NVIDIA tana Ajiye Chiplets don Mafi Kyau

Idan kun yi imani da kalaman NVIDIA Babban Mashawarcin Kimiyyar Kimiyya Bill Dally a cikin wata hira da albarkatun Injiniya Semiconductor, Kamfanin ya haɓaka fasaha don ƙirƙirar na'ura mai mahimmanci tare da tsarin guntu mai yawa shekaru shida da suka wuce, amma har yanzu bai shirya yin amfani da shi ba wajen samar da taro. A gefe guda, kamfanin ya kuma fara sanya kwakwalwan ƙwaƙwalwar ajiyar nau'in HBM a kusanci kusa da GPU shekaru da yawa da suka gabata, don haka ba za a iya zarge shi gaba ɗaya yin watsi da "salon ga chiplets."

Har ya zuwa yanzu an yi ta cewa samfurin NVIDIA na buƙatar na'ura mai mahimmanci na 36-core tare da gine-gine na RISC-V don gwada hanyoyin da za a iya yin ƙima a cikin masu haɓaka ƙididdiga, da kuma shirya don gabatarwar sababbin hanyoyin tattarawa. Duk wannan ƙwarewar, a cewar wakilan NVIDIA, kamfanin na iya buƙatar buƙata a lokacin da ya zama mai yuwuwar tattalin arziki don ƙirƙirar GPUs daga “chiplets” guda ɗaya. Irin wannan lokacin bai riga ya isa ba, kuma NVIDIA ba ta ma yin hasashen lokacin da hakan zai faru ba.

Bill Dally kuma ya lura cewa dogaro da lithography zuwa sikelin aikin sarrafawa bai daɗe ba. Tsakanin matakai biyu masu kusa na tsarin fasaha, haɓakar aikin transistor ana auna shi da kashi 20%, a cikin mafi kyawun yanayin, kuma ƙirar gine-gine da software na iya haɓaka aikin na'urori masu sarrafa hoto sau da yawa. A wannan ma'anar, gine-ginen ya mamaye lithography daga ra'ayi na NVIDIA.

An tabbatar da wannan matsayi akai-akai a cikin maganganunsa ta hanyar wanda ya kafa NVIDIA Jensen Huang. Har ya zuwa yanzu, ya yi iya ƙoƙarinsa don tabbatar da ci gaban tsarin da ake bi don ƙirƙirar lu'ulu'u na monolithic, ya yi magana da ɓatanci ga masu fafatawa da ke neman sabbin hanyoyin fasaha, har ma da wasa da kwatanta "chiplets" tare da ɗanɗano mai baƙar fata ("chiclets"), yana bayyana cewa. kawai yana son sabon fassarar wannan kalmar. Koyaya, kalamai daga kwararrun NVIDIA kusa da haɓaka samfura suna ba mu damar yin imani cewa a ƙarshe kamfanin zai canza zuwa shimfidar guntu da yawa. Intel, alal misali, bai ɓoye niyyarsa na yin 7nm GPU multi-chip ta amfani da shimfidar Foveros ba. AMD yana amfani da "chiplets" a rayayye yayin ƙirƙirar na'urori masu sarrafawa na tsakiya, amma a cikin sashin zane-zane ya iyakance kansa zuwa "sharing" nau'in ƙwaƙwalwar ajiya na HBM2.

source: 3dnews.ru