{"id":13045,"date":"2026-08-07T20:50:18","date_gmt":"2026-08-07T17:50:18","guid":{"rendered":"https:\/\/gtmteknoloji.com\/b2b\/xai-colossus-100-bin-gpu-kume-analizi\/"},"modified":"2026-08-07T20:52:58","modified_gmt":"2026-08-07T17:52:58","slug":"xai-colossus-100-bin-gpu-kume-analizi","status":"publish","type":"post","link":"https:\/\/gtmteknoloji.com\/b2b\/xai-colossus-100-bin-gpu-kume-analizi\/","title":{"rendered":"122 G\u00fcnde 100.000 GPU: xAI Colossus K\u00fcmesinin Mimarisi"},"content":{"rendered":"
\n
\n
\n
\n
\n
Ba\u015far\u0131 Hik\u00e2yesi \u00b7 AI S\u00fcperbilgisayar<\/div>\n

122 g\u00fcnde 100.000 GPU<\/em>: xAI Colossus k\u00fcmesi nas\u0131l kuruldu?<\/h1>\n

Elon Musk’\u0131n Memphis’teki AI s\u00fcperbilgisayar\u0131, yaln\u0131zca b\u00fcy\u00fckl\u00fc\u011f\u00fcyle de\u011fil kurulum h\u0131z\u0131yla da dikkat \u00e7ekiyor. K\u00fcmenin yap\u0131 ta\u015f\u0131 Supermicro’nun s\u0131v\u0131 so\u011futmal\u0131 raf\u0131: raf ba\u015f\u0131na 64 NVIDIA H100, sekizerli gruplarda 512 GPU’luk mini k\u00fcmeler. Bu yaz\u0131da mimarinin nas\u0131l katmanland\u0131\u011f\u0131n\u0131 ve kendi projenize ne s\u00f6yledi\u011fini inceliyoruz.<\/p>\n

Supermicro<\/b> \u00d7 NVIDIA<\/b> \u00d7 xAI<\/b> Teknik analiz \u00b7 GTM Teknoloji<\/span><\/div>\n<\/p><\/div>\n
\"Supermicro<\/div>\n<\/p><\/div>\n

G\u00f6rsel: GTM Teknoloji katalo\u011fundaki 4U s\u0131v\u0131 so\u011futmal\u0131 8-GPU HGX sistemi \u2014 Colossus’ta kullan\u0131lan yap\u0131 ta\u015f\u0131n\u0131n ayn\u0131 ailesi.<\/p>\n<\/p><\/div>\n<\/header>\n

\n
100.000<\/b>NVIDIA H100 GPU<\/span><\/div>\n
122 g\u00fcn<\/b>Kurulum s\u00fcresi<\/span><\/div>\n
64<\/b>Raf ba\u015f\u0131na GPU (8 sunucu \u00d7 8)<\/span><\/div>\n
3,6 Tb\/s<\/b>GPU sunucusu ba\u015f\u0131na a\u011f bant geni\u015fli\u011fi<\/span><\/div>\n<\/p><\/div>\n
\n
\n

Neden bu kurulum \u00f6nemli?<\/h2>\n

Colossus’u kayda de\u011fer k\u0131lan yaln\u0131zca 100.000 GPU’luk \u00f6l\u00e7e\u011fi de\u011fil; bu \u00f6l\u00e7e\u011fe 122 g\u00fcnde<\/strong> ula\u015f\u0131lm\u0131\u015f olmas\u0131. Bu h\u0131z, tek tek \u00fcr\u00fcnlerin ba\u015far\u0131s\u0131ndan \u00e7ok bir yap\u0131 ta\u015f\u0131<\/em> yakla\u015f\u0131m\u0131n\u0131n sonucu: ayn\u0131 s\u0131v\u0131 so\u011futmal\u0131 raf tasar\u0131m\u0131 defalarca \u00e7o\u011falt\u0131l\u0131yor, her katman bir \u00f6ncekinin kat\u0131 olarak b\u00fcy\u00fcyor. A\u015fa\u011f\u0131daki tablo, tek bir 4U sunucudan 100.000 GPU’ya kadar \u00f6l\u00e7e\u011fin nas\u0131l katland\u0131\u011f\u0131n\u0131 g\u00f6steriyor.<\/p>\n

\n
4U sunucu8 GPU<\/span><\/b><\/p>\n

S\u0131v\u0131 so\u011futmal\u0131 Supermicro 4U Universal GPU sistemi. NVIDIA HGX H100 tepsisi \u00f6zel s\u0131v\u0131 so\u011futma bloklar\u0131yla so\u011futulur; CPU tepsisinde iki x86 i\u015flemci blo\u011fu bulunur.<\/p>\n<\/div>\n

Raf64 GPU<\/span><\/b><\/p>\n

8 adet 4U sunucu + Supermicro CDU (so\u011futucu da\u011f\u0131t\u0131m \u00fcnitesi) ve yard\u0131mc\u0131 donan\u0131m bir GPU hesaplama raf\u0131n\u0131 olu\u015fturur.<\/p>\n<\/div>\n

Raf grubu512 GPU<\/span><\/b><\/p>\n

Raflar sekizerli gruplar h\u00e2linde konumland\u0131r\u0131l\u0131r; kendi a\u011f\u0131yla birlikte, \u00e7ok daha b\u00fcy\u00fck sistemin i\u00e7inde bir mini k\u00fcme gibi \u00e7al\u0131\u015f\u0131r.<\/p>\n<\/div>\n

Veri salonu25.000 GPU<\/span><\/b><\/p>\n

Ziyaret s\u0131ras\u0131nda d\u00f6rt veri salonu devredeydi; her biri 25.000 GPU bar\u0131nd\u0131r\u0131yordu.<\/p>\n<\/div>\n

Colossus100.000 GPU<\/span><\/b><\/p>\n

D\u00f6rt salonun toplam\u0131. Geni\u015fleme \u00e7al\u0131\u015fmalar\u0131 h\u0131zla s\u00fcr\u00fcyor.<\/p>\n<\/div>\n<\/div><\/div>\n<\/section>\n

\n
\n

Sunucu tasar\u0131m\u0131nda \u00f6ne \u00e7\u0131kan d\u00f6rt karar<\/h2>\n

Colossus’ta kullan\u0131lan 4U Universal GPU sistemi, piyasadaki di\u011fer HGX sunucular\u0131ndan birka\u00e7 somut noktada ayr\u0131\u015f\u0131yor.<\/p>\n

\n
Ba\u015ftan s\u0131v\u0131 so\u011futmal\u0131, tek elden<\/b><\/p>\n

Sekt\u00f6rdeki yayg\u0131n yakla\u015f\u0131m, hava so\u011futmal\u0131 bir tasar\u0131m\u0131n \u00fcst\u00fcne s\u0131v\u0131 so\u011futma eklemektir. Supermicro’nun 4U Universal GPU sistemi ise do\u011frudan s\u0131v\u0131 so\u011futulmak \u00fczere tasarland\u0131 ve bile\u015fenlerin tamam\u0131 tek \u00fcreticiden geliyor.<\/p>\n<\/div>\n

PCIe anahtarlar\u0131 anakartta<\/b><\/p>\n

Neredeyse her HGX sunucusunda kullan\u0131lan d\u00f6rt Broadcom PCIe anahtar\u0131, ayr\u0131 bir kart yerine do\u011frudan Supermicro anakart\u0131na t\u00fcmle\u015fik. Bu anahtarlar da \u00f6zel bir s\u0131v\u0131 so\u011futma blo\u011fuyla so\u011futuluyor \u2014 sekt\u00f6rde e\u015fi olmayan bir tasar\u0131m tercihi.<\/p>\n<\/div>\n

Raftan \u00e7\u0131karmadan servis<\/b><\/p>\n

Sistemler tepsiler \u00fczerinde; 1U raf manifoldu her sunucuya so\u011fuk s\u0131v\u0131y\u0131 getirip \u0131s\u0131nm\u0131\u015f s\u0131v\u0131y\u0131 geri al\u0131yor. H\u0131zl\u0131 ay\u0131r\u0131c\u0131lar (quick disconnect) sayesinde s\u0131v\u0131 ba\u011flant\u0131s\u0131 tek elle s\u00f6k\u00fclebiliyor ve tepsi servis i\u00e7in d\u0131\u015far\u0131 \u00e7ekilebiliyor \u2014 sunucuyu raftan \u00e7\u0131karmaya gerek yok.<\/p>\n<\/div>\n

Alan de\u011fi\u015ftirilebilir CDU<\/b><\/p>\n

Her raf\u0131n alt\u0131ndaki CDU, dev bir \u0131s\u0131 e\u015fanj\u00f6r\u00fc gibi \u00e7al\u0131\u015f\u0131p raftaki t\u00fcm GPU sunucular\u0131n\u0131 besleyen bir s\u0131v\u0131 d\u00f6ng\u00fcs\u00fc kuruyor. Pompalar\u0131 ve g\u00fc\u00e7 kaynaklar\u0131 yedekli; biri ar\u0131zalan\u0131rsa raf\u0131 kapatmadan sahada de\u011fi\u015ftirilebiliyor.<\/p>\n<\/div>\n<\/div><\/div>\n<\/section>\n

\n
\n

Is\u0131 y\u00f6netimi: s\u0131v\u0131 so\u011futma tek ba\u015f\u0131na yetmiyor<\/h2>\n

Bu kurulumun en \u00f6\u011fretici taraf\u0131, s\u0131v\u0131 so\u011futman\u0131n hik\u00e2yenin yaln\u0131zca yar\u0131s\u0131 olmas\u0131. Geri kalan\u0131, raf\u0131n \u00fcretti\u011fi art\u0131k \u0131s\u0131n\u0131n veri salonuna hi\u00e7 verilmeden d\u0131\u015far\u0131 ta\u015f\u0131nmas\u0131 \u00fczerine kurulu.<\/p>\n

\n
Raf\u0131n \u0131s\u0131 n\u00f6tr olmas\u0131 \u015fart<\/b><\/p>\n

S\u0131v\u0131 so\u011futmaya ra\u011fmen sunucularda h\u00e2l\u00e2 fan var: DIMM’ler, g\u00fc\u00e7 kaynaklar\u0131, BMC ve a\u011f kartlar\u0131 gibi bile\u015fenler havayla so\u011futuluyor. Colossus’ta her raf\u0131n veri salonuna g\u00f6re \u0131s\u0131 n\u00f6tr<\/strong> olmas\u0131 hedeflenmi\u015f \u2014 b\u00f6ylece salona devasa klima santralleri kurmak gerekmiyor.<\/p>\n<\/div>\n

Arka kap\u0131 \u0131s\u0131 e\u015fanj\u00f6rleri<\/b><\/p>\n

Sunucu fanlar\u0131 \u00f6nden so\u011fuk havay\u0131 al\u0131p arkadan at\u0131yor; bu hava raf\u0131n arkas\u0131ndaki \u0131s\u0131 e\u015fanj\u00f6r\u00fcnden ge\u00e7iriliyor. Mant\u0131k otomobil radyat\u00f6r\u00fcyle ayn\u0131: i\u00e7inden s\u0131v\u0131 ge\u00e7en petekli bir yap\u0131, \u0131s\u0131y\u0131 tesis su d\u00f6ng\u00fcs\u00fcne aktar\u0131yor.<\/p>\n<\/div>\n

Yaln\u0131zca GPU de\u011fil<\/b><\/p>\n

Arka kap\u0131 e\u015fanj\u00f6rleri sadece s\u0131v\u0131 so\u011futmal\u0131 GPU sunucular\u0131n\u0131n art\u0131k \u0131s\u0131s\u0131n\u0131 de\u011fil; depolama, CPU hesaplama k\u00fcmeleri ve a\u011f bile\u015fenlerinin \u0131s\u0131s\u0131n\u0131 da al\u0131yor. Bu sayede ayn\u0131 salonda s\u0131v\u0131 so\u011futmal\u0131 ve hava so\u011futmal\u0131 ekipman bir arada \u00e7al\u0131\u015fabiliyor.<\/p>\n<\/div>\n

Tesis suyu ve \u00e7illerler<\/b><\/p>\n

So\u011fuk su tesise al\u0131n\u0131p her raftaki CDU’dan dola\u015ft\u0131r\u0131l\u0131yor; GPU’lardan ve arka kap\u0131 e\u015fanj\u00f6rlerinden toplanan \u0131s\u0131 burada tesis su d\u00f6ng\u00fcs\u00fcne aktar\u0131l\u0131yor. Is\u0131nan su d\u0131\u015far\u0131daki \u00e7illerlerde tekrar kullan\u0131labilecek s\u0131cakl\u0131\u011fa indiriliyor.<\/p>\n<\/div>\n<\/div><\/div>\n<\/section>\n

\n
\n

A\u011f: 400GbE’nin dokuz kat\u0131<\/h2>\n

K\u00fcmenin en dikkat \u00e7ekici tercihlerinden biri a\u011f taraf\u0131nda. GPU’lar aras\u0131 RDMA trafi\u011fi i\u00e7in InfiniBand yerine Ethernet<\/strong> se\u00e7ilmi\u015f. Gerek\u00e7e \u00f6l\u00e7ek: Ethernet internetin omurgas\u0131 oldu\u011fu i\u00e7in son derece \u00f6l\u00e7eklenebilir bir teknoloji ve bu b\u00fcy\u00fckl\u00fckteki AI k\u00fcmeleri, daha egzotik teknolojilerin hen\u00fcz ula\u015famad\u0131\u011f\u0131 \u00f6l\u00e7eklere \u00e7\u0131k\u0131yor.<\/p>\n

\n
400GbE<\/b>Optik ba\u011flant\u0131 ba\u015f\u0131na
(yayg\u0131n 1GbE’nin 400 kat\u0131)<\/span><\/div>\n
\u00d79<\/b>Sistem ba\u015f\u0131na ba\u011flant\u0131 say\u0131s\u0131<\/span><\/div>\n
3,6 Tb\/s<\/b>GPU sunucusu ba\u015f\u0131na
toplam bant geni\u015fli\u011fi<\/span><\/div>\n<\/p><\/div>\n

Bant geni\u015fli\u011finin b\u00fcy\u00fck k\u0131sm\u0131n\u0131 GPU RDMA a\u011f\u0131 kullan\u0131yor; her GPU’nun kendi a\u011f kart\u0131 var. Kurulumda NVIDIA BlueField-3 SuperNIC ve Spectrum-X a\u011f teknolojisi kullan\u0131l\u0131yor. CPU’lar ise tamamen ayr\u0131 bir anahtar dokusu \u00fczerinden 400GbE al\u0131yor \u2014 GPU a\u011f\u0131 ve k\u00fcmenin geri kalan\u0131 i\u00e7in ayr\u0131 a\u011flar kurmak, y\u00fcksek ba\u015far\u0131ml\u0131 hesaplamada yayg\u0131n bir tasar\u0131m tercihidir.<\/p>\n

S\u0131radaki sorun: anahtarlar\u0131n so\u011futulmas\u0131<\/h3>\n

Kaynak belgede dile getirilen \u00f6nemli bir tespit: bu tesiste en \u00e7ok g\u00f6ze \u00e7arpan eksik, s\u0131v\u0131 so\u011futmal\u0131 a\u011f anahtarlar\u0131<\/strong>. Modern bir anahtarda optikler, anahtar yongas\u0131ndan daha fazla g\u00fc\u00e7 t\u00fcketebiliyor. Bu \u00f6l\u00e7ekteki kurulumlar\u0131n sekt\u00f6r\u00fc ortak paketlenmi\u015f optiklere (co-packaged optics) ve anahtar taraf\u0131nda da s\u0131v\u0131 so\u011futmaya do\u011fru itmesi bekleniyor.<\/p>\n<\/p><\/div>\n<\/section>\n

\n
\n

K\u00fcmenin geri kalan\u0131: depolama, CPU ve g\u00fc\u00e7<\/h2>\n
\n
Depolama \u2014 flash’a ge\u00e7i\u015f<\/b><\/p>\n

Farkl\u0131 \u00fcreticilerin depolama yaz\u0131l\u0131mlar\u0131 \u00e7al\u0131\u015fsa da donan\u0131m\u0131n neredeyse tamam\u0131 Supermicro; a\u011f\u0131rl\u0131kl\u0131 olarak 2,5″ NVMe yuval\u0131 1U d\u00fc\u011f\u00fcmler. B\u00fcy\u00fck AI k\u00fcmeleri diskten flash’a kay\u0131yor: petabayt ba\u015f\u0131na maliyet daha y\u00fcksek olsa da enerji tasarrufu, performans ve yo\u011funluk toplam sahip olma maliyetinde flash’\u0131 \u00f6ne ge\u00e7iriyor.<\/p>\n<\/div>\n

CPU hesaplama<\/b><\/p>\n

Veri haz\u0131rlama ve i\u015fleme g\u00f6revleri h\u00e2l\u00e2 CPU’da \u00e7ok iyi \u00e7al\u0131\u015f\u0131yor; GPU’lar\u0131 e\u011fitim ve \u00e7\u0131kar\u0131m i\u00e7in ay\u0131rmak mant\u0131kl\u0131. Bunun i\u00e7in hava so\u011futmal\u0131 1U sunucu raflar\u0131 kullan\u0131l\u0131yor \u2014 \u00f6n panelin yakla\u015f\u0131k \u00fc\u00e7te biri hava giri\u015fine ayr\u0131lm\u0131\u015f durumda.<\/p>\n<\/div>\n

G\u00fc\u00e7 dalgalanmas\u0131<\/b><\/p>\n

AI sunucular\u0131 s\u00fcrekli tam g\u00fc\u00e7te \u00e7al\u0131\u015fmaz; i\u015f y\u00fck\u00fc da\u011f\u0131t\u0131ld\u0131k\u00e7a ve sonu\u00e7lar topland\u0131k\u00e7a g\u00fc\u00e7 t\u00fcketimi tepe ve dip yapar. Ekip, bu milisaniyelik dalgalanmalar\u0131 tamponlamak i\u00e7in araya Tesla Megapack<\/strong> bataryalar\u0131 koymu\u015f.<\/p>\n<\/div>\n

Yap\u0131land\u0131r\u0131lm\u0131\u015f kablolama<\/b><\/p>\n

Bu a\u011f yo\u011funlu\u011fu devasa fiber \u00e7ekimi demek. Her fiber do\u011fru uzunlukta kesilip sonland\u0131r\u0131lm\u0131\u015f ve etiketlenmi\u015f; ayr\u0131ca y\u00f6netim aray\u00fczleri ve ortam sens\u00f6rleri i\u00e7in daha d\u00fc\u015f\u00fck h\u0131zl\u0131 ayr\u0131 bir a\u011f bulunuyor.<\/p>\n<\/div><\/div>\n<\/p><\/div>\n<\/section>\n

\n
\n

Bu mimari sizin projeniz i\u00e7in ne anlama geliyor?<\/h2>\n

Colossus’un as\u0131l dersi 100.000 GPU de\u011fil, tekrarlanabilir yap\u0131 ta\u015f\u0131<\/strong> mant\u0131\u011f\u0131. Ayn\u0131 tasar\u0131m tek rafl\u0131k bir ba\u015flang\u0131\u00e7 i\u00e7in de ge\u00e7erli: s\u0131v\u0131 so\u011futmal\u0131 8-GPU sunucu, raf\u0131n alt\u0131nda CDU, arkada \u0131s\u0131 e\u015fanj\u00f6r\u00fc. \u00d6l\u00e7e\u011fi b\u00fcy\u00fctmek, ayn\u0131 blo\u011fu \u00e7o\u011faltmaktan ibaret h\u00e2le geliyor.<\/p>\n

Colossus’ta kullan\u0131lan 4U Universal GPU sisteminin muadili, GTM Teknoloji katalo\u011funda haz\u0131r:<\/p>\n