{"id":118683,"date":"2026-01-27T10:24:13","date_gmt":"2026-01-27T10:24:13","guid":{"rendered":"https:\/\/ekamu.net\/?p=118683"},"modified":"2026-01-27T10:24:13","modified_gmt":"2026-01-27T10:24:13","slug":"nvidiaya-optik-rakip-neurophostan-56-ghzlik-fotonik-yapay-zeka-cipi","status":"publish","type":"post","link":"https:\/\/ekamu.net\/index.php\/2026\/01\/27\/nvidiaya-optik-rakip-neurophostan-56-ghzlik-fotonik-yapay-zeka-cipi\/","title":{"rendered":"Nvidia\u2019ya optik rakip: Neurophos\u2019tan 56 GHz\u2019lik fotonik yapay zeka \u00e7ipi"},"content":{"rendered":"

<\/span> Merkezi ABD\u2019de bulunan ve Bill Gates taraf\u0131ndan desteklenen Neurophos<\/strong>, silikon fotonik tabanl\u0131 yeni bir optik i\u015flem birimi<\/strong> (OPU) geli\u015ftirdi\u011fini duyurdu. \u015eirket, bu \u00e7ipin \u00f6zellikle FP4 ve INT4 hesaplama i\u015f y\u00fcklerinde Nvidia\u2019n\u0131n en yeni Vera Rubin NVL72<\/strong> yapay zeka s\u00fcper bilgisayar\u0131ndan yakla\u015f\u0131k 10 kat daha y\u00fcksek performans<\/strong> sundu\u011funu bildiriyor. \u00dcstelik bunu benzer g\u00fc\u00e7 t\u00fcketimiyle yapt\u0131\u011f\u0131n\u0131 iddia ediyor. <\/figure>\n

Radikal mimari de\u011fi\u015fim<\/b><\/p>\n

Neurophos\u2019un fark yaratt\u0131\u011f\u0131 nokta, optik hesaplamay\u0131 kulland\u0131\u011f\u0131 matris boyutunda yat\u0131yor. \u015eirketin CEO\u2019su Patrick Bowen\u2019a g\u00f6re \u00e7ip \u00fczerinde 1.000 x 1.000<\/strong> boyutunda tek bir fotonik sens\u00f6r<\/strong> yer al\u0131yor. Bu yap\u0131, g\u00fcn\u00fcm\u00fczde \u00e7o\u011fu yapay zeka GPU\u2019sunda kullan\u0131lan 256 x 256 matrislere k\u0131yasla yakla\u015f\u0131k 15 kat daha b\u00fcy\u00fck bir hesaplama alan\u0131<\/strong> sunuyor. Daha b\u00fcy\u00fck matris sayesinde ayn\u0131 anda \u00e7ok daha fazla \u00e7arpma ve toplama i\u015flemi yap\u0131labiliyor.<\/p>\n

Bu mimarinin \u00f6n\u00fcndeki en b\u00fcy\u00fck engel, optik transist\u00f6rlerin fiziksel boyutlar\u0131yd\u0131. Bowen\u2019\u0131n aktard\u0131\u011f\u0131na g\u00f6re, g\u00fcn\u00fcm\u00fczde silikon fotonik fabrikalar\u0131nda \u00fcretilen optik transist\u00f6rlerin uzunlu\u011fu yakla\u015f\u0131k 2 milimetre seviyesinde bulunuyor. Bu da bir \u00e7ip \u00fczerine yeterli say\u0131da transist\u00f6r yerle\u015ftirmeyi neredeyse imkans\u0131z hale getiriyor. Neurophos ise bu sorunu a\u015farak mevcut \u00e7\u00f6z\u00fcmlere k\u0131yasla yakla\u015f\u0131k 10.000 kat daha k\u00fc\u00e7\u00fck optik transist\u00f6rler geli\u015ftirmeyi<\/strong> ba\u015fard\u0131. Bu k\u00fc\u00e7\u00fclme, optik hesaplaman\u0131n dijital CMOS teknolojileriyle rekabet edebilecek bir yo\u011funlu\u011fa ula\u015fmas\u0131n\u0131n \u00f6n\u00fcn\u00fc a\u00e7\u0131yor.<\/p>\n

<\/span> Neurophos\u2019un ilk nesil h\u0131zland\u0131r\u0131c\u0131s\u0131, \u201coptik e\u015fde\u011fer\u201d olarak tan\u0131mlanan tek bir tens\u00f6r \u00e7ekirde\u011fini bar\u0131nd\u0131r\u0131yor ve bu yap\u0131 yakla\u015f\u0131k 25 milimetrekarelik bir alan<\/strong> kapl\u0131yor. Ka\u011f\u0131t \u00fczerinde bak\u0131ld\u0131\u011f\u0131nda bu, 576 tens\u00f6r \u00e7ekirde\u011fi i\u00e7eren Nvidia Vera Rubin \u00e7ipinin olduk\u00e7a gerisinde g\u00f6r\u00fcn\u00fcyor. Ancak fark, Neurophos\u2019un fotonik yongay\u0131 nas\u0131l kulland\u0131\u011f\u0131nda yat\u0131yor. <\/figure>\n

\u015eirketin Tulkas T100<\/strong> ad\u0131n\u0131 verdi\u011fi ilk OPU\u2019su, 56 GHz<\/strong> gibi son derece y\u00fcksek bir saat h\u0131z\u0131nda \u00e7al\u0131\u015f\u0131yor. Y\u00fcksek saat h\u0131z\u0131 ve b\u00fcy\u00fck matris yap\u0131s\u0131 birle\u015fti\u011finde ka\u011f\u0131t \u00fczerinde zay\u0131f g\u00f6r\u00fcnen donan\u0131m\u0131n Nvidia\u2019n\u0131n g\u00fc\u00e7l\u00fc yapay zeka GPU\u2019lar\u0131n\u0131 geride b\u0131rakabildi\u011fi belirtiliyor. \u015eirket, \u00e7ipinin saniyede 235 Peta Operasyon (POPS) zirve performans\u0131 sa\u011flayabilece\u011fini ve 675 watt g\u00fc\u00e7 t\u00fcketece\u011fini belirtiyor. Buna kar\u015f\u0131l\u0131k, B200 \u00e7ipi 1.000 watt g\u00fc\u00e7te 9 POPS performans sa\u011flayabiliyor.<\/p>\n

Mevcut \u00fcretim teknolojileriyle uyumlu<\/b><\/p>\n

<\/span> Neurophos\u2019un bir di\u011fer kritik iddias\u0131 ise \u00fcretim taraf\u0131nda geliyor. Bowen\u2019a g\u00f6re \u015firket, optik transist\u00f6rlerini mevcut yar\u0131 iletken \u00fcretim teknolojileriyle geli\u015ftirdi. Bu da teoride Intel veya TSMC gibi b\u00fcy\u00fck yar\u0131 iletken \u00fcreticilerinin devreye al\u0131narak seri \u00fcretime ge\u00e7ilebilece\u011fi anlam\u0131na geliyor. Ancak \u00e7ipler \u015fu anda test a\u015famas\u0131nda ve hacimli \u00fcretimin 2028 y\u0131l\u0131ndan \u00f6nce ba\u015flamas\u0131 beklenmiyor<\/strong>. Bununla birlikte, \u00e7\u00f6z\u00fclmesi gereken \u00f6nemli teknik zorluklar da bulunuyor. \u00d6zellikle \u00e7ok say\u0131da vekt\u00f6r i\u015flem birimi ihtiyac\u0131 ve statik bellek (SRAM) gereksinimleri, \u015firketin \u00f6n\u00fcndeki ba\u015fl\u0131ca m\u00fchendislik sorunlar\u0131 aras\u0131nda yer al\u0131yor. <\/figure>\n

\u00d6te yandan silikon fotonik, son d\u00f6nemde yar\u0131 iletken d\u00fcnyas\u0131n\u0131n en \u00e7ok ilgi g\u00f6ren alanlar\u0131ndan biri haline gelmi\u015f durumda. Nvidia, Rubin platformunda Spectrum-X Ethernet fotonik anahtar sistemlerini halihaz\u0131rda kullan\u0131rken AMD\u2019nin de silikon fotonik ara\u015ft\u0131rmalar\u0131na odaklanan 280 milyon dolarl\u0131k bir merkez kurmaya haz\u0131rland\u0131\u011f\u0131 biliniyor. Neurophos\u2019un bu son hamlesi, fotonik tabanl\u0131 hesaplaman\u0131n art\u0131k teoriden \u00e7\u0131k\u0131p pratik uygulamalara yakla\u015fmaya ba\u015flad\u0131\u011f\u0131n\u0131 g\u00f6steriyor.<\/p>\n

Fotonik hesaplama nedir, nas\u0131l \u00e7al\u0131\u015f\u0131r?<\/b><\/p>\n

Fotonik hesaplama, geleneksel yar\u0131 iletkenlerde oldu\u011fu gibi elektronlar yerine \u0131\u015f\u0131k<\/strong> (fotonlar) kullan\u0131larak veri i\u015flenmesine dayan\u0131yor. Klasik dijital \u00e7iplerde bilgi, transist\u00f6rler \u00fczerinden ge\u00e7en elektrik ak\u0131m\u0131yla temsil edilirken fotonik tabanl\u0131 sistemlerde bu g\u00f6rev lazer kaynaklar\u0131, dalga k\u0131lavuzlar\u0131 ve optik mod\u00fclat\u00f6rler arac\u0131l\u0131\u011f\u0131yla ger\u00e7ekle\u015ftiriliyor. Hesaplama i\u015flemleri, \u0131\u015f\u0131\u011f\u0131n genli\u011fi, faz\u0131 veya dalga boyu gibi fiziksel \u00f6zellikleri kullan\u0131larak yap\u0131l\u0131yor.<\/p>\n

Bu yakla\u015f\u0131m \u00f6zellikle matris \u00e7arp\u0131mlar\u0131 gibi yapay zekan\u0131n temelini olu\u015fturan i\u015flemlerde avantaj sa\u011fl\u0131yor. I\u015f\u0131k, bir optik yap\u0131dan ge\u00e7erken ayn\u0131 anda \u00e7ok say\u0131da veriyi paralel bi\u00e7imde i\u015fleyebildi\u011fi<\/strong> i\u00e7in fotonik devreler teoride son derece y\u00fcksek paralellik ve d\u00fc\u015f\u00fck gecikme sunabiliyor. Neurophos\u2019un 1.000 x 1.000 boyutundaki optik matris yap\u0131s\u0131 da tam olarak bu prensip \u00fczerine kuruluyor.<\/p>\n

Geleneksel CMOS tabanl\u0131 \u00e7ipler, transist\u00f6r \u00f6l\u00e7ekleri k\u00fc\u00e7\u00fcld\u00fck\u00e7e \u0131s\u0131, g\u00fc\u00e7 t\u00fcketimi ve veri aktar\u0131m h\u0131zlar\u0131 gibi fiziksel s\u0131n\u0131rlara giderek daha fazla tak\u0131lmaya ba\u015flad\u0131. \u00d6zellikle b\u00fcy\u00fck dil modelleri ve yapay zeka sistemleri, devasa matris i\u015flemleri nedeniyle hem y\u00fcksek enerji t\u00fcketiyor hem de veri merkezlerinde ciddi so\u011futma sorunlar\u0131na yol a\u00e7\u0131yor.<\/p>\n

Fotonik hesaplama ise bu noktada \u00f6ne \u00e7\u0131k\u0131yor. I\u015f\u0131kla \u00e7al\u0131\u015fan sistemler, elektriksel diren\u00e7 kaynakl\u0131 \u0131s\u0131 \u00fcretimini b\u00fcy\u00fck \u00f6l\u00e7\u00fcde ortadan kald\u0131r\u0131yor ve veri iletiminde bak\u0131r ba\u011flant\u0131lara k\u0131yasla \u00e7ok daha y\u00fcksek bant geni\u015fli\u011fi sunuyor. Bu da ayn\u0131 g\u00fc\u00e7 zarf\u0131 i\u00e7inde daha fazla hesaplama yap\u0131labilmesini m\u00fcmk\u00fcn k\u0131l\u0131yor.<\/p>\n\n

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

Kaynak :\u00a0https:\/\/www.donanimhaber.com\/nvidia-ya-optik-rakip-neurophos-tan-ilginc-fotonik-ai-cipi–201329<\/span><\/p>\n","protected":false},"excerpt":{"rendered":"

Merkezi ABD\u2019de bulunan ve Bill Gates taraf\u0131ndan desteklenen Neurophos, silikon fotonik tabanl\u0131 yeni bir optik i\u015flem birimi (OPU) geli\u015ftirdi\u011fini duyurdu. \u015eirket, bu \u00e7ipin \u00f6zellikle FP4 ve INT4 hesaplama i\u015f y\u00fcklerinde Nvidia\u2019n\u0131n en yeni Vera Rubin …<\/p>\n","protected":false},"author":1,"featured_media":118684,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[498,5087,7096,2452,106],"class_list":["post-118683","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-teknoloji","tag-cip","tag-hesaplama","tag-matris","tag-optik","tag-yapi"],"_links":{"self":[{"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/posts\/118683","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=118683"}],"version-history":[{"count":1,"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/posts\/118683\/revisions"}],"predecessor-version":[{"id":118688,"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/posts\/118683\/revisions\/118688"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/media\/118684"}],"wp:attachment":[{"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/media?parent=118683"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/categories?post=118683"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/tags?post=118683"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}