[China, Shenzhen, Julayi 14, 2023] Namhlanje, uHuawei utyhile isisombululo sakhe esitsha sokugcina i-AI kwixesha leemodeli ezinkulu, ebonelela ngezisombululo ezifanelekileyo zokugcinwa kwesiseko soqeqesho olusisiseko, uqeqesho lwemodeli ethile kwishishini, kunye nenkcazo kwiimeko ezicandekileyo, ngaloo ndlela. ukhulula amandla amatsha e-AI.
Kuphuhliso kunye nokuphunyezwa kwezicelo zemodeli enkulu, amashishini ajongene nemingeni emine emikhulu:
Okokuqala, ixesha elifunekayo lokulungiswa kwedatha lide, imithombo yedatha isasazekile, kwaye udibaniso lucotha, kuthatha malunga neentsuku ezili-10 ukulungiselela kwangaphambili amakhulu edatha yedatha. Okwesibini, kwiimodeli ezinkulu zeemodyuli ezininzi ezinombhalo omkhulu kunye neeseti zedatha yemifanekiso, isantya sangoku sokulayisha kwiifayile ezincinci ezinkulu singaphantsi kwe-100MB/s, okukhokelela ekusebenzeni okuphantsi koqeqesho lokulayishwa kweseti. Okwesithathu, uhlengahlengiso lweparameter rhoqo kwiimodeli ezinkulu, kunye namaqonga oqeqesho angazinzanga, kubangela ukuphazamiseka koqeqesho malunga neentsuku ezi-2, kunyanzeleke indlela yokuHlola ukuba iqalise uqeqesho, kunye nokuthatha usuku olulula. Okokugqibela, imida yophumezo ephezulu yeemodeli ezinkulu, ukuseta inkqubo entsonkothileyo, imiceli mngeni yokucwangcisa izixhobo, kunye nokusetyenziswa kwezixhobo ze-GPU rhoqo ngaphantsi kwe-40%.
IHuawei ihambelana nomkhwa wophuhliso lwe-AI ngeli xesha leemodeli ezinkulu, ezibonelela ngezisombululo ezilungiselelwe amashishini kunye neemeko ezahlukeneyo. Yazisa i-OceanStor A310 yeDatha eNzululwazi yokuFunda iLake Storage kunye neFusionCube A3000 Training/Inference Super-Converged Appliance. OceanStor A310 Deep Learning Data Lake Storage ekujoliswe kuzo zombini ezisisiseko kunye neshishini-level imodeli data enkulu iimeko echibini, ekufezekiseni olubanzi AI ulawulo lwedatha ukusuka ukuhlanganiswa data, preprocessing ukuya kuqeqesho imodeli, kunye nezicelo inference. I-OceanStor A310, kwi-rack eyodwa ye-5U, isekela i-400GB / s i-bandwidth ehamba phambili kwi-industry kunye nokufika kwi-12 yezigidi ze-IOPS, kunye ne-linear scalability ukuya kuma-4096 nodes, eyenza unxibelelwano olungenamthungo lweprotocol. I-Global File System (GFS) iququzelela ukuthungwa kwedatha ngobukrelekrele kuyo yonke imimandla, iphucula iinkqubo zokudityaniswa kwedatha. Icomputing ekufutshane nokugcinwa kwedatha iqonda i-preprocessing ekufutshane-data, ukunciphisa ukunyakaza kwedatha, kunye nokuphucula ukusebenza kakuhle kwangaphambili nge-30%.
I-FusionCube A3000 Training/Inference Super-Converged Appliance, eyenzelwe i-industry-level enkulu yemodeli yoqeqesho / inference scenarios, ibonelela kwizicelo ezibandakanya iimodeli ezineebhiliyoni zeeparamitha. Idibanisa iindawo zokugcina eziphezulu ze-OceanStor A300, uqeqesho / i-inference nodes, izixhobo zokutshintsha, i-software ye-platform ye-AI, kunye nesofthiwe yokulawula kunye nokusebenza, ukubonelela amaqabane amakhulu omzekelo kunye namava okusasazwa kweplagi kunye ne-play-stop ukuhanjiswa okukodwa. Ilungele ukusetyenziswa, ingafakwa kwiiyure ezi-2. Zombini uqeqesho / intelekelelo kunye neendawo zokugcina zinokwandiswa ngokuzimeleyo kwaye ngokuthe tye ukuze zihambelane neemfuno zesikali semodeli. Ngeli xesha, i-FusionCube A3000 isebenzisa izikhongozeli ezisebenza ngokuphezulu ukwenza uqeqesho lwemodeli emininzi kunye nemisebenzi yokuqonda ukwabelana ngeGPUs, ukwandisa ukusetyenziswa kwezixhobo ukusuka kwi-40% ukuya ngaphezulu kwe-70%. I-FusionCube A3000 ixhasa iimodeli ezimbini zeshishini eziguquguqukayo: IHuawei inyukela kwiSisombululo esine-One-Stop kunye neqabane lesithathu isisombululo esine-computing evulekileyo, inethiwekhi, kunye nesoftware yeqonga le-AI.
Umongameli weHuawei weMveliso yoGcino lweDatha, uZhou Yuefeng, uthe, "Ngexesha leemodeli ezinkulu, idatha imisela ubude bobukrelekrele be-AI. Njengomphathi wedatha, ukugcinwa kwedatha kuba sisiseko sesiseko sesiseko se-AI yeemodeli ezinkulu. I-Huawei Data Storage iya kuqhubeka nokusungula, ibonelela ngezisombululo ezahlukeneyo kunye neemveliso zexesha leemodeli ezinkulu ze-AI, ukusebenzisana namaqabane ukuqhuba ukuxhotyiswa kwe-AI kuluhlu olubanzi lwamashishini. "
Ixesha lokuposa: Aug-01-2023
