Ngokukhuphuka okusheshayo kwezinhlelo zokusebenza ze-AI, eziholwa amamodeli afana ne-ChatGPT, isidingo samandla ekhompiyutha senyuke kakhulu.Ukuze kuhlangatshezwane nezidingo ezikhulayo zekhompiyutha zenkathi ye-AI, i-H3C Group, ngaphansi kwesambulela se-Tsinghua Unigroup, isanda kwethula imikhiqizo emisha engu-11 ochungechungeni lwe-H3C UniServer G6 kanye ne-HPE Gen11 eNgqungqutheleni yabaHoli be-NAVIGATE ka-2023.Le mikhiqizo yeseva emisha idala i-matrix ebanzi ye-AI kuzo zonke izimo ezihlukahlukene, ihlinzeka ngenkundla enamandla eyisisekelo yokuphatha idatha enkulu namamodeli we-algorithms, kanye nokuqinisekisa ukutholakala okwanele kwezinsiza zekhompyutha ze-AI.
I-Matrix Yomkhiqizo Ehlukahlukene Yokubhekana Nezidingo Ezihlukahlukene Zekhompyutha ye-AI
Njengomholi kukhompyutha ehlakaniphile, i-H3C Group ibambe iqhaza ngokujulile emkhakheni we-AI iminyaka eminingi.Ngo-2022, i-H3C yazuza izinga lokukhula eliphakeme kakhulu emakethe yamakhompiyutha yaseShayina esheshisiwe futhi yaqongelela isamba samazinga okuqala omhlaba angu-132 kubhentshimakhi ye-AI eyaziwayo emhlabeni wonke iMLPerf, ebonisa ubuchwepheshe bayo obuqinile namandla.
Isebenzisa ukwakheka kwekhompyutha okuthuthukisiwe namandla okuphatha amandla ekhompuyutha ahlakaniphile eyakhelwe esisekelweni sekhompyutha ehlakaniphile, i-H3C ithuthukise ifulegi lekhompyutha elihlakaniphile i-H3C UniServer R5500 G6, elakhelwe ngokukhethekile ukuqeqeshwa kwamamodeli amakhulu.Baphinde bethula i-H3C UniServer R5300 G6, injini eyi-hybrid computing elungele izimo ezinkulu zokucabanga/zokuqeqesha.Le mikhiqizo iqhubeka nokuhlangabezana nezidingo ezahlukahlukene zekhompuyutha ezimeni ezahlukahlukene ze-AI, ihlinzeka ngekhava yekhompuyutha ye-AI ephelele.
I-Intelligent Computing Flagship Edizayinelwe Ukuqeqeshwa Kwemodeli Enkulu
I-H3C UniServer R5500 G6 ihlanganisa amandla, ukusetshenziswa kwamandla okuphansi, nobuhlakani.Uma kuqhathaniswa nesizukulwane sangaphambilini, sinikeza amandla okuhlanganisa aphindwe kathathu, sinciphisa isikhathi sokuqeqesha ngo-70% ku-GPT-4 yezimo zokuqeqeshwa kwemodeli yezinga elikhulu.Isebenza ezimeni ezahlukahlukene zebhizinisi le-AI, njengokuqeqeshwa kwenani elikhulu, ukubonwa kwenkulumo, ukuhlukaniswa kwezithombe, nokuhumusha ngomshini.
Amandla: I-R5500 G6 isekela ama-CPU cores afika kwangu-96, iletha ukukhuphuka okungu-150% ekusebenzeni okuyisisekelo.Ifakwe imojula entsha ye-NVIDIA HGX H800 8-GPU, ehlinzeka ngama-PFLOPS angu-32 wamandla okuhlanganisa, okuphumela ekuthuthukisweni okungu-9x kumodeli yesivinini esikhulu sokuqeqeshwa kwe-AI kanye nokuthuthukiswa okungu-30x ekusebenzeni kwemodeli yezinga elikhulu le-AI.Ukwengeza, ngokusekelwa kwenethiwekhi ye-PCIe 5.0 kanye ne-400G, abasebenzisi bangasebenzisa amaqoqo ekhompiyutha e-AI asebenza kahle kakhulu, asheshise ukwamukelwa nokusebenzisa i-AI emabhizinisini.
Ubuhlakani: I-R5500 G6 isekela ukucushwa kwe-topology okubili, ukuzivumelanisa ngobuhlakani nezimo ezihlukahlukene zohlelo lwe-AI kanye nokusheshisa ukufunda okujulile nezinhlelo zokusebenza zekhompuyutha zesayensi, kuthuthukisa kakhulu ukusetshenziswa kwensiza ye-GPU.Ngenxa yesici se-GPU ye-multi-instance module ye-H800, i-H800 eyodwa ingahlukaniswa ibe yizimo ze-GPU ezingu-7, okungenzeka kube yizimo ze-GPU ezingu-56, ngayinye ibe nezinsiza ezizimele zekhompuyutha nenkumbulo.Lokhu kuthuthukisa kakhulu ukuguquguquka kwezinsiza ze-AI.
I-Low Carbon Footprint: I-R5500 G6 isekela ngokugcwele ukupholisa uketshezi, kuhlanganise nokupholisa uketshezi kokubili kwe-CPU ne-GPU.Nge-PUE (Ukusebenza Kokusetshenziswa Kwamandla) engaphansi kuka-1.1, inika amandla “ikhompyutha epholile” ekushiseni kokuhlinzwa kokubala.
Kuyafaneleka ukubalula ukuthi i-R5500 G6 yaqashelwa njengenye “Yamaseva Aphezulu Okusebenza Kakhulu ayi-10 ango-2023” ku-“2023 Power Ranking for Computational Performance” lapho ikhululwa.
I-Hybrid Computing Engine Yokufanisa Okuguquguqukayo Kokuqeqeshwa kanye Nezimfuno Eziqondisayo
I-H3C UniServer R5300 G6, njengeseva ye-AI yesizukulwane esilandelayo, inikeza intuthuko enkulu ekucacisweni kwe-CPU ne-GPU uma kuqhathaniswa neyandulelayo.Isebenza kahle kakhulu, i-topology ehlakaniphile, namandla ahlanganisiwe ekhompuyutha nokugcina, okuyenza ifanelekele ukuqeqeshwa kwemodeli yokufunda okujulile, inkomba yokufunda ejulile, nezinye izimo zohlelo lokusebenza lwe-AI, ukuqeqeshwa okuvumelana nezimo kanye nezidingo zekhompuyutha zokuqondisisa.
Ukusebenza Okuvelele: I-R5300 G6 ihambisana nesizukulwane sakamuva sama-GPU we-NVIDIA enterprise-grade, ihlinzeka ngokuthuthuka kokusebenza okungu-4.85x uma kuqhathaniswa nesizukulwane sangaphambilini.Isekela izinhlobo ezahlukahlukene zamakhadi wokusheshisa we-AI, njengama-GPU, ama-DPU, nama-NPU, ukuze kuhlangatshezwane nezidingo zamandla ekhompuyutha ahlukahlukene we-AI ezimeni ezahlukahlukene, okunika amandla inkathi yobuhlakani.
I-Intelligent Topology: I-R5300 G6 inikeza izilungiselelo ezinhlanu ze-GPU topology, okuhlanganisa i-HPC, i-parallel AI, i-serial AI, ukufinyelela okuqondile kwamakhadi angu-4, nokufinyelela okuqondile kwamakhadi angu-8.Lokhu kuguquguquka okungakaze kubonwe kuthuthukisa kakhulu ukuguquguquka kwezimo ezihlukene zohlelo lwabasebenzisi, kwabela izinsiza ngobuhlakani, futhi kuqhubekisela phambili ukusebenza kahle kwamandla ekhompuyutha.
I-Computing Edidiyelwe Nesitoreji: I-R5300 G6 iwamukela kalula amakhadi okusheshisa i-AI nama-NIC ahlakaniphile, ahlanganisa ukuqeqeshwa namandla okucabanga.Isekela ama-GPU angu-10 anobubanzi obuphindwe kabili kanye nezikhala ezingama-24 LFF (I-Large Form Factor) hard drive, evumela ukuqeqeshwa kanyekanye kanye nencazelo kuseva eyodwa futhi inikeze injini yekhompuyutha engabizi yokuthuthukiswa nokuhlola izindawo.Ngomthamo wokugcina ongafika ku-400TB, ihlangabezana ngokugcwele nezidingo zesikhala sokugcina idatha ye-AI.
Njengoba i-AI boom iqhubekela phambili, amandla ekhompuyutha ahlala elolongwa kabusha futhi ephonselwa inselelo.Ukukhishwa kwamaseva esizukulwane esilandelayo se-AI kuphawula enye ingqopha-mlando ekuzibophezeleni kwe-H3C Group kubuchwepheshe “bobuhlakani bemvelo” kanye nomfutho wayo oqhubekayo wokuvela kwekhompyutha ehlakaniphile.
Uma sibheka ikusasa, siqondiswa isu le-“Cloud-Native Intelligence”, Iqembu le-H3C linamathela emcabangweni “wokwenza izinto ngendlela ecophelelayo, enikeza inkathi ubuhlakani.”Bazoqhubeka nokuhlakulela inhlabathi evundile yekhompuyutha ehlakaniphile, bahlole izimo zohlelo lokusebenza lwe-AI ezisezingeni elijulile, futhi basheshise ukufika komhlaba ohlakaniphile onamandla alungele ikusasa, navumelana nezimo.
Isikhathi sokuthumela: Jul-04-2023
