Ngenkathi abahlaziyi bemboni begijima ukubikezela usuku lokukhishwa kwama-GPU okuqala e-NVIDIA angu-7nm, abaphathi benkampani bakhetha ukuzikhawulela ekuchazeni zonke izimemezelo ezisemthethweni ezihlobene ngokuthi "ngokuzumayo." Izinhlelo zokusiza abashayeli ezisebenzayo ezisekelwe kuprosesa ye-Tegra yesizukulwane se-Orin zizoqala ukuvela ngo-2022, kodwa ngisho naleyo prosesa ngeke ikhiqizwe kusetshenziswa ubuchwepheshe be-7nm. Kuvela ukuthi i-NVIDIA izosebenzisana ne-Samsung, enobuchwepheshe be-8nm, ukukhiqiza laba prosesa.
Kwabalingani abavela kusayithi Sikwazile ukuthola imibono esemthethweni evela ku-NVIDIA mayelana nenqubo yokukhiqiza esetshenziselwa i-Orin. Nakuba unjiniyela engabonisi ukubandakanyeka kwe-Samsung kula ma-oda, ukusetshenziswa kwenqubo ye-nanometer engu-8 kuyaqinisekiswa. I-TSMC inikeza amakhasimende ayo izinqubo ze-nanometer ezingu-7, 6, noma 5, futhi lokhu kulandelana kulandelana kwesikhathi. Ngakho-ke, konke kukhomba ukuthi i-NVIDIA ibambisana ne-Samsung ekudalweni kwama-processor e-Orin—ikakhulukazi njengoba ukwakheka kwawo kufanele kube lula kune-GPU yesizukulwane esilandelayo esicatshangelwayo, okungenzeka ukuthi izokhiqizwa yi-TSMC. Isidingo esikhulu sezinsizakalo ze-TSMC siphoqa i-NVIDIA ukuthi icabangele i-Samsung.
Ukwakhiwa kwe-ARM Hercules, okuzosetshenziswa ngabaprosesa be-Orin, kwaqale kwaklanywa ubuchwepheshe benqubo ye-7- noma ye-5-nm, kodwa ubuchwepheshe be-Samsung be-8-nm cishe buzofana ngamapharamitha. Ekucushweni okuphezulu, abaprosesa be-Orin bazoba nama-cores e-computing ayishumi nambili asekelwe ku-Hercules, kodwa imininingwane mayelana nama-cores ehluzo igcinwa iyimfihlo.
Njengangaphambili, izinhlelo zokusiza abashayeli ezisebenzayo zokuzimela kweLevel 2 zizokwazi ukusebenzisa iprosesa eyodwa ye-Orin. Inguqulo eyisisekelo kakhulu, enokusetshenziswa kwamandla okungadluli ama-watts ayi-15, izonikwa amandla yikhamera eyodwa, futhi izofinyelela izinga lokusebenza lemisebenzi eyizigidigidi ezingama-36 ngomzuzwana. Inguqulo ye-Orin ethuthuke kakhulu izokwazi ukukhulisa ukusebenza kube imisebenzi eyizigidigidi eziyi-100 ngomzuzwana, isebenzise amakhamera amane ngesikhathi esisodwa, futhi ingasebenzisi ama-watts angaphezu kwama-40.
I-duo yeprosesa ye-Orin izothuthukisa ukusebenza kwayo kube yi-400 trillion operations ngomzuzwana (TOPS) ngenkathi ingasebenzisi amandla angaphezu kwama-watts angu-130. Lokhu kwanele ezinhlelweni zokusiza abashayeli zezinga lesi-3. Izinhlelo ezifanayo ezisekelwe ku-Xavier zisebenzise i-GPU ehlukile kodwa zisebenzise amandla afinyelela kuma-watts angu-230, futhi ukusebenza kwazo akudluli ama-TOPS angu-160.
Uhlelo oluphambili lwe-Level 5 luzohlanganisa amaprosesa amabili e-Orin nama-GPU e-NVIDIA ahlukile. Lokhu kuzokhulisa ukusebenza kube ama-tOP angu-2000, kodwa futhi kuzokhulisa ukusetshenziswa kwamandla kube ama-watts angu-750. Uhlelo olufanayo olusekelwe kumaprosesa amabili e-Xavier kanye nama-GPU amabili esizukulwane se-Volta lunikeze ukusebenza okungaphezu kwama-tOP angu-320 ngokusetshenziswa kwamandla kwama-watts angu-460. Kusenesikhathi kakhulu ukunquma ukuthi ama-GPU e-NVIDIA ahlukile alesi sigaba azosebenzisa yini inkumbulo ye-HBM yesizukulwane esifanayo, kodwa umfanekiso ukhomba lokho.
Source: 3dnews.ru
