Umhlaziyi uMing-Chi Kuo uthe izolo ucwaningo lwakhe lwakamuva lwemboni lubonisa ukuthi i-NVIDIA isiqalise kabusha iphrojekthi ye-GPU i-Rubin CPX yokusheshisa i-prefill ye-artificial intelligence inference, kodwa ukwakhiwa kwayo kushintshwe kakhulu. Kulindeleke ukuthi ukukhiqizwa kuqale ngo-2027Q1.
Ezingeni le-chip, i-Rubin CPX entsha inamandla okucubungula acishe afane na-standard Rubin GPU, kanti ukusebenza kwe-prefill kuthuthukisiwe. Amandla aphezulu asetshenziswa iyunithi eyodwa nawo angu-2300W, futhi ifakwe inkumbulo engu-168GB HBM4 (inothi: i-Rubin CPX endala ibine-128GB GDDR7; i-168GB kungenzeka ibhekisele ekwakhiweni kwe-7×24GB).
Ezingeni lethreyi nerack, ama-Rubin CPX amasha angu-8 akha ithreyi eyodwa yokubala; amathreyi okubala angu-8 akha imojuli eyodwa ye-rack, exhunywe ngaphakathi nge-Spectrum-6 Ethernet yethusi elimsulwa ukuze kwandiswe ngokuvundlile (Scale-out). I-Rubin CPX entsha isebenzisa i-rack ezimele ye-MGX ETL, kanti i-rack ngayinye iqukethe amaqoqo angu-1~4 amamojuli e-rack. Ukwandiswa okuvundlile phakathi kwamamojuli e-rack kusebenzisa i-Spectrum-6 kanye nesixazululo se-OSFP optical.
Ezingeni lesistimu, i-NVIDIA iphakamisa ukuthi i-Rubin CPX ne-standard Rubin kubhangqwe ngesilinganiso esingu-1:1. Phakathi kwe-rack ye-Rubin CPX ETL nerack ye-Vera Rubin NVL72 kusetshenziswa uxhumano lwe-RDMA olusekelwe ku-Ethernet.
Lo mhlaziyi uphawule wathi okwamanje ngaphezu kuka-50% womthwalo womsebenzi we-AI inference uvela ekucubunguleni okokufaka nasekwakheni i-KV Cache ehambisanayo. I-Rubin CPX entsha ingakwazi ukuthwala umthwalo we-prefill ngendlela yokuyifaka esebenza kalula futhi ngezindleko eziphansi; ithreyi yokubala ye-Rubin CPX entsha inenkumbulo engu-1.34TB HBM, eyanele ukuhlangabezana nezidingo eziningi zokwenza i-prefill yezimo ezinde kanye nokwakha i-KV Cache.