\u00c7inli \u00e7ip \u00fcreticisi Iluvatar CoreX, yapay zek\u00e2 h\u0131zland\u0131r\u0131c\u0131s\u0131 Tiangai 300\u2019\u00fcn belirli test ko\u015fullar\u0131nda Nvidia H200\u2019den daha y\u00fcksek performans g\u00f6sterdi\u011fini a\u00e7\u0131klad\u0131. \u015eanghay\u2019da bas\u0131n mensuplar\u0131na y\u00f6nelik d\u00fczenlenen g\u00f6sterimde payla\u015f\u0131lan sonu\u00e7lara g\u00f6re Tiangai 300, standart test yap\u0131land\u0131rmas\u0131nda kart ba\u015f\u0131na verim bak\u0131m\u0131ndan H200\u2019\u00fcn 1,05 kat\u0131na ula\u015ft\u0131. Bu oran, s\u00f6z konusu \u00f6l\u00e7\u00fcmde Tiangai 300 lehine yakla\u015f\u0131k y\u00fczde 5\u2019lik bir fark bulundu\u011funu ifade ediyor.<\/p>\n
\u015eirket, kar\u015f\u0131la\u015ft\u0131rmay\u0131 ger\u00e7ek yapay zek\u00e2 i\u015f y\u00fckleri \u00fczerinden ger\u00e7ekle\u015ftirdi\u011fini belirtti. Bununla birlikte, a\u00e7\u0131klanan sonu\u00e7lar\u0131n kapsam\u0131n\u0131 do\u011fru de\u011ferlendirmek gerekiyor: Performans, kullan\u0131lan model ve yaz\u0131l\u0131ma, g\u00f6rev t\u00fcr\u00fcne, sistem yap\u0131land\u0131rmas\u0131na ve \u00f6l\u00e7\u00fcm y\u00f6ntemine g\u00f6re de\u011fi\u015febilir. Bu nedenle bildirilen fark, her uygulamada Tiangai 300\u2019\u00fcn H200\u2019den daha h\u0131zl\u0131 olaca\u011f\u0131n\u0131 kan\u0131tlam\u0131yor; yaln\u0131zca \u015firketin payla\u015ft\u0131\u011f\u0131 belirli test senaryolar\u0131ndaki sonucu g\u00f6steriyor.<\/p>\n
Standart test d\u00fczeninde hesaplama bile\u015fenleri ile hizmet bile\u015fenleri birbirinden ayr\u0131 konumland\u0131r\u0131ld\u0131. Iluvatar CoreX\u2019in a\u00e7\u0131klamas\u0131na g\u00f6re bu yap\u0131land\u0131rmada tek bir Tiangai 300 kart\u0131n\u0131n verimi, tek bir Nvidia H200 kart\u0131n\u0131n 1,05 kat\u0131 olarak \u00f6l\u00e7\u00fcld\u00fc. Dolay\u0131s\u0131yla kar\u015f\u0131la\u015ft\u0131rma, toplam sistem g\u00fcc\u00fcnden ziyade kart ba\u015f\u0131na elde edilen sonu\u00e7 \u00fczerinden yap\u0131ld\u0131.<\/p>\n
Burada kullan\u0131lan \u201cverim\u201d ifadesi, kart\u0131n test s\u0131ras\u0131nda belirlenen i\u015f y\u00fck\u00fcn\u00fc ne \u00f6l\u00e7\u00fcde ger\u00e7ekle\u015ftirebildi\u011fini anlat\u0131yor. Ancak \u015firketin payla\u015ft\u0131\u011f\u0131 bilgilerde testin t\u00fcm teknik ayr\u0131nt\u0131lar\u0131 yer alm\u0131yor. \u00d6rne\u011fin g\u00f6revlerin tam olarak nas\u0131l tan\u0131mland\u0131\u011f\u0131, kullan\u0131lan yaz\u0131l\u0131m s\u00fcr\u00fcmleri, \u00f6l\u00e7\u00fcm s\u00fcresi ve sonu\u00e7lar\u0131n hangi performans \u00f6l\u00e7\u00fct\u00fcyle hesapland\u0131\u011f\u0131 a\u00e7\u0131klamada ayr\u0131nt\u0131l\u0131 bi\u00e7imde belirtilmedi. Bu bilgiler olmadan 1,05 katl\u0131k sonucu ba\u015fka testlerle do\u011frudan kar\u015f\u0131la\u015ft\u0131rmak veya ba\u011f\u0131ms\u0131z bi\u00e7imde yeniden \u00fcretmek m\u00fcmk\u00fcn olmayabilir.<\/p>\n
Y\u00fczde 5\u2019lik fark, tek ba\u015f\u0131na her ko\u015fulda daha iyi performans anlam\u0131na gelmez.<\/strong> Yapay zek\u00e2 h\u0131zland\u0131r\u0131c\u0131lar\u0131n\u0131n sonu\u00e7lar\u0131; modelin donan\u0131ma ne kadar iyi uyarland\u0131\u011f\u0131, veri aktar\u0131m\u0131n\u0131n ne kadar verimli oldu\u011fu ve sistemin yaz\u0131l\u0131m y\u0131\u011f\u0131n\u0131ndan nas\u0131l yararland\u0131\u011f\u0131 gibi unsurlardan etkilenir. Farkl\u0131 bir model, g\u00f6rev veya donan\u0131m d\u00fczeninde performans s\u0131ralamas\u0131 de\u011fi\u015febilir. Bu nedenle test sonucu, kar\u015f\u0131la\u015ft\u0131rman\u0131n yap\u0131ld\u0131\u011f\u0131 senaryoyla birlikte okunmal\u0131d\u0131r.<\/p>\n Tiangai 300, yaln\u0131zca standart test yap\u0131land\u0131rmas\u0131nda de\u011fil, \u00e7ok kipli yapay zek\u00e2 i\u015f y\u00fcklerinde de de\u011ferlendirildi. \u00c7ok kipli sistemler, metin, g\u00f6r\u00fcnt\u00fc ve video gibi farkl\u0131 veri t\u00fcrlerini birlikte i\u015fleyebilir. G\u00f6sterimde \u00f6ne \u00e7\u0131kan g\u00f6revler aras\u0131nda metin istemlerinden g\u00f6r\u00fcnt\u00fc ve video \u00fcretmek bulunuyordu. Bu t\u00fcr i\u015flemler, modelin girdiyi yorumlamas\u0131ndan g\u00f6r\u00fcnt\u00fc veya video \u00e7\u0131kt\u0131s\u0131n\u0131n haz\u0131rlanmas\u0131na kadar birden fazla hesaplama a\u015famas\u0131 i\u00e7erdi\u011fi i\u00e7in h\u0131zland\u0131r\u0131c\u0131 \u00fczerinde yo\u011fun y\u00fck olu\u015fturabilir.<\/p>\n Iluvatar CoreX, bu karma\u015f\u0131k g\u00f6revlerde de Tiangai 300\u2019\u00fcn Nvidia H200\u2019\u00fcn \u00f6n\u00fcnde sonu\u00e7 verdi\u011fini bildirdi. Testlerde b\u00fcy\u00fck dil modellerinden yararlan\u0131ld\u0131 ve Zhipu AI taraf\u0131ndan geli\u015ftirilen GLM-5.3 de kullan\u0131lan modeller aras\u0131nda yer ald\u0131. B\u00fcy\u00fck dil modelleri metni anlay\u0131p \u00fcretebilir; uygun model ve yaz\u0131l\u0131m bile\u015fenleriyle daha geni\u015f \u00e7ok kipli i\u015f ak\u0131\u015flar\u0131n\u0131n par\u00e7as\u0131 h\u00e2line gelebilir. Ancak metinden g\u00f6r\u00fcnt\u00fc ya da video \u00fcretimi, yaln\u0131zca dil modelinin \u00e7al\u0131\u015fmas\u0131ndan ibaret de\u011fildir. G\u00f6rev zincirinin farkl\u0131 a\u015famalar\u0131, kullan\u0131lan ek modeller ve yaz\u0131l\u0131m optimizasyonlar\u0131 da toplam s\u00fcreyi ve elde edilen sonucu etkileyebilir.<\/p>\n Bu nedenle Tiangai 300\u2019\u00fcn ilgili testlerde H200\u2019\u00fc ge\u00e7ti\u011fine ili\u015fkin a\u00e7\u0131klama, belirli model ve i\u015f ak\u0131\u015f\u0131 ko\u015fullar\u0131yla s\u0131n\u0131rl\u0131 bi\u00e7imde yorumlanmal\u0131. \u015eirket, performans kar\u015f\u0131la\u015ft\u0131rmas\u0131n\u0131n sonucunu payla\u015fm\u0131\u015f olsa da her bir g\u00f6rev i\u00e7in ayr\u0131nt\u0131l\u0131 s\u00fcre, \u00e7\u0131kt\u0131 miktar\u0131, g\u00fc\u00e7 t\u00fcketimi ve hata pay\u0131 gibi verilerin tamam\u0131 haberde yer alm\u0131yor. Bu t\u00fcr \u00f6l\u00e7\u00fcmler yay\u0131mland\u0131\u011f\u0131nda, h\u0131z\u0131n yan\u0131 s\u0131ra enerji verimlili\u011fini ve sonu\u00e7lar\u0131n tutarl\u0131l\u0131\u011f\u0131n\u0131 da de\u011ferlendirmek m\u00fcmk\u00fcn olabilir.<\/p>\n Nvidia H200, yapay zek\u00e2 h\u0131zland\u0131r\u0131c\u0131lar\u0131 alan\u0131nda g\u00fc\u00e7l\u00fc bir referans \u00fcr\u00fcn olarak de\u011ferlendiriliyor. Haberde aktar\u0131lan bilgilere g\u00f6re H200, ABD\u2019nin y\u00fcr\u00fcrl\u00fckteki ihracat k\u0131s\u0131tlamalar\u0131 kapsam\u0131nda anakara \u00c7in\u2019e sevkiyat\u0131na izin verilen en g\u00fc\u00e7l\u00fc Nvidia h\u0131zland\u0131r\u0131c\u0131lar\u0131ndan biri. Bu ko\u015fullar, H200\u2019\u00fc \u00c7in\u2019de geli\u015ftirilen alternatiflerin performans\u0131n\u0131 kar\u015f\u0131la\u015ft\u0131rmak i\u00e7in dikkat \u00e7ekici bir k\u0131yas noktas\u0131 h\u00e2line getiriyor.<\/p>\n \u00c7inli teknoloji \u015firketleri, yapay zek\u00e2 sistemlerinde kullan\u0131lan donan\u0131mlar i\u00e7in yerli se\u00e7enekler geli\u015ftirmeye \u00e7al\u0131\u015f\u0131yor. Bu \u00e7abalar\u0131n \u00f6nemi yaln\u0131zca i\u015flem g\u00fcc\u00fcyle s\u0131n\u0131rl\u0131 de\u011fil: Donan\u0131ma uygun yaz\u0131l\u0131m deste\u011fi, \u00fcretim kapasitesi, sistemlerle uyumluluk ve b\u00fcy\u00fck \u00f6l\u00e7ekli da\u011f\u0131t\u0131m olanaklar\u0131 da bir h\u0131zland\u0131r\u0131c\u0131n\u0131n pratik de\u011ferini belirliyor. Iluvatar CoreX\u2019in Tiangai 300\u2019\u00fc H200 ile kar\u015f\u0131la\u015ft\u0131rmas\u0131, yerli bir \u00fcr\u00fcn\u00fcn belirli i\u015f y\u00fcklerinde nas\u0131l sonu\u00e7 verdi\u011fine ili\u015fkin bir g\u00f6sterge sunuyor; ancak tek ba\u015f\u0131na bu g\u00f6sterim \u00fcr\u00fcn\u00fcn t\u00fcm bu alanlardaki durumunu ortaya koymuyor.<\/p>\n Bir h\u0131zland\u0131r\u0131c\u0131n\u0131n ger\u00e7ek d\u00fcnyadaki ba\u015far\u0131s\u0131n\u0131 anlamak i\u00e7in yaln\u0131zca tek kart\u0131n test sonucuna de\u011fil, farkl\u0131 sistemlerdeki performansa da bakmak gerekir. Veri merkezlerinde \u00e7ok say\u0131da kart\u0131n birlikte \u00e7al\u0131\u015fmas\u0131, a\u011f ba\u011flant\u0131lar\u0131 ve bellek eri\u015fimi gibi ek unsurlar\u0131 g\u00fcndeme getirir. Ayr\u0131ca model geli\u015ftiricilerinin ve yaz\u0131l\u0131m ekiplerinin \u00fcr\u00fcne y\u00f6nelik optimizasyonlar\u0131, zaman i\u00e7inde performans\u0131 art\u0131rabilir veya belirli g\u00f6revlerde farkl\u0131 sonu\u00e7lar do\u011furabilir. Bu y\u00fczden kart ba\u015f\u0131na verim, kapsaml\u0131 \u00fcr\u00fcn de\u011ferlendirmesinin \u00f6nemli fakat tek ba\u015f\u0131na yeterli olmayan bir par\u00e7as\u0131d\u0131r.<\/p>\n \u015eanghay\u2019daki g\u00f6sterimin temel bulgusu, standart yap\u0131land\u0131rmada Tiangai 300\u2019\u00fcn kart ba\u015f\u0131na H200\u2019\u00fcn 1,05 kat\u0131 verime ula\u015ft\u0131\u011f\u0131n\u0131n a\u00e7\u0131klanmas\u0131yd\u0131. \u015eirket ayr\u0131ca GLM-5.3 gibi modellerin yer ald\u0131\u011f\u0131 metinden g\u00f6r\u00fcnt\u00fc ve video \u00fcretimi g\u00f6revlerinde Tiangai 300\u2019\u00fcn H200\u2019\u00fc geride b\u0131rakt\u0131\u011f\u0131n\u0131 bildirdi. Bu sonu\u00e7lar, Tiangai 300\u2019\u00fcn belirtilen i\u015f y\u00fcklerinde Nvidia\u2019n\u0131n g\u00fc\u00e7l\u00fc bir \u00fcr\u00fcn\u00fcyle rekabet edebildi\u011fine i\u015faret ediyor.<\/p>\n \u00d6te yandan, test ayr\u0131nt\u0131lar\u0131n\u0131n tamam\u0131 kamuya a\u00e7\u0131k de\u011filse sonu\u00e7lar\u0131n ba\u011f\u0131ms\u0131z do\u011frulanmas\u0131 ve farkl\u0131 ko\u015fullara genellenmesi g\u00fc\u00e7le\u015fir. Ayn\u0131 g\u00f6revlerin farkl\u0131 yaz\u0131l\u0131m s\u00fcr\u00fcmleriyle veya farkl\u0131 sistem ayarlar\u0131yla y\u00fcr\u00fct\u00fclmesi \u00f6l\u00e7\u00fcmleri de\u011fi\u015ftirebilir. Ayr\u0131ca y\u00fcksek i\u015flem performans\u0131, g\u00fc\u00e7 t\u00fcketimi, sat\u0131n alma maliyeti, \u00fcr\u00fcn\u00fcn bulunabilirli\u011fi ya da uzun s\u00fcreli kararl\u0131l\u0131k hakk\u0131nda do\u011frudan bilgi vermez. Bunlar, bir h\u0131zland\u0131r\u0131c\u0131n\u0131n i\u015fletmeler ve veri merkezi kullan\u0131c\u0131lar\u0131 a\u00e7\u0131s\u0131ndan ne kadar uygun oldu\u011funu belirleyen ayr\u0131 \u00f6l\u00e7\u00fctlerdir.<\/p>\n Bu nedenle Tiangai 300 ile H200 aras\u0131ndaki kar\u015f\u0131la\u015ft\u0131rma, a\u00e7\u0131klanan senaryolar \u00e7er\u00e7evesinde anlaml\u0131 bir performans g\u00f6stergesi olarak g\u00f6r\u00fclebilir. Ancak \u00fcr\u00fcnlerin genel \u00fcst\u00fcnl\u00fc\u011f\u00fc hakk\u0131nda kesin bir sonuca varmak i\u00e7in test y\u00f6ntemlerinin, mutlak sonu\u00e7lar\u0131n ve ek \u00f6l\u00e7\u00fcmlerin payla\u015f\u0131lmas\u0131 gerekir. Farkl\u0131 modellerde, daha b\u00fcy\u00fck sistemlerde ve uzun s\u00fcreli kullan\u0131mda elde edilecek veriler, ilk g\u00f6sterimin ne \u00f6l\u00e7\u00fcde genellenebildi\u011fini ortaya koyacakt\u0131r.<\/p>\n Iluvatar CoreX, Ocak 2026\u2019dan bu yana Hong Kong borsas\u0131nda i\u015flem g\u00f6r\u00fcyor. \u015eirketin Tiangai 300\u2019e ili\u015fkin a\u00e7\u0131klamas\u0131, hem kendi yapay zek\u00e2 donan\u0131m\u0131 \u00fcr\u00fcnlerini hem de \u00c7in\u2019in yerli h\u0131zland\u0131r\u0131c\u0131 geli\u015ftirme \u00e7al\u0131\u015fmalar\u0131n\u0131 yeniden g\u00fcndeme ta\u015f\u0131d\u0131. Bildirilen test sonu\u00e7lar\u0131, yerli bir \u00fcr\u00fcn\u00fcn baz\u0131 yapay zek\u00e2 g\u00f6revlerinde H200\u2019\u00fc a\u015fabildi\u011fini ortaya koyuyor; bununla birlikte performans\u0131n kapsaml\u0131 bi\u00e7imde de\u011ferlendirilmesi i\u00e7in daha fazla teknik veriye ihtiya\u00e7 var.<\/p>\n \u00d6n\u00fcm\u00fczdeki d\u00f6nemde ba\u011f\u0131ms\u0131z testlerin, farkl\u0131 model ve yaz\u0131l\u0131m yap\u0131land\u0131rmalar\u0131ndan elde edilen sonu\u00e7lar\u0131n ve enerji verimlili\u011fi \u00f6l\u00e7\u00fcmlerinin yay\u0131mlanmas\u0131, Tiangai 300\u2019\u00fcn pratik kapasitesini daha net g\u00f6sterebilir. \u015eimdilik en do\u011fru de\u011ferlendirme, \u015firketin a\u00e7\u0131klad\u0131\u011f\u0131 sonu\u00e7lar\u0131 belirli test ko\u015fullar\u0131na ait veriler olarak ele almak ve bunlar\u0131 t\u00fcm uygulamalara genellememektir. Tiangai 300\u2019\u00fcn H200 kar\u015f\u0131s\u0131ndaki performans\u0131 dikkat \u00e7ekici bir iddia olsa da, \u00fcr\u00fcn\u00fcn genel rekabet g\u00fcc\u00fc daha geni\u015f ve \u015feffaf kar\u015f\u0131la\u015ft\u0131rmalarla de\u011ferlendirilebilir.<\/strong><\/p>\nMetinden g\u00f6r\u00fcnt\u00fc ve video \u00fcretimi testleri<\/h2>\n
H200\u2019\u00fcn k\u0131yas noktas\u0131 olarak \u00f6nemi<\/h2>\n
Test sonu\u00e7lar\u0131n\u0131n kapsam\u0131 ve s\u0131n\u0131rlamalar\u0131<\/h2>\n
Iluvatar CoreX\u2019in konumu<\/h2>\n