Kutshanje, umbutho wovavanyo lomgangatho we-AI ogunyazisiweyo wehlabathi jikelele iMLPerf™ ukhuphe inqanaba le-AI Inference V3.1 yamva nje. Itotali ye-25 semiconductor, iseva, kunye nabavelisi be-algorithm emhlabeni jikelele bathathe inxaxheba kolu vavanyo. Kukhuphiswano oluqatha, i-H3C yagqama kudidi lweseva ye-AI yaze yaphumelela i-25 yehlabathi yokuqala, ibonisa i-H3C yobuchule obuqinileyo bobuchule kunye nophuhliso lwemveliso kwibala le-AI.
I-MLPerf™ yaphehlelelwa ngophumelele iMbasa yeTuring uDavid Patterson ebambisene namaziko emfundo aphezulu. Lolona vavanyo lwaziwayo noluthathelwe inxaxheba kubukrelekrele bokwenziwa. Kubandakanya ukusetyenzwa kolwimi lwendalo, ukwahlulwa kwemifanekiso yezonyango, iingcebiso ezikrelekrele kunye nezinye iingoma zemodeli yakudala. Ibonelela ngovavanyo olufanelekileyo lwehardware yomenzi, isoftware, uqeqesho lwenkonzo kunye nokusebenza okuthelekelelwayo. Iziphumo zovavanyo zinesicelo esibanzi kunye nexabiso lereferensi. Kukhuphiswano lwangoku lweziseko ezingundoqo ze-AI, iMLPerf inokubonelela ngesikhokelo esigunyazisiweyo nesisebenzayo sedatha yokulinganisa ukusebenza kwezixhobo, ibe “lilitye lokuchophela” lamandla obugcisa babavelisi kwibala le-AI. Ngeminyaka yokugxila kunye namandla anamandla, i-H3C iphumelele iintshatsheli ze-157 kwi-MLPerf.
Kulo vavanyo lwe-benchmark ye-AI, i-H3C R5300 iseva ye-G6 iqhube kakuhle, ibeka kuqala kwiimpawu ze-23 kumaziko edatha kunye neemeko ezinqamlekileyo, kwaye okokuqala kwi-1 yoqwalaselo olupheleleyo, ebonisa inkxaso yayo eqinile kwizinga elikhulu, elihlukeneyo, kunye nezicelo eziphambili. . Iimeko zekhompuyutha ezintsonkothileyo.
Kwimodeli yemodeli ye-ResNet50, umncedisi we-R5300 we-G6 unokwahlula imifanekiso engama-282,029 ngexesha langempela ngesekhondi, ukubonelela ngokufanelekileyo nangokuchanekileyo ukusetyenzwa komfanekiso kunye namandla okuqaphela.
Kumkhondo wemodeli ye-RetinaNet, iseva ye-R5300 ye-G6 inokuchonga izinto kwimifanekiso engama-5,268.21 ngomzuzwana, ibonelela ngesiseko sekhompuyutha yeemeko ezifana nokuqhuba ngokuzimeleyo, ukuthengisa ngobuchule, kunye nokuvelisa ngobuchule.
Kumzila wemodeli ye-3D-UNet, i-R5300 iseva ye-G6 ingahlula i-26.91 imifanekiso ye-3D yonyango ngomzuzwana, kunye nemfuno echanekileyo ye-99.9%, ukuncedisa oogqirha ekuxilongeni ngokukhawuleza kunye nokuphucula ukuxilongwa kunye nomgangatho.
Njengoko iflegi yezakhono ezininzi zekhompyuter kwixesha elikrelekrele, i-R5300 iseva ye-G6 inomsebenzi ogqwesileyo, ulwakhiwo oluguquguqukayo, ukuqina okunamandla, kunye nokuthembeka okuphezulu. Ixhasa iindidi ezininzi ze-AI zamakhadi e-accelerator, kunye ne-CPU kunye ne-GPU yokufakela umlinganiselo we-1: 4 kunye ne-1: 8, kwaye ibonelela ngeentlobo ze-5 ze-topology ze-GPU ukulungelelanisa neemfuno zeemeko ezahlukeneyo ze-AI. Ngaphezu koko, i-R5300 G6 ithatha uyilo oludibeneyo lwamandla ekhompyuter kunye nokugcinwa, ukuxhasa ukuya kuthi ga kwi-10 ebanzi kabini ye-GPU kunye ne-400TB yokugcina enkulu ukuhlangabezana neemfuno zendawo yokugcina idatha ye-AI.
Ngexesha elifanayo, kunye noyilo lwenkqubo ye-AI ephuculweyo kunye nobuchule obugcweleyo be-stack, i-R5350 iseva ye-G6 ibekwe kwindawo yokuqala kunye noqwalaselo olufanayo kwi-ResNet50 (ukuhlelwa komfanekiso) umsebenzi wokuvavanya kolu vavanyo lwe-benchmark. Xa kuthelekiswa nemveliso yemveliso yangaphambili, i-R5350 G6 iphumeza ukuphuculwa kwentsebenzo ngama-90% kunye nokwanda kwe-50% kumanani angundoqo. Ixhotyiswe ngememori ye-12-channel, umthamo wememori unokufikelela kwi-6TB. Ukongeza, i-R5350 G6 ixhasa ukuya kuthi ga kwi-24 2.5/3.5-inch hard drives, 12 PCIe5.0 slots kunye ne-400GE network cards ukuhlangabezana nemfuno ye-AI yogcino lwedatha enkulu kunye nesantya esiphezulu sothungelwano bandwidth. Ingasetyenziswa kwiimeko ezahlukeneyo ezifana noqeqesho lwemodeli yokufunda nzulu, intelekelelo yokufunda enzulu, ukusebenza okuphezulu kwekhompyuter, kunye nohlalutyo lwedatha.
Yonke impumelelo kunye nokusebenza okurekhodiweyo kubonisa ugqaliselo lweQela le-H3C kwiimeko zezicelo zabathengi kunye nokuqokelela kwayo amava asebenzayo kunye nobuchule bobugcisa. Kwixesha elizayo, i-H3C iya kubambelela kumbono "wokuchaneka kwezolimo, ukuxhobisa ixesha lobukrelekrele", idibanise ngokusondeleyo ukuveliswa kwemveliso kunye neemeko zokusetyenziswa kobukrelekrele bokwenziwa, kwaye izise inguquko eqhubekayo yamandla ekhompyuter yobukrelekrele kuzo zonke iindlela zobomi.
Ixesha lokuposa: Sep-13-2023
