[I-China, Shenzhen, Julayi 14, 2023] Namuhla, uHuawei wethule isisombululo sakhe esisha se-AI senkathi yamamodeli amakhulu, ehlinzeka ngezixazululo eziphelele zokugcina ukuqeqeshwa kwamamodeli ayisisekelo, ukuqeqeshwa kwemodeli ethile embonini, kanye nokuchazwa kwezimo ezihlukene, ngaleyo ndlela. ukuvula amakhono amasha e-AI.
Ekuthuthukisweni nasekusetshenzisweni kwezicelo zamamodeli amakhulu, amabhizinisi abhekana nezinselelo ezinkulu ezine:
Okokuqala, isikhathi esidingekayo sokulungiswa kwedatha side, imithombo yedatha iyahlakazeka, futhi ukuhlanganisa kuhamba kancane, kuthatha cishe izinsuku eziyi-10 ukucubungula ngaphambili amakhulu ama-terabytes edatha. Okwesibili, kumamodeli amakhulu anezinhlobo eziningi anombhalo omkhulu nedathasethi yezithombe, isivinini samanje sokulayisha samafayela amancane amakhulu singaphansi kuka-100MB/s, okuholela ekusebenzeni okuphansi kokulayishwa kwesethi yokuqeqeshwa. Okwesithathu, ukulungiswa okuvamile kwepharamitha yamamodeli amakhulu, kanye nezinkundla zokuqeqesha ezingazinzile, kubangela ukuphazamiseka kokuqeqeshwa cishe njalo ezinsukwini ezi-2, okudingekile ukuthi indlela yokuhlola iqalise kabusha ukuqeqeshwa, nokululama kuthatha usuku. Okokugcina, imikhawulo ephezulu yokuqaliswa yamamodeli amakhulu, ukusethwa kwesistimu eyinkimbinkimbi, izinselele zokuhlela izinsiza, kanye nokusetshenziswa kwezinsiza ze-GPU ngokuvamile ngaphansi kwama-40%.
I-Huawei ihambisana nokuthambekela kwentuthuko ye-AI esikhathini samamodeli amakhulu, enikeza izixazululo ezenzelwe izimboni nezimo ezahlukahlukene. Yethula i-OceanStor A310 Deep Learning Data Lake Storage kanye ne-FusionCube A3000 Training/Inference Super-Converged Appliance. I-OceanStor A310 Deep Learning Data Lake Storage ihlose kokubili imodeli yedatha eyisisekelo nezinga lemboni yezimo zedatha yechibi, ukufeza ukulawulwa okuphelele kwedatha ye-AI kusukela ekuhlanganisweni kwedatha, ukucutshungulwa kusengaphambili kuya ekuqeqeshweni okuyimodeli, kanye nezinhlelo zokusebenza eziqondisayo. I-OceanStor A310, endaweni yokubeka eyodwa engu-5U, isekela umkhawulokudonsa ongu-400GB/s ohamba phambili embonini futhi kufika ezigidini eziyi-12 ze-IOPS, ngomugqa wokukala kuze kufike ku-4096 nodes, okuvumela ukuxhumana okungenamthungo kwe-cross-protocol. I-Global File System (GFS) isiza ukwelukwa kwedatha okuhlakaniphile kuzo zonke izifunda, iqondise izinqubo zokuhlanganisa idatha. Ikhompyutha ye-Near-storage computing ibona ukucutshungulwa kwangaphambili kwedatha, ukunciphisa ukunyakaza kwedatha, nokuthuthukisa ukusebenza kahle kokucubungula kusengaphambili ngo-30%.
I-FusionCube A3000 Training/Inference Super-Converged Appliance, edizayinelwe izimo zokuqeqeshwa/okucatshangwayo ezinkulu zezinga lemboni, ibhekelela izinhlelo zokusebenza ezibandakanya amamodeli anezigidigidi zamapharamitha. Ihlanganisa izindawo zokugcina ezisebenza kahle kakhulu ze-OceanStor A300, izindawo zokuqeqesha/zeziqondiso, imishini yokushintsha, isofthiwe yesikhulumi se-AI, kanye nesofthiwe yokuphatha nokusebenza, ihlinzeka ngozakwethu abangamamodeli abakhulu ngesipiliyoni sokuphakelwa kwe-plug-and-play ukuze kuthunyelwe indawo eyodwa. Ilungele ukusetshenziswa, ingafakwa phakathi namahora ama-2. Kokubili ama-node okuqeqesha/okulinganisa nawokugcina anganwetshwa ngokuzimela nangokuvundlile ukuze afane nezidingo zesikali samamodeli ahlukahlukene. Ngaleso sikhathi, i-FusionCube A3000 isebenzisa iziqukathi ezisebenza kahle kakhulu ukuze inike amandla ukuqeqeshwa kwamamodeli amaningi kanye nemisebenzi yokukhomba ukwabelana ngama-GPU, ikhulisa ukusetshenziswa kwezinsiza kusuka ku-40% kuya ngaphezulu kwama-70%. I-FusionCube A3000 isekela amamodeli ebhizinisi amabili aguquguqukayo: I-Huawei Ascend One-Stop Solution kanye nesixazululo sokumisa esisodwa sikazakwethu wenkampani yangaphandle ngekhompyutha evulekile, inethiwekhi, nesofthiwe yesikhulumi se-AI.
UMongameli weHuawei Wolayini Womkhiqizo Wokugcinwa Kwedatha, uZhou Yuefeng, uthe, “Enkathini yamamodeli amakhulu, idatha inquma ukuphakama kobuhlakani be-AI. Njengomphathi wedatha, ukugcinwa kwedatha kuba ingqalasizinda eyisisekelo yamamodeli amakhulu e-AI. I-Huawei Data Storage izoqhubeka nokusungula izinto ezintsha, ihlinzeka ngezixazululo nemikhiqizo ehlukahlukene yenkathi yamamodeli amakhulu e-AI, isebenzisana nozakwethu ukushayela ukuhlonyiswa kwe-AI kuzo zonke izimboni eziningi."
Isikhathi sokuthumela: Aug-01-2023
