Muva nje, inhlangano egunyazwe emhlabeni jikelele yokuhlola ibhentshimakhi ye-AI i-MLPerf™ ikhiphe izinga lakamuva le-AI Inference V3.1. Isamba esingama-25 semiconductor, iseva, nabakhiqizi be-algorithm emhlabeni jikelele ababambe iqhaza kulokhu kuhlola. Emqhudelwaneni oshubile, i-H3C yagqama esigabeni seseva ye-AI futhi yazuza okokuqala emhlabeni, ikhombisa ukuqanjwa okuqinile kwe-H3C namakhono okuthuthukisa umkhiqizo emkhakheni we-AI.
I-MLPerf™ yethulwa owine iTuring Award uDavid Patterson ngokubambisana nezikhungo zemfundo ephakeme. Iwukuhlolwa kwebhentshimakhi yobuhlakani bokwenziwa eyaziwa kakhulu futhi ebambe iqhaza. Kubandakanya ukucutshungulwa kolimi lwemvelo, ukuhlukaniswa kwezithombe zezokwelapha, izincomo ezihlakaniphile namanye amathrekhi angamamodeli akudala. Ihlinzeka ngokuhlolwa okufanele kwezingxenyekazi zekhompuyutha zomkhiqizi, isofthiwe, ukuqeqeshwa kwesevisi kanye nokusebenza okucatshangelwayo. Imiphumela yokuhlolwa inokusebenza okubanzi kanye nenani eliyireferensi. Emqhudelwaneni wamanje wengqalasizinda ye-AI, i-MLPerf inganikeza umhlahlandlela wedatha ogunyaziwe futhi osebenzayo wokulinganisa ukusebenza kwemishini, ibe “itshe lokuthinta” lamandla obuchwepheshe babakhiqizi emkhakheni we-AI. Ngokugxila kweminyaka namandla aqinile, i-H3C iwine imiqhudelwano engu-157 ku-MLPerf.
Kulolu vivinyo lwe-benchmark ye-AI Inference, iseva ye-H3C R5300 G6 yenze kahle, ibekwe kuqala ekucushweni okungu-23 ezikhungweni zedatha nezimo ezinqenqemeni, futhi okokuqala ekucushweni oku-1 okuphelele, okufakazela ukwesekwa kwayo okunamandla kwezinhlelo zokusebenza ezinkulu, ezihlukahlukene, nezithuthukile. . Izimo zekhompyutha eziyinkimbinkimbi.
Kuthrekhi yemodeli ye-ResNet50, iseva ye-R5300 G6 ingahlukanisa izithombe ezingu-282,029 ngesikhathi sangempela ngomzuzwana, inikeze amakhono asebenzayo nanembile okucubungula izithombe kanye nokuqashelwa.
Kumodeli yethrekhi ye-RetinaNet, iseva ye-R5300 G6 ingakwazi ukuhlonza izinto ezithombeni ezingu-5,268.21 ngomzuzwana, inikeze isisekelo sekhompiyutha sezimo ezifana nokushayela okuzenzakalelayo, ukudayisa ngobuhlakani, kanye nokukhiqiza okuhlakaniphile.
Kumodeli yethrekhi ye-3D-UNet, iseva ye-G6 ye-R5300 ingahlukanisa izithombe zezokwelapha ze-3D ezingu-26.91 ngomzuzwana, ngemfuneko yokunemba engu-99.9%, isiza odokotela ekuxilongeni ngokushesha nasekuthuthukiseni ukusebenza kahle kokuxilongwa kanye nekhwalithi.
Njengokuhamba phambili kwamakhono amakhompiyutha amaningi enkathini yobuhlakani, iseva ye-R5300 G6 inokusebenza okuhle kakhulu, izakhiwo ezivumelana nezimo, ukuqina okuqinile, nokuthembeka okuphezulu. Isekela izinhlobo eziningi zamakhadi okusheshisa i-AI, anezilinganiso zokufakwa kwe-CPU ne-GPU okungu-1:4 no-1:8, futhi ihlinzeka ngezinhlobo ezingu-5 ze-topology ye-GPU ukuze ivumelane nezidingo zezimo ze-AI ezahlukene. Ngaphezu kwalokho, i-R5300 G6 yamukela idizayini edidiyelwe yamandla ekhompiyutha nokugcinwa, isekela ama-GPU abanzi afinyelela kwayi-10 kanye nama-400TB wesitoreji esikhulu ukuhlangabezana nezidingo zendawo yokugcina idatha ye-AI.
Ngesikhathi esifanayo, ngokuklama kwayo okuthuthukisiwe kwesistimu ye-AI namandla okwenza isitaki esigcwele, iseva ye-R5350 G6 ibekwe kuqala ngokucushwa okufanayo kumsebenzi wokuhlola we-ResNet50 (ukuhlelwa kwezithombe) kulolu hlolo lokulinganiswa. Uma kuqhathaniswa nomkhiqizo wesizukulwane sangaphambilini, u-R5350 G6 uzuza ukuthuthukiswa kokusebenza okungu-90% kanye nokwenyuka kwesibalo esingu-50%. Ifakwe inkumbulo yeziteshi eziyi-12, umthamo wememori ungafinyelela ku-6TB. Ngaphezu kwalokho, i-R5350 G6 isekela ama-hard drive afinyelela kwangu-24 2.5/3.5-intshi, izindawo eziyi-12 PCIe5.0 namakhadi enethiwekhi angu-400GE ukuhlangabezana nesidingo se-AI sokugcinwa kwedatha okukhulu kanye nomkhawulokudonsa wenethiwekhi wesivinini esikhulu. Ingasetshenziswa ezimeni ezihlukahlukene ezifana nokuqeqeshwa kwemodeli yokufunda okujulile, inkomba yokufunda ejulile, ikhompuyutha esebenza kahle kakhulu, nokuhlaziywa kwedatha.
Konke ukuphumelela nokusebenza okwephula irekhodi kubonisa ukugxila kwe-H3C Group ezimeni zohlelo lwamakhasimende kanye nokuqongelela kwayo ulwazi olungokoqobo namandla obuchwepheshe. Ngokuzayo, i-H3C izohambisana nomqondo "wokunemba kwezolimo, ukunika amandla inkathi yobuhlakani", ihlanganise eduze ukuqanjwa kabusha komkhiqizo nezimo zokusetshenziswa kobuhlakani bokwenziwa, futhi ilethe inguquko eqhubekayo yamandla ekhompyutha ahlakaniphile kuzo zonke izigaba zokuphila.
Isikhathi sokuthumela: Sep-13-2023
