Ngokuhlanganisa ukuhlaziywa kwe-Eugene Investment & Securities yaseNingizimu Korea okwabiwe ngumthombo wolwazi u-@jukan05, kanye nendatshana eshicilelwe namuhla yisikhungo i-TrendForce, i-NVIDIA ihlola ukwehlisa ukucushwa kwe-HBM ye-Rubin Ultra AI GPU yangonyaka ozayo, futhi ihlola ngokuhambisanayo ezinye izixazululo ngaphandle kwe-12Hi HBM4E ebihlelwe ekuqaleni.
Izindlela okungenzeka i-NVIDIA yehlise ngazo izinga zihlanganisa ukusuka ku-HBM4E kuye ku-HBM4, kanye nokusuka ku-12Hi HBM kuye ku-8Hi HBM. Eyokuqala idalwa wukungaqiniseki ngenqubekelaphambili yokuqinisekiswa kwe-HBM4E ngabakhiqizi abathathu abakhulu be-DRAM; akukacaci ukuthi ukuqinisekiswa nokukhiqizwa ngobuningi kuzophothulwa ngesikhathi yini.
Uma kuncishiswa inani lezingqimba ze-DRAM Die ezinqwabelene ku-HBM stack, inani elifanayo lama-HBM DRAM Die lingakha ama-HBM stack amaningi, ngaleyo ndlela kuhlangatshezwane nezidingo zememori zama-Rubin Ultra amaningi. Leli qhinga lisiza ukuthumela imikhiqizo eminingi lapho inani eliphelele lokuhlinzekwa kwe-DRAM lilinganiselwe; isimo esifanayo senzekile naku-LPDDR5X SOCAMM2 ye-Vera CPU.
I-TrendForce ikholelwa ukuthi iphuzu eliyinhloko lokuthuthukiswa kwe-Rubin Ultra lisesezingeni lesivinini se-I/O, ngakho-ke i-NVIDIA kungenzeka ikhethe isu lokunciphisa inqwaba ye-HBM. Ngaphezu kwalokho, kuthiwa abanye abenzi be-AI ASIC nabo bacabangela umthamo ophansi we-HBM.