0131nd\u0131rmak iki ayr\u0131 i\u015ftir: e\u011fitim neredeyse her zaman GPU ister, haz\u0131r bir modelle cevap \u00fcretmek (\u00e7\u0131kar\u0131m) ise k\u00fc\u00e7\u00fck modellerde i\u015flemciyle de m\u00fcmk\u00fcnd\u00fcr.<\/li>\n<li>Bellek hesab\u0131 basittir: 16 bit hassasiyette milyar parametre ba\u015f\u0131na yakla\u015f\u0131k 2 GB gerekir; 7 milyar parametreli bir model 14 GB civar\u0131 ister, 4 bit nicemlemeyle bu ihtiya\u00e7 kabaca d\u00f6rtte birine iner.<\/li>\n<li>\u0130\u015flemcide \u00e7al\u0131\u015fan modelde darbo\u011faz bellek h\u0131z\u0131 ve \u00e7ekirdek say\u0131s\u0131d\u0131r; tek kullan\u0131c\u0131l\u0131, gecikmeye toleransl\u0131 i\u015fler i\u00e7in uygundur, \u00e7ok kullan\u0131c\u0131l\u0131 canl\u0131 sohbet i\u00e7in yetersiz kal\u0131r.<\/li>\n<li>Vekt\u00f6r veritaban\u0131, g\u00f6mme (embedding) \u00fcretimi, API tabanl\u0131 sohbet botu, belge arama ve i\u015f ak\u0131\u015f\u0131 otomasyonu GPU&#8217;suz sunucuda rahat \u00e7al\u0131\u015fan yapay zek\u00e2 i\u015fleridir.<\/li>\n<li>B\u00fcy\u00fck model e\u011fitimi, ince ayar (fine-tuning) ve y\u00fcksek e\u015fzamanl\u0131 \u00e7\u0131kar\u0131m i\u00e7in GPU&#8217;lu donan\u0131m gerekir; bu i\u015flerde i\u015flemci sunucusu do\u011fru ara\u00e7 de\u011fildir.<\/li>\n<\/ul>\n<\/div>\n<p>Bir yapay zek\u00e2 projesi planl\u0131yorsunuz ve ilk buldu\u011funuz tavsiye &#8220;GPU&#8217;lu sunucu kirala&#8221; oluyor. Fiyatlara bak\u0131nca proje daha ba\u015flamadan b\u00fct\u00e7e a\u015f\u0131l\u0131yor. Oysa yapay zeka sunucu kiralama sorusunun cevab\u0131 \u00e7o\u011fu zaman beklenenden sadedir: i\u015fin \u00f6nemli bir k\u0131sm\u0131 GPU olmadan, s\u0131radan bir sunucuda y\u00fcr\u00fcr.<\/p>\n<h2>Yapay Zeka Sunucu Kiralama \u00d6ncesi \u0130lk Soru, API mi Kendi Model mi<\/h2>\n<p>Yapay zek\u00e2 kullanan uygulamalar\u0131n b\u00fcy\u00fck k\u0131sm\u0131 modeli kendisi \u00e7al\u0131\u015ft\u0131rmaz. Sohbet botu, metin \u00f6zetleyen bir ara\u00e7 ya da m\u00fc\u015fteri e-postalar\u0131n\u0131 s\u0131n\u0131fland\u0131ran bir sistem, \u00e7o\u011funlukla bir yapay zek\u00e2 sa\u011flay\u0131c\u0131s\u0131n\u0131n API&#8217;sine istek g\u00f6nderir ve gelen cevab\u0131 i\u015fler. Bu mimaride a\u011f\u0131r hesaplama sa\u011flay\u0131c\u0131n\u0131n veri merkezinde yap\u0131l\u0131r. Sizin sunucunuzun g\u00f6revi web uygulamas\u0131n\u0131 \u00e7al\u0131\u015ft\u0131rmak, iste\u011fi haz\u0131rlamak, cevab\u0131 veritaban\u0131na yazmak ve kullan\u0131c\u0131ya g\u00f6stermektir.<\/p>\n<p>Bu i\u015f i\u00e7in gereken \u015fey bir web uygulamas\u0131n\u0131n ihtiyac\u0131ndan farkl\u0131 de\u011fildir: birka\u00e7 \u00e7ekirdek, uygulaman\u0131n ve veritaban\u0131n\u0131n s\u0131\u011faca\u011f\u0131 kadar bellek, kal\u0131c\u0131 \u00e7al\u0131\u015fan s\u00fcre\u00e7ler i\u00e7in k\u00f6k eri\u015fimi. Python ya da Node.js ile yaz\u0131lm\u0131\u015f bu t\u00fcr bir uygulaman\u0131n bar\u0131nd\u0131rma se\u00e7eneklerini <a href=\"https:\/\/narhost.com\/blog\/python-nodejs-icin-hangi-hosting\/\" title=\"Python ve Node.js \u0130\u00e7in Hangi Hosting \u2014 Narhost Blog\">Python ve Node.js i\u00e7in hangi hosting yaz\u0131s\u0131nda<\/a> bulabilirsiniz. GPU bu senaryoda hi\u00e7bir i\u015fe yaramaz, yaln\u0131zca faturay\u0131 b\u00fcy\u00fct\u00fcr.<\/p>\n<div class=\"highlight-box\">\n<p class=\"label\">Bilgi<\/p>\n<p>API ile \u00e7al\u0131\u015fan bir uygulamada maliyet kalemi sunucu de\u011fil, sa\u011flay\u0131c\u0131ya g\u00f6nderilen istek hacmidir. Sunucuyu b\u00fcy\u00fctmek cevab\u0131 h\u0131zland\u0131rmaz; cevap s\u00fcresini sa\u011flay\u0131c\u0131n\u0131n modeli belirler.<\/p>\n<\/div>\n<h2>Model Bar\u0131nd\u0131rmak, E\u011fitim ve \u00c7\u0131kar\u0131m Ayr\u0131m\u0131<\/h2>\n<p>Modeli kendi sunucunuzda \u00e7al\u0131\u015ft\u0131rmak istemenizin ge\u00e7erli nedenleri olabilir: veri d\u0131\u015far\u0131 \u00e7\u0131kmamal\u0131d\u0131r, istek hacmi API maliyetini a\u015fm\u0131\u015ft\u0131r ya da internet ba\u011flant\u0131s\u0131 olmadan \u00e7al\u0131\u015fan bir sistem gerekir. Bu noktada iki i\u015fi ay\u0131rmak gerekir.<\/p>\n<p>E\u011fitim, modelin a\u011f\u0131rl\u0131klar\u0131n\u0131 veriden \u00f6\u011frenerek olu\u015fturmakt\u0131r. S\u0131f\u0131rdan e\u011fitim ya da mevcut bir modelin ince ayar\u0131, milyarlarca say\u0131 \u00fczerinde tekrar tekrar yap\u0131lan matris i\u015flemleri demektir ve bu i\u015f i\u00e7in GPU&#8217;nun paralel hesaplama g\u00fcc\u00fc ger\u00e7ek anlamda \u015fartt\u0131r. \u00c7\u0131kar\u0131m ise e\u011fitilmi\u015f bir modelle cevap \u00fcretmektir. Burada hesaplama y\u00fck\u00fc \u00e7ok daha k\u00fc\u00e7\u00fckt\u00fcr ve k\u00fc\u00e7\u00fck modellerde i\u015flemci de i\u015fi g\u00f6r\u00fcr; ancak h\u0131z GPU&#8217;ya g\u00f6re belirgin bi\u00e7imde d\u00fc\u015f\u00fckt\u00fcr.<\/p>\n<table>\n<thead>\n<tr>\n<th>\u0130\u015f<\/th>\n<th>GPU gerekir mi<\/th>\n<th>\u0130\u015flemci sunucusunda durum<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>API ile sohbet botu ya da metin i\u015fleme<\/td>\n<td>Hay\u0131r<\/td>\n<td>Rahat \u00e7al\u0131\u015f\u0131r, s\u0131radan web uygulamas\u0131 gibi<\/td>\n<\/tr>\n<tr>\n<td>Vekt\u00f6r veritaban\u0131 ve belge arama<\/td>\n<td>Hay\u0131r<\/td>\n<td>Bellek ve disk h\u0131z\u0131 belirleyicidir<\/td>\n<\/tr>\n<tr>\n<td>K\u00fc\u00e7\u00fck g\u00f6mme modeliyle vekt\u00f6r \u00fcretimi<\/td>\n<td>Genellikle hay\u0131r<\/td>\n<td>Toplu i\u015flerde yava\u015f ama kabul edilebilir<\/td>\n<\/tr>\n<tr>\n<td>7\u20138 milyar parametreli modelle \u00e7\u0131kar\u0131m<\/td>\n<td>H\u0131z i\u00e7in evet<\/td>\n<td>4 bit nicemlemeyle \u00e7al\u0131\u015f\u0131r, tek kullan\u0131c\u0131da kullan\u0131labilir<\/td>\n<\/tr>\n<tr>\n<td>\u0130nce ayar ve e\u011fitim<\/td>\n<td>Evet<\/td>\n<td>Pratikte uygun de\u011fil<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h2>Bellek Hesab\u0131, Model Boyutu Ne Kadar RAM \u0130ster<\/h2>\n<p>Bir modelin bellek ihtiyac\u0131n\u0131 tahmin etmek i\u00e7in parametre say\u0131s\u0131na ve a\u011f\u0131rl\u0131klar\u0131n hassasiyetine bak\u0131l\u0131r. Hugging Face&#8217;in <a href=\"https:\/\/huggingface.co\/docs\/transformers\/main\/en\/llm_tutorial_optimization\" title=\"Hugging Face Transformers \u2014 b\u00fcy\u00fck dil modellerinde h\u0131z ve bellek optimizasyonu\" target=\"_blank\" rel=\"noopener\">b\u00fcy\u00fck dil modeli optimizasyon rehberi<\/a> kural\u0131 a\u00e7\u0131k\u00e7a verir: X milyar parametreli bir modelin a\u011f\u0131rl\u0131klar\u0131n\u0131 16 bit hassasiyette y\u00fcklemek yakla\u015f\u0131k 2 \u00d7 X GB bellek ister. Ayn\u0131 rehbere g\u00f6re 8 ve 4 bit nicemleme (quantization) bu ihtiyac\u0131 belirgin bi\u00e7imde d\u00fc\u015f\u00fcr\u00fcr; kar\u015f\u0131l\u0131\u011f\u0131 do\u011frulukta k\u00fc\u00e7\u00fck bir kay\u0131pt\u0131r.<\/p>\n<p>Pratikte bunun anlam\u0131 \u015fudur: 7 milyar parametreli bir model 16 bit h\u00e2linde 14 GB civar\u0131, 4 bit nicemlenmi\u015f h\u00e2linde ise 4\u20135 GB civar\u0131 bellek ister. Buna modelin uzun metinleri i\u015flerken tuttu\u011fu ara bellek, i\u015fletim sistemi ve uygulaman\u0131z\u0131n kendisi eklenir. \u0130\u015flemcide \u00e7al\u0131\u015fan modelde bu bellek GPU belle\u011fi de\u011fil sunucunun RAM&#8217;idir; bu y\u00fczden i\u015flemciyle model \u00e7al\u0131\u015ft\u0131racaksan\u0131z RAM, \u00e7ekirdek say\u0131s\u0131ndan \u00f6nce gelen kalemdir.<\/p>\n<div class=\"warn-box\">\n<p class=\"label\">Dikkat<\/p>\n<p>Belle\u011fe s\u0131\u011fmayan bir modeli takas alan\u0131yla (swap) \u00e7al\u0131\u015ft\u0131rmaya \u00e7al\u0131\u015fmay\u0131n. Model her cevapta a\u011f\u0131rl\u0131klar\u0131n tamam\u0131n\u0131 okur; a\u011f\u0131rl\u0131klar diske ta\u015farsa cevap \u00fcretimi dakikalara uzar ve sunucudaki di\u011fer her \u015fey de yava\u015flar. Modelin nicemlenmi\u015f boyutu, sunucunun bo\u015ftaki belle\u011finin rahat\u00e7a alt\u0131nda kalmal\u0131d\u0131r.<\/p>\n<\/div>\n<h2>\u0130\u015flemciyle Model \u00c7al\u0131\u015ft\u0131rman\u0131n S\u0131n\u0131rlar\u0131<\/h2>\n<p>\u0130\u015flemcide \u00e7\u0131kar\u0131m i\u00e7in en yayg\u0131n ara\u00e7 <a href=\"https:\/\/github.com\/ggml-org\/llama.cpp\" title=\"llama.cpp resm\u00ee GitHub deposu\" target=\"_blank\" rel=\"noopener\">llama.cpp<\/a> ve onu temel alan ara\u00e7lard\u0131r. Proje, amac\u0131n\u0131 geni\u015f bir donan\u0131m yelpazesinde en az kurulumla model \u00e7al\u0131\u015ft\u0131rmak olarak tan\u0131mlar; x86 i\u015flemcilerde AVX, AVX2 ve AVX-512 komut setlerini kullan\u0131r ve 1,5 bitten 8 bite kadar nicemleme destekler. Yani 7\u20138 milyar parametreli bir model, yeterli belle\u011fi olan \u00e7ok \u00e7ekirdekli bir sunucuda GPU olmadan \u00e7al\u0131\u015f\u0131r.<\/p>\n<p>S\u0131n\u0131r h\u0131zdad\u0131r. \u0130\u015flemcide cevap \u00fcretimi GPU&#8217;ya g\u00f6re belirgin bi\u00e7imde yava\u015ft\u0131r ve h\u0131z, bellek bant geni\u015fli\u011fi ile \u00e7ekirdek say\u0131s\u0131yla s\u0131n\u0131rl\u0131d\u0131r. Tek bir kullan\u0131c\u0131n\u0131n gece \u00e7al\u0131\u015fan bir \u00f6zetleme g\u00f6revi, i\u00e7 belgeler \u00fczerinde soru cevap ya da cevab\u0131n birka\u00e7 saniye gecikmesinin sorun olmad\u0131\u011f\u0131 bir ara\u00e7 i\u00e7in bu h\u0131z yeterlidir. Ayn\u0131 anda onlarca kullan\u0131c\u0131n\u0131n canl\u0131 sohbet etti\u011fi bir \u00fcr\u00fcn i\u00e7in ise yeterli de\u011fildir; b\u00f6yle bir \u00fcr\u00fcn ya API&#8217;ye ya da GPU&#8217;lu donan\u0131ma dayanmal\u0131d\u0131r.<\/p>\n<div class=\"senaryo-block\">\n<p class=\"label\">\u00d6rnek senaryo<\/p>\n<p>Yirmi ki\u015filik bir muhasebe b\u00fcrosu, m\u00fc\u015fteri s\u00f6zle\u015fmelerinden belirli maddeleri bulan bir i\u00e7 ara\u00e7 istedi ve belgelerin d\u0131\u015far\u0131 \u00e7\u0131kmamas\u0131n\u0131 \u015fart ko\u015ftu. Ekip 8 \u00e7ekirdek ve 32 GB bellekli bir sunucuya 4 bit nicemlenmi\u015f 8 milyar parametreli bir model, k\u00fc\u00e7\u00fck bir g\u00f6mme modeli ve bir vekt\u00f6r veritaban\u0131 kurdu. Belgeler gece toplu i\u015flendi; g\u00fcnd\u00fcz gelen sorulara cevap birka\u00e7 saniye s\u00fcrd\u00fc. Ayn\u0131 anda en fazla \u00fc\u00e7 ki\u015fi kulland\u0131\u011f\u0131 i\u00e7in GPU&#8217;ya gerek kalmad\u0131.<\/p>\n<\/div>\n<figure class=\"nh-sema\"><img loading=\"lazy\" src=\"https:\/\/narhost.com\/blog\/wp-content\/uploads\/2026\/09\/yapay-zeka-sunucu-kiralama-sema-narhost.webp\" alt=\"yapay zeka sunucu kiralama karar\u0131n\u0131 API kullan\u0131m\u0131, vekt\u00f6r veritaban\u0131, i\u015flemcide k\u00fc\u00e7\u00fck model ve GPU gerektiren e\u011fitim ad\u0131mlar\u0131yla g\u00f6steren \u015fema\" width=\"1200\" height=\"500\" loading=\"lazy\" decoding=\"async\"><figcaption>API kullan\u0131m\u0131ndan vekt\u00f6r veritaban\u0131na, i\u015flemcide k\u00fc\u00e7\u00fck modelden GPU gerektiren e\u011fitime giden karar s\u0131ras\u0131.<\/figcaption><\/figure>\n<h2>GPU Olmadan Sunucuda Yap\u0131labilecek Yapay Zek\u00e2 \u0130\u015fleri<\/h2>\n<p>GPU&#8217;suz bir sunucu, yapay zek\u00e2 projelerinin g\u00f6r\u00fcnmeyen ama i\u015f y\u00fck\u00fcn\u00fcn \u00e7o\u011funu ta\u015f\u0131yan k\u0131sm\u0131n\u0131 rahat\u00e7a \u00fcstlenir. Bunlar\u0131n ba\u015f\u0131nda API tabanl\u0131 uygulamalar gelir: sohbet aray\u00fcz\u00fc, m\u00fc\u015fteri destek asistan\u0131, i\u00e7erik s\u0131n\u0131fland\u0131rma ya da form verisini ayr\u0131\u015ft\u0131ran bir servis. \u0130kinci grup vekt\u00f6r veritabanlar\u0131d\u0131r; belgeleri par\u00e7alara b\u00f6l\u00fcp anlam vekt\u00f6rleri olarak saklayan ve soruya en yak\u0131n par\u00e7ay\u0131 bulan bu sistemler bellek ve h\u0131zl\u0131 disk ister, GPU istemez.<\/p>\n<p>\u00dc\u00e7\u00fcnc\u00fc grup k\u00fc\u00e7\u00fck modellerdir: metin s\u0131n\u0131fland\u0131rma, dil tespiti, g\u00f6mme \u00fcretimi ve k\u00fc\u00e7\u00fck boyutlu dil modelleri i\u015flemcide makul h\u0131zda \u00e7al\u0131\u015f\u0131r. D\u00f6rd\u00fcnc\u00fc grup ise i\u015f ak\u0131\u015f\u0131 otomasyonudur; bir e-postay\u0131 al\u0131p yapay zek\u00e2 servisine g\u00f6nderen, cevab\u0131 bir tabloya yazan ve ilgili ki\u015fiye bildirim atan ak\u0131\u015flar tamamen i\u015flemci ve a\u011f i\u015fidir. Bu servislerin \u00e7o\u011fu Docker ile kurulur; kurulum ad\u0131mlar\u0131 <a href=\"https:\/\/narhost.com\/destek\/ubuntu-docker-kurulumu\/\" title=\"Ubuntu&#039;da Docker Kurulumu \u2014 Narhost Destek Merkezi\">Ubuntu&#8217;da Docker kurulumu kart\u0131nda<\/a> anlat\u0131l\u0131yor.<\/p>\n<h2>Kaynak Tablosu, Hangi \u0130\u015f \u0130\u00e7in Ne Kadar Sunucu<\/h2>\n<p>A\u015fa\u011f\u0131daki de\u011ferler temkinli ba\u015flang\u0131\u00e7 noktalar\u0131d\u0131r. Model bar\u0131nd\u0131ran senaryolarda belirleyici olan RAM&#8217;dir; paket se\u00e7erken \u00f6nce modelin nicemlenmi\u015f boyutunu, sonra e\u015fzamanl\u0131 kullan\u0131c\u0131 say\u0131s\u0131n\u0131 hesaba kat\u0131n. Genel kaynak hesab\u0131n\u0131n mant\u0131\u011f\u0131n\u0131 <a href=\"https:\/\/narhost.com\/blog\/vps-ne-kadar-ram-cpu-gerekir\/\" title=\"VPS Ne Kadar RAM CPU Gerekir \u2014 Narhost Blog\">VPS i\u00e7in ne kadar RAM ve CPU gerekir yaz\u0131s\u0131nda<\/a> ayr\u0131nt\u0131l\u0131 bulabilirsiniz.<\/p>\n<table>\n<thead>\n<tr>\n<th>Senaryo<\/th>\n<th>vCPU<\/th>\n<th>RAM<\/th>\n<th>Not<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>API tabanl\u0131 uygulama ve veritaban\u0131<\/td>\n<td>2<\/td>\n<td>2\u20134 GB<\/td>\n<td>S\u0131radan web uygulamas\u0131 gibi<\/td>\n<\/tr>\n<tr>\n<td>Vekt\u00f6r veritaban\u0131, birka\u00e7 bin belge<\/td>\n<td>2\u20134<\/td>\n<td>4\u20138 GB<\/td>\n<td>NVMe disk arama h\u0131z\u0131n\u0131 etkiler<\/td>\n<\/tr>\n<tr>\n<td>G\u00f6mme modeli ve belge i\u015fleme<\/td>\n<td>4<\/td>\n<td>8 GB<\/td>\n<td>Toplu i\u015fleri gece \u00e7al\u0131\u015ft\u0131r\u0131n<\/td>\n<\/tr>\n<tr>\n<td>4 bit nicemlenmi\u015f 7\u20138 milyar parametreli model<\/td>\n<td>8 ve \u00fczeri<\/td>\n<td>16\u201332 GB<\/td>\n<td>Tek ya da birka\u00e7 kullan\u0131c\u0131<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h2>Narhost&#8217;ta \u0130\u015flemci Tabanl\u0131 Yapay Zek\u00e2 Sunucusu<\/h2>\n<p>Model bar\u0131nd\u0131ran senaryolarda \u00e7ekirdeklerin size ayr\u0131lm\u0131\u015f olmas\u0131 \u00f6nemlidir; \u00e7\u0131kar\u0131m s\u0131ras\u0131nda t\u00fcm \u00e7ekirdekler ayn\u0131 anda y\u00fcklenir ve payla\u015f\u0131lan \u00e7ekirdekte h\u0131z dalgalan\u0131r. Bu y\u00fczden yapay zeka sunucu kiralama plan\u0131nda i\u015flemciyle model \u00e7al\u0131\u015ft\u0131racaksan\u0131z VDS daha tutarl\u0131 sonu\u00e7 verir. Narhost VDS&#8217;te kaynaklar size ayr\u0131l\u0131r, NVMe disk ve 50 Gbps port standartt\u0131r, Ubuntu, Debian ve AlmaLinux gibi da\u011f\u0131t\u0131mlar sipari\u015fte se\u00e7ilir ve sunucu an\u0131nda kurulur.<\/p>\n<p>VDS y\u00f6netimsiz bir \u00fcr\u00fcnd\u00fcr: model, \u00e7al\u0131\u015fma ortam\u0131, g\u00fcvenlik ve yedek sizin sorumlulu\u011funuzdad\u0131r. Paketler ve g\u00fcncel kaynak se\u00e7enekleri <a href=\"https:\/\/narhost.com\/vds-sunucu\" title=\"Narhost VDS Sunucu paketleri\">VDS sunucu<\/a> sayfas\u0131nda; payla\u015f\u0131ml\u0131 \u00e7ekirdek ile ayr\u0131lm\u0131\u015f \u00e7ekirdek fark\u0131 ise <a href=\"https:\/\/narhost.com\/blog\/vps-mi-vds-mi-bulut-sunucu-mu\/\" title=\"VPS mi VDS mi Bulut Sunucu mu \u2014 Narhost Blog\">VPS mi VDS mi bulut sunucu mu kar\u015f\u0131la\u015ft\u0131rmas\u0131nda<\/a> ayr\u0131nt\u0131l\u0131 anlat\u0131l\u0131yor.<\/p>\n<div class=\"success-box\">\n<p class=\"label\">Sonu\u00e7<\/p>\n<p>GPU, e\u011fitim ve yo\u011fun canl\u0131 \u00e7\u0131kar\u0131m i\u00e7in \u015fartt\u0131r; geri kalan i\u015flerin \u00e7o\u011fu i\u00e7in de\u011fildir. API ile \u00e7al\u0131\u015f\u0131yorsan\u0131z s\u0131radan bir sunucu yeter. K\u00fc\u00e7\u00fck bir modeli kendiniz bar\u0131nd\u0131racaksan\u0131z RAM&#8217;i modelin nicemlenmi\u015f boyutuna g\u00f6re se\u00e7in, \u00e7ekirdekleri size ayr\u0131lm\u0131\u015f bir sunucu tercih edin ve e\u015fzamanl\u0131 kullan\u0131c\u0131 say\u0131s\u0131n\u0131 ger\u00e7ek\u00e7i tahmin edin.<\/p>\n<\/div>\n<h2>Kontrol Listesi, Sunucuyu Se\u00e7meden \u00d6nce<\/h2>\n<ul class=\"kontrol-listesi\">\n<li>Modeli kendiniz mi \u00e7al\u0131\u015ft\u0131racaks\u0131n\u0131z, yoksa bir API mi kullanacaks\u0131n\u0131z, netle\u015ftirin.<\/li>\n<li>E\u011fitim ya da ince ayar gerekip gerekmedi\u011fini belirleyin; gerekiyorsa GPU&#8217;lu donan\u0131m planlay\u0131n.<\/li>\n<li>Kullanaca\u011f\u0131n\u0131z modelin parametre say\u0131s\u0131n\u0131 ve nicemlenmi\u015f dosya boyutunu yaz\u0131n.<\/li>\n<li>RAM&#8217;i model boyutu, ba\u011flam belle\u011fi, i\u015fletim sistemi ve uygulaman\u0131n toplam\u0131na g\u00f6re se\u00e7in.<\/li>\n<li>Ayn\u0131 anda ka\u00e7 ki\u015finin cevap bekleyece\u011fini ve kabul edilebilir cevap s\u00fcresini belirleyin.<\/li>\n<li>Verinin sunucu d\u0131\u015f\u0131na \u00e7\u0131k\u0131p \u00e7\u0131kamayaca\u011f\u0131n\u0131 hukuk ve m\u00fc\u015fteri s\u00f6zle\u015fmeleri a\u00e7\u0131s\u0131ndan kontrol edin.<\/li>\n<li>Model dosyalar\u0131 ve vekt\u00f6r veritaban\u0131 i\u00e7in yedek plan\u0131 yap\u0131n.<\/li>\n<\/ul>\n<div class=\"kb-faq\">\n<h2>S\u0131k\u00e7a sorulan sorular<\/h2>\n<h3>Yapay zeka sunucu kiralama i\u00e7in mutlaka GPU&#8217;lu paket mi gerekir?<\/h3>\n<p>Hay\u0131r. Bir yapay zek\u00e2 servisini API ile kullanan uygulamalar, vekt\u00f6r veritabanlar\u0131, belge arama ve otomasyon ak\u0131\u015flar\u0131 GPU&#8217;suz sunucuda \u00e7al\u0131\u015f\u0131r. GPU e\u011fitim, ince ayar ve \u00e7ok kullan\u0131c\u0131l\u0131 canl\u0131 \u00e7\u0131kar\u0131m i\u00e7in gerekir.<\/p>\n<h3>ChatGPT benzeri bir modeli kendi sunucumda \u00e7al\u0131\u015ft\u0131rabilir miyim?<\/h3>\n<p>A\u00e7\u0131k a\u011f\u0131rl\u0131kl\u0131 k\u00fc\u00e7\u00fck ve orta boy modelleri \u00e7al\u0131\u015ft\u0131rabilirsiniz. 7\u20138 milyar parametreli bir model 4 bit nicemlemeyle 16 GB ve \u00fczeri belle\u011fi olan \u00e7ok \u00e7ekirdekli bir sunucuda \u00e7al\u0131\u015f\u0131r. B\u00fcy\u00fck ticari modellerin kalitesini ve h\u0131z\u0131n\u0131 beklemeyin; tek ya da birka\u00e7 kullan\u0131c\u0131l\u0131 i\u00e7 ara\u00e7lar i\u00e7in uygundur.<\/p>\n<h3>\u0130\u015flemcide \u00e7al\u0131\u015fan model neden yava\u015f?<\/h3>\n<p>Her yeni kelime \u00fcretilirken modelin a\u011f\u0131rl\u0131klar\u0131n\u0131n b\u00fcy\u00fck k\u0131sm\u0131 bellekten okunur. GPU bunu \u00e7ok y\u00fcksek bant geni\u015fli\u011fiyle yapar, i\u015flemci ise sunucu belle\u011finin h\u0131z\u0131yla s\u0131n\u0131rl\u0131d\u0131r. Nicemleme okunacak veriyi k\u00fc\u00e7\u00fcltt\u00fc\u011f\u00fc i\u00e7in h\u0131z\u0131 art\u0131r\u0131r; \u00e7ekirdek say\u0131s\u0131 da bir noktaya kadar yard\u0131mc\u0131 olur.<\/p>\n<h3>Payla\u015f\u0131ml\u0131 hostingte yapay zek\u00e2 uygulamas\u0131 \u00e7al\u0131\u015f\u0131r m\u0131?<\/h3>\n<p>API&#8217;ye istek g\u00f6nderen basit bir PHP beti\u011fi \u00e7al\u0131\u015fabilir. Model bar\u0131nd\u0131rmak, vekt\u00f6r veritaban\u0131 kurmak ya da s\u00fcrekli \u00e7al\u0131\u015fan bir Python servisi a\u00e7mak ise payla\u015f\u0131ml\u0131 hostingin kaynak s\u0131n\u0131rlar\u0131na ve SSH olmamas\u0131na tak\u0131l\u0131r. Bu i\u015fler i\u00e7in k\u00f6k eri\u015fimli bir VPS ya da VDS gerekir.<\/p>\n<h3>Vekt\u00f6r veritaban\u0131 i\u00e7in ne kadar bellek gerekir?<\/h3>\n<p>Belge say\u0131s\u0131na, vekt\u00f6r boyutuna ve kullan\u0131lan dizin t\u00fcr\u00fcne ba\u011fl\u0131d\u0131r. Birka\u00e7 bin belgelik bir i\u00e7 arama sistemi 4\u20138 GB bellekle rahat \u00e7al\u0131\u015f\u0131r. Belge say\u0131s\u0131 y\u00fcz binlere \u00e7\u0131kt\u0131\u011f\u0131nda dizinin belle\u011fe s\u0131\u011fmas\u0131 i\u00e7in RAM&#8217;i b\u00fcy\u00fctmek ya da disk tabanl\u0131 dizinlere ge\u00e7mek gerekir.<\/p>\n<\/div>\n<div class=\"cta-box\">\n<p class=\"label\">Narhost&#8217;ta<\/p>\n<p>VDS sunucular\u0131m\u0131zda \u00e7ekirdekler size ayr\u0131l\u0131r; NVMe disk, 50 Gbps port ve an\u0131nda kurulumla API tabanl\u0131 uygulaman\u0131z\u0131, vekt\u00f6r veritaban\u0131n\u0131z\u0131 ya da k\u00fc\u00e7\u00fck modelinizi \u00e7al\u0131\u015ft\u0131rabilirsiniz.<\/p>\n<p><a class=\"cta-btn\" href=\"https:\/\/narhost.com\/vds-sunucu\" title=\"Narhost VDS Sunucu paketleri\">VDS paketlerini incele<\/a><\/div>\n","protected":false},"excerpt":{"rendered":"<p>Yapay zeka sunucu kiralama karar\u0131nda ilk soru GPU de\u011fil, modeli kimin \u00e7al\u0131\u015ft\u0131raca\u011f\u0131d\u0131r. API ile \u00e7al\u0131\u015fan uygulamada GPU gerekmez; k\u00fc\u00e7\u00fck modeller i\u015flemcide \u00e7al\u0131\u015f\u0131r, e\u011fitim ise GPU ister. Bellek hesab\u0131 ve kaynak tablosu bu yaz\u0131da.<\/p>\n","protected":false},"author":1,"featured_media":2785,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[6],"tags":[],"class_list":["post-2784","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-sunucu"],"_links":{"self":[{"href":"https:\/\/narhost.com\/blog\/wp-json\/wp\/v2\/posts\/2784","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/narhost.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/narhost.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/narhost.com\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/narhost.com\/blog\/wp-json\/wp\/v2\/comments?post=2784"}],"version-history":[{"count":1,"href":"https:\/\/narhost.com\/blog\/wp-json\/wp\/v2\/posts\/2784\/revisions"}],"predecessor-version":[{"id":2786,"href":"https:\/\/narhost.com\/blog\/wp-json\/wp\/v2\/posts\/2784\/revisions\/2786"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/narhost.com\/blog\/wp-json\/wp\/v2\/media\/2785"}],"wp:attachment":[{"href":"https:\/\/narhost.com\/blog\/wp-json\/wp\/v2\/media?parent=2784"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/narhost.com\/blog\/wp-json\/wp\/v2\/categories?post=2784"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/narhost.com\/blog\/wp-json\/wp\/v2\/tags?post=2784"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}