I-NVIDIA ilondoloza amandla okusebenzisa ama-chiplets ngezikhathi ezingcono

Uma ukholelwa izitatimende zika-NVIDIA Chief Scientific Advisor Bill Dally kungxoxo nesisetshenziswa Ubunjiniyela be-Semiconductor, inkampani ithuthukise ubuchwepheshe bokudala iprosesa enama-multi-core ngesakhiwo se-multi-chip eminyakeni eyisithupha edlule, kodwa namanje ayikakulungeli ukuyisebenzisa ekukhiqizeni ngobuningi. Ngakolunye uhlangothi, inkampani iphinde yaqala ukubeka ama-memory chips ohlobo lwe-HBM eduze ne-GPU eminyakeni embalwa edlule, ngakho-ke ayinakusolwa ngokushaya indiva ngokuphelele "imfashini yama-chiplets."

Kuze kube manje kusaphikiswana ngokuthi isibonelo I-NVIDIA idinga iphrosesa engu-36-core enezakhiwo ze-RISC-V ukuze ihlole izindlela zokukala ukusebenza kwezisheshisi zekhompyutha, kanye nokulungiselela ukwethulwa kwezixazululo ezintsha zokupakisha. Konke lokhu okuhlangenwe nakho, ngokusho kwabamele i-NVIDIA, kungase kudingwe inkampani ngesikhathi lapho kuba nokwenzeka ngokomnotho ukudala ama-GPU "ku-chiplets" ngazinye. Isikhathi esinjalo asikakafiki, futhi i-NVIDIA ayifuni ngisho nokubikezela ukuthi lokhu kuzokwenzeka nini.

UBill Dally uphinde waphawula ukuthi ukuthembela ku-lithography ukukala ukusebenza kweprosesa kudala kungenangqondo. Phakathi kwezigaba ezimbili eziseduze zenqubo yobuchwepheshe, ukukhuphuka kokusebenza kwe-transistor kukalwa ku-20%, esimweni esihle kakhulu, kanye nokuqanjwa kwezakhiwo nesofthiwe kungakhuphula ukusebenza kwama-GPU izikhathi ezimbalwa. Ngalo mqondo, i-architecture ibusa i-lithography ngokombono we-NVIDIA.

Lesi sikhundla siqinisekiswe ngokuphindaphindiwe esitatimendeni sakhe ngumsunguli we-NVIDIA uJensen Huang. Kuze kube manje, wenze konke okusemandleni akhe ukuze afakazele inqubekela phambili yendlela yokudala amakristalu e-monolithic, wakhuluma ngokudelela abancintisana nabo abajaha izinqubo ezintsha zobuchwepheshe, waze waqhathanisa ngokuhlekisa "ama-chiplets" nongwaqa wokuhlafuna ushingamu ("ama-chiclets"), echaza lokho. uthanda kuphela incazelo yakamuva yaleli gama. Kodwa-ke, izitatimende ezivela kochwepheshe be-NVIDIA eduze nokuthuthukiswa komkhiqizo zisivumela ukuthi sikholwe ukuthi inkampani izogcina ishintshele kusakhiwo esinama-chip amaningi. I-Intel, ngokwesibonelo, ayizange ifihle izinhloso zayo zokwenza i-7nm GPU multi-chip isebenzisa ukwakheka kweFoveros. I-AMD isebenzisa ngenkuthalo “ama-chiplets” lapho idala amaphrosesa amaphakathi, kodwa engxenyeni yezithombe isizikhawulele kuze kube manje “ekwabelaneni” kwenkumbulo yohlobo lwe-HBM2.

Source: 3dnews.ru