Meta, yapay zek\u00e2 altyap\u0131s\u0131na y\u00f6nelik yat\u0131r\u0131mlar\u0131n\u0131 s\u00fcrd\u00fcr\u00fcrken AMD teknolojilerini veri merkezlerinde daha geni\u015f \u00f6l\u00e7ekte kullanmaya haz\u0131rlan\u0131yor. \u0130ki \u015firket aras\u0131ndaki i\u015f birli\u011fi; i\u015flemciler, grafik i\u015flemcileri, raf \u00f6l\u00e7ekli sistemler, a\u011f \u00e7\u00f6z\u00fcmleri ve yaz\u0131l\u0131m optimizasyonlar\u0131n\u0131 kaps\u0131yor. AMD\u2019nin a\u00e7\u0131klamas\u0131na g\u00f6re Meta, ilk Helios raf \u00f6l\u00e7ekli sistemlerini veri merkezlerinde \u00e7al\u0131\u015ft\u0131rmaya ba\u015flad\u0131. Bu sistemler, geni\u015f kapsaml\u0131 da\u011f\u0131t\u0131mdan \u00f6nce ger\u00e7ek kullan\u0131m ko\u015fullar\u0131nda de\u011ferlendirilecek.<\/p>\n
Ortakl\u0131\u011f\u0131n bir di\u011fer aya\u011f\u0131n\u0131 Meta\u2019n\u0131n Muse yapay zek\u00e2 ailesini destekleyen bilgi i\u015flem altyap\u0131s\u0131 olu\u015fturuyor. Meta\u2019n\u0131n ki\u015fisel yapay zek\u00e2 arac\u0131s\u0131 olarak tan\u0131mlad\u0131\u011f\u0131 Muse i\u00e7in kullan\u0131lan baz\u0131 \u00f6zel bilgi i\u015flem ortamlar\u0131 AMD EPYC i\u015flemcilerden g\u00fc\u00e7 al\u0131yor. \u015eirketler, bu \u00e7al\u0131\u015fmalar\u0131 gigawatt \u00f6l\u00e7e\u011finde entegre yapay zek\u00e2 sistemleri geli\u015ftirme hedefi do\u011frultusunda derinle\u015ftirirken AMD\u2019nin Meta\u2019n\u0131n hisse senedi arz\u0131na 1 milyar dolarl\u0131k yat\u0131r\u0131m yapaca\u011f\u0131 da duyuruldu.<\/p>\n
Helios, tek bir sunucudan ziyade birden fazla bile\u015fenin birlikte \u00e7al\u0131\u015ft\u0131\u011f\u0131 raf \u00f6l\u00e7ekli bir sistem yakla\u015f\u0131m\u0131n\u0131 temsil ediyor. B\u00f6yle bir yap\u0131da i\u015flem g\u00fcc\u00fc, h\u0131zland\u0131r\u0131c\u0131lar, a\u011f ba\u011flant\u0131lar\u0131 ve sistem yaz\u0131l\u0131mlar\u0131 ortak tasar\u0131m\u0131n par\u00e7alar\u0131 olarak ele al\u0131n\u0131yor. Meta\u2019n\u0131n ilk Helios sistemlerini devreye almas\u0131, teknolojinin b\u00fcy\u00fck \u00f6l\u00e7ekli kullan\u0131ma ge\u00e7meden \u00f6nce test edilmesine ve Meta\u2019n\u0131n i\u015f y\u00fcklerine g\u00f6re de\u011ferlendirilmesine imk\u00e2n sa\u011flayacak.<\/p>\n
AMD EPYC i\u015flemciler de Meta\u2019n\u0131n altyap\u0131s\u0131ndaki \u00f6nemli bile\u015fenlerden biri olmay\u0131 s\u00fcrd\u00fcr\u00fcyor. Meta\u2019n\u0131n bug\u00fcne kadar milyonlarca EPYC i\u015flemciyi sistemlerine entegre etti\u011fi belirtiliyor. \u015eirket ayr\u0131ca AMD\u2019nin 6. Nesil EPYC mimarisi i\u00e7in \u00f6nc\u00fc m\u00fc\u015fteriler aras\u0131nda yer al\u0131yor. Bu t\u00fcr erken d\u00f6nem \u00e7al\u0131\u015fmalar, yeni nesil i\u015flemcilerin veri merkezi gereksinimlerine uygunlu\u011funun de\u011ferlendirilmesine ve sistem tasar\u0131m\u0131n\u0131n ger\u00e7ek i\u015f y\u00fcklerinden elde edilen geri bildirimlerle geli\u015ftirilmesine yard\u0131mc\u0131 oluyor.<\/p>\n
\u0130ki \u015firketin ortak m\u00fchendislik \u00e7al\u0131\u015fmalar\u0131, \u201cVerano\u201d kod adl\u0131 gelecekteki i\u015flemci mimarilerini de kaps\u0131yor. Bu \u00e7al\u0131\u015fmalar\u0131n amac\u0131 yaln\u0131zca i\u015flemci performans\u0131n\u0131 art\u0131rmak de\u011fil; g\u00fc\u00e7 t\u00fcketimi, sistem yo\u011funlu\u011fu ve b\u00fcy\u00fck \u00f6l\u00e7ekli veri merkezi i\u015fletimi gibi konular\u0131 da birlikte ele almak. Ancak yeni mimarilerin teknik \u00f6zellikleri, piyasaya \u00e7\u0131k\u0131\u015f tarihleri ve Meta altyap\u0131s\u0131nda ne \u00f6l\u00e7ekte kullan\u0131laca\u011f\u0131na ili\u015fkin ayr\u0131nt\u0131lar a\u00e7\u0131klamada belirtilmedi.<\/p>\n
Meta, AMD Instinct GPU\u2019lardan yapay zek\u00e2 modeli e\u011fitimi, i\u00e7erik s\u0131ralama ve \u00f6neri sistemleri gibi farkl\u0131 i\u015f y\u00fcklerinde yararlan\u0131yor. Bu g\u00f6revler y\u00fcksek miktarda paralel hesaplama gerektirdi\u011finden, GPU\u2019lar b\u00fcy\u00fck yapay zek\u00e2 modellerinin geli\u015ftirilmesi ve hizmetlerin \u00e7al\u0131\u015ft\u0131r\u0131lmas\u0131 i\u00e7in kritik bir rol \u00fcstleniyor. Meta ile AMD\u2019nin i\u015f birli\u011fi, mevcut donan\u0131m\u0131n kullan\u0131m\u0131n\u0131n yan\u0131 s\u0131ra gelecek nesil h\u0131zland\u0131r\u0131c\u0131lar\u0131n Meta\u2019n\u0131n ihtiya\u00e7lar\u0131na g\u00f6re \u015fekillendirilmesini de i\u00e7eriyor.<\/p>\n
Yapay zek\u00e2 donan\u0131m\u0131n\u0131n performans\u0131, onu kullanan yaz\u0131l\u0131m ara\u00e7lar\u0131n\u0131n deste\u011fine de ba\u011fl\u0131. Bu nedenle AMD ve Meta, ROCm yaz\u0131l\u0131m platformunun PyTorch ile uyumunu ve performans\u0131n\u0131 geli\u015ftirmek i\u00e7in \u00e7al\u0131\u015f\u0131yor. PyTorch, yapay zek\u00e2 modellerinin geli\u015ftirilmesi ve e\u011fitilmesinde yayg\u0131n bi\u00e7imde kullan\u0131lan bir \u00e7er\u00e7eve; ROCm ise AMD h\u0131zland\u0131r\u0131c\u0131lar\u0131n\u0131n bu t\u00fcr i\u015f y\u00fcklerinde kullan\u0131lmas\u0131n\u0131 sa\u011flayan yaz\u0131l\u0131m ekosisteminin temel par\u00e7alar\u0131ndan biri.<\/p>\n
Ortak \u00e7al\u0131\u015fmalar\u0131n a\u00e7\u0131k kaynak projelerine de katk\u0131 sa\u011flamas\u0131 hedefleniyor. AMD ve Meta, Triton ve PyTorch k\u00fct\u00fcphanelerinde iyile\u015ftirmeler \u00fczerinde \u00e7al\u0131\u015f\u0131yor. Bu t\u00fcr geli\u015ftirmeler, geli\u015ftiricilerin AMD donan\u0131mlar\u0131nda modellerini \u00e7al\u0131\u015ft\u0131rmas\u0131n\u0131 kolayla\u015ft\u0131rabilir ve farkl\u0131 donan\u0131m platformlar\u0131 aras\u0131nda ge\u00e7i\u015f yaparken kar\u015f\u0131la\u015f\u0131lan uyumluluk sorunlar\u0131n\u0131 azaltabilir. Bununla birlikte, performans kazan\u0131mlar\u0131n\u0131n uygulamaya ve i\u015f y\u00fck\u00fcne g\u00f6re de\u011fi\u015febilece\u011fi g\u00f6z \u00f6n\u00fcnde bulundurulmal\u0131.<\/p>\n
\u0130\u015f birli\u011fi yaln\u0131zca veri merkezleriyle s\u0131n\u0131rl\u0131 de\u011fil. AMD, Instinct GPU ailesinde Meta\u2019n\u0131n Muse Spark a\u00e7\u0131k modellerine; AMD Ryzen AI destekli bilgisayarlar ve Radeon ekran kartlar\u0131nda ise Muse Glimmer modeline S\u0131f\u0131r\u0131nc\u0131 G\u00fcn deste\u011fi sunmay\u0131 planl\u0131yor. S\u0131f\u0131r\u0131nc\u0131 G\u00fcn deste\u011fi, ilgili modelin veya yaz\u0131l\u0131m bile\u015feninin kullan\u0131ma sunuldu\u011fu d\u00f6nemde donan\u0131m ve yaz\u0131l\u0131m taraf\u0131nda \u00e7al\u0131\u015ft\u0131r\u0131labilmesini ama\u00e7l\u0131yor. Bu destek, modellerin her \u00fcr\u00fcn ve yap\u0131land\u0131rmada ayn\u0131 performans\u0131 g\u00f6sterece\u011fi anlam\u0131na gelmiyor; kullan\u0131m deneyimi donan\u0131m \u00f6zelliklerine ve yaz\u0131l\u0131m s\u00fcr\u00fcmlerine ba\u011fl\u0131 olacak.<\/p>\n
Meta ile AMD aras\u0131ndaki geni\u015fleyen ortakl\u0131k, yapay zek\u00e2 altyap\u0131s\u0131n\u0131n tek bir i\u015flemci veya h\u0131zland\u0131r\u0131c\u0131dan ibaret olmad\u0131\u011f\u0131n\u0131 g\u00f6steriyor. Gigawatt \u00f6l\u00e7e\u011fini hedefleyen sistemlerde enerji, so\u011futma, a\u011f ba\u011flant\u0131lar\u0131, yaz\u0131l\u0131m uyumlulu\u011fu ve donan\u0131m\u0131n birlikte tasarlanmas\u0131 \u00f6nem ta\u015f\u0131yor. \u015eirketlerin silikon seviyesinden raf \u00f6l\u00e7ekli sistemlere ve a\u00e7\u0131k kaynak yaz\u0131l\u0131mlara kadar uzanan \u00e7al\u0131\u015fmalar\u0131, Meta\u2019n\u0131n yapay zek\u00e2 hizmetlerini b\u00fcy\u00fctme hedefini desteklemeyi ama\u00e7l\u0131yor. Yat\u0131r\u0131m\u0131n ve yeni sistemlerin uzun vadeli etkisi ise da\u011f\u0131t\u0131m\u0131n kapsam\u0131, teknik sonu\u00e7lar ve veri merkezi operasyonlar\u0131ndaki verimlilik gibi unsurlar netle\u015ftik\u00e7e daha iyi de\u011ferlendirilebilecek.<\/p>\n