{"id":116861,"date":"2026-01-06T09:40:21","date_gmt":"2026-01-06T09:40:21","guid":{"rendered":"https:\/\/ekamu.net\/?p=116861"},"modified":"2026-01-06T09:40:21","modified_gmt":"2026-01-06T09:40:21","slug":"gezegenin-en-guclu-ai-platformu-karsinizda-nvidia-rubin-ve-detaylari","status":"publish","type":"post","link":"https:\/\/ekamu.net\/index.php\/2026\/01\/06\/gezegenin-en-guclu-ai-platformu-karsinizda-nvidia-rubin-ve-detaylari\/","title":{"rendered":"Gezegenin en g\u00fc\u00e7l\u00fc AI platformu: Kar\u015f\u0131n\u0131zda Nvidia Rubin ve detaylar\u0131"},"content":{"rendered":"

<\/span> Nvidia<\/strong>, yapay zeka odakl\u0131 veri merkezleri i\u00e7in geli\u015ftirdi\u011fi Rubin<\/strong> platformunu resmen duyurdu. \u015eirketin mevcut Blackwell<\/strong> mimarisine k\u0131yasla 5 kata varan performans<\/strong> art\u0131\u015f\u0131 sunan Rubin, Nvidia\u2019n\u0131n bug\u00fcne kadar geli\u015ftirdi\u011fi en kapsaml\u0131 ve en g\u00fc\u00e7l\u00fc yapay zeka platformu olarak konumlan\u0131yor. <\/figure>\n

Alt\u0131 farkl\u0131 \u00e7ipten olu\u015fan dev platform<\/b><\/p>\n

<\/span> Rubin platformu, \u00fcretimden \u00e7\u0131km\u0131\u015f ve Nvidia laboratuvarlar\u0131nda test a\u015famas\u0131nda olan toplam alt\u0131 ayr\u0131 \u00e7ipten olu\u015fuyor. Bu bile\u015fenler Rubin GPU, Vera CPU, NVLink 6 Switch<\/strong>, ConnectX-9 ve BlueField-4<\/strong> a\u011f \u00e7\u00f6z\u00fcmleri ile Spectrum-X 102.4T CPO<\/strong> silikon fotonik altyap\u0131s\u0131ndan olu\u015fuyor. T\u00fcm bu bile\u015fenler DGX, HGX ve MGX<\/strong> sistemlerinde birlikte \u00e7al\u0131\u015facak \u015fekilde tasarland\u0131. <\/figure>\n
<\/span> Veri merkezlerinin merkezinde ise Nvidia Vera Rubin Superchip<\/strong> yer al\u0131yor. Bu s\u00fcper \u00e7ip, iki Rubin GPU, bir Vera CPU ve y\u00fcksek kapasiteli HBM4 ile LPDDR5x bellek yap\u0131land\u0131rmalar\u0131n\u0131 bir araya getiriyor. <\/figure>\n

Rubin GPU<\/b><\/p>\n

<\/span> Rubin GPU, her biri ba\u011f\u0131ms\u0131z i\u015flem ve tensor \u00e7ekirdeklerine sahip iki kal\u0131pl\u0131 tasar\u0131mla geliyor. Yapay zeka i\u015f y\u00fckleri i\u00e7in \u00f6zel olarak geli\u015ftirilen bu GPU, NVFP4 format\u0131nda 50 PFLOPs<\/strong> \u00e7\u0131kar\u0131m ve 35 PFLOPs e\u011fitim<\/strong> performans\u0131 sunuyor. Bu de\u011ferler, Blackwell\u2019e k\u0131yasla s\u0131ras\u0131yla 5 kat ve 3,5 kat art\u0131\u015f<\/strong> anlam\u0131na geliyor. <\/figure>\n

Bellek taraf\u0131nda ise Rubin GPU, HBM4<\/strong> teknolojisine ge\u00e7i\u015f yap\u0131yor. \u00c7ip ba\u015f\u0131na 22 TB\/s<\/strong> bellek bant geni\u015fli\u011fi sunan bu yap\u0131, Blackwell\u2019e g\u00f6re 2,8 kat art\u0131\u015f<\/strong> sa\u011fl\u0131yor. Ayr\u0131ca her CPU i\u00e7in 3,6 TB\/s NVLink bant geni\u015fli\u011fi sunuluyor ve bu da \u00f6nceki neslin iki kat\u0131 seviyesinde.<\/p>\n

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

<\/span> Rubin platformunun i\u015flemci taraf\u0131nda yer alan Vera CPU<\/strong>, Nvidia\u2019n\u0131n \u00f6zel olarak geli\u015ftirdi\u011fi Olympus <\/strong>kod adl\u0131 yeni nesil Arm <\/strong>mimarisine dayan\u0131yor. Toplam 88 \u00e7ekirdek ve 176 i\u015f par\u00e7ac\u0131\u011f\u0131 <\/strong>sunan bu i\u015flemci, Nvidia Spatial Multi-Threading teknolojisini kullan\u0131yor. <\/figure>\n

Vera CPU 1,8 TB\/s NVLink-C2C tutarl\u0131 bellek ba\u011flant\u0131s\u0131, 1,5 TB sistem belle\u011fi (Grace\u2019e g\u00f6re 3 kat art\u0131\u015f) ve 1,2 TB\/s LPDDR5x bellek bant geni\u015fli\u011fi ile dikkat \u00e7ekiyor. T\u00fcm bu yap\u0131, veri i\u015fleme, s\u0131k\u0131\u015ft\u0131rma ve CI\/CD i\u015f y\u00fcklerinde Grace mimarisine k\u0131yasla 2 kat performans art\u0131\u015f\u0131 sa\u011fl\u0131yor. Ayr\u0131ca platform, raf \u00f6l\u00e7e\u011finde gizli hesaplama deste\u011fi sunuyor.<\/p>\n

NVLink 6 ve a\u011f altyap\u0131s\u0131<\/b><\/p>\n

<\/span> <\/figure>\n<\/p>\n

Rubin platformunda kullan\u0131lan NVLink 6 Switch<\/strong>, 400G SerDes altyap\u0131s\u0131 sayesinde CPU ba\u015f\u0131na 3,6 TB\/s<\/strong> t\u00fcmden t\u00fcm ba\u011flant\u0131 bant geni\u015fli\u011fi sa\u011fl\u0131yor. Toplamda 28,8 TB\/s<\/strong> a\u011f kapasitesine ula\u015fan bu anahtarlar, 14,4 TFLOPS FP8<\/strong> a\u011f i\u00e7i i\u015flem g\u00fcc\u00fc sunuyor ve tamamen s\u0131v\u0131 so\u011futmal\u0131<\/strong> olarak tasarlanm\u0131\u015f durumda.<\/p>\n

<\/span> <\/figure>\n<\/p>\n

A\u011f taraf\u0131nda ConnectX-9 SuperNIC<\/strong>, 1,6 TB\/s bant geni\u015fli\u011fi<\/strong>, 200G PAM4 SerDes, programlanabilir RDMA ve geli\u015fmi\u015f g\u00fcvenlik \u00f6zellikleriyle \u00f6ne \u00e7\u0131k\u0131yor. BlueField-4 DPU<\/strong> ise 64 \u00e7ekirdekli Grace CPU\u2019yu ConnectX-9 ile entegre ederek \u00f6nceki nesle k\u0131yasla 2 kat a\u011f<\/strong>, 6 kat i\u015flem<\/strong> ve 3 kat bellek bant geni\u015fli\u011fi<\/strong> sunuyor.<\/p>\n

NVL72 raf sistemi s\u0131n\u0131rlar\u0131 zorluyor<\/b><\/p>\n

<\/span> <\/figure>\n<\/p>\n

Nvidia\u2019n\u0131n bu t\u00fcm unsurlar\u0131 tek bir platformda, yani Nvidia Vera Rubin NVL72<\/strong> raf sisteminde bir araya geliyor. Bu yap\u0131, Blackwell tabanl\u0131 sistemlere g\u00f6re 5 kat NVFP4 \u00e7\u0131kar\u0131m<\/strong> (3,6 EFLOPS) ve 3,5 kat NVFP4<\/strong> e\u011fitim<\/strong> performans\u0131 (2,5 EFLOPS) sa\u011fl\u0131yor. Bellek taraf\u0131nda LPDDR5x<\/strong> kapasitesi 54 TB<\/strong>, HBM4<\/strong> kapasitesi 20,7 TB<\/strong> seviyesine \u00e7\u0131karken HBM4 bant geni\u015fli\u011fi 1,6 PB\/s ile 2,8 kat art\u0131\u015f sunuyor. \u00d6l\u00e7ekleme bant geni\u015fli\u011fi ise 260 TB\/s<\/strong> ile iki kat\u0131na ula\u015f\u0131yor.<\/p>\n

Nvidia, Rubin ile birlikte Spectrum-X Ethernet Co-Packaged Optics \u00e7\u00f6z\u00fcm\u00fcn\u00fc de duyurdu. 102,4 Tb\/s \u00f6l\u00e7eklenebilir anahtarlama altyap\u0131s\u0131 sunan bu sistem, 200G silikon fotonik entegrasyonu sayesinde \u00f6l\u00e7ek alt\u0131nda %95 etkin bant geni\u015fli\u011fi sa\u011fl\u0131yor. Nvidia\u2019ya g\u00f6re bu \u00e7\u00f6z\u00fcm 5 kat verimlilik, 10 kat g\u00fcvenilirlik ve 5 kat daha y\u00fcksek uygulama \u00e7al\u0131\u015fma s\u00fcresi sunuyor.<\/p>\n

<\/span> <\/figure>\n<\/p>\n

Rubin SuperPOD i\u00e7in tan\u0131t\u0131lan Inference Context Memory Storage platformu ise gigaskala \u00e7\u0131kar\u0131m senaryolar\u0131na odaklan\u0131yor ve Dynamo, NIXL ve DOCA gibi Nvidia yaz\u0131l\u0131m \u00e7\u00f6z\u00fcmleriyle tam entegre \u00e7al\u0131\u015f\u0131yor. Nvidia, Rubin platformunu 8 adet NVL72 raf<\/strong> i\u00e7eren yeni nesil DGX SuperPOD<\/strong> sistemlerinde konumland\u0131r\u0131rken daha geni\u015f veri merkezleri i\u00e7in DGX Rubin NVL8<\/strong> se\u00e7ene\u011fini de sunacak.<\/p>\n

Nvidia\u2019ya g\u00f6re Rubin platformu, Blackwell GB200\u2019e k\u0131yasla \u00e7\u0131kar\u0131m token maliyetini 10 kat<\/strong>, MoE model e\u011fitiminde kullan\u0131lan GPU say\u0131s\u0131n\u0131 ise 4 kat azalt\u0131yor<\/strong>. Geni\u015f bir i\u015f orta\u011f\u0131 ekosistemiyle desteklenen Rubin platformu halihaz\u0131rda tam \u00fcretim a\u015famas\u0131nda bulunuyor ve ilk m\u00fc\u015fterilerin \u00e7ipleri bu y\u0131l i\u00e7inde teslim almaya<\/strong> ba\u015flamas\u0131 planlan\u0131yor.<\/p>\n\n

<\/span><\/p>\n

Kaynak :\u00a0https:\/\/www.donanimhaber.com\/nvidia-rubin-sahneye-cikti-gezegenin-en-guclu-ai-platformu–200521<\/span><\/p>\n","protected":false},"excerpt":{"rendered":"

Nvidia, yapay zeka odakl\u0131 veri merkezleri i\u00e7in geli\u015ftirdi\u011fi Rubin platformunu resmen duyurdu. \u015eirketin mevcut Blackwell mimarisine k\u0131yasla 5 kata varan performans art\u0131\u015f\u0131 sunan Rubin, Nvidia\u2019n\u0131n bug\u00fcne kadar geli\u015ftirdi\u011fi en kapsaml\u0131 ve en g\u00fc\u00e7l\u00fc …<\/p>\n","protected":false},"author":1,"featured_media":116862,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[6978,2654,104,284,641],"class_list":["post-116861","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-teknoloji","tag-5-kat","tag-bant-genisligi","tag-kat","tag-nvidia","tag-sistem"],"_links":{"self":[{"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/posts\/116861","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/comments?post=116861"}],"version-history":[{"count":1,"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/posts\/116861\/revisions"}],"predecessor-version":[{"id":116872,"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/posts\/116861\/revisions\/116872"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/media\/116862"}],"wp:attachment":[{"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/media?parent=116861"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/categories?post=116861"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/tags?post=116861"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}