Akunakwenzeka ukuba nabani na aphikisane nento yokuba ukuyila ii-LSIs zesiko (ASICs) kude nenkqubo elula kwaye ekhawulezayo. Kodwa ndifuna kwaye ndiyidinga ukuba ikhawuleze: namhlanje ndikhuphe i-algorithm, kwaye emva kweveki ndathatha iprojekthi yedijithali egqityiweyo. Inyani yeyokuba ii-LSI ezikhethekileyo ziphantse zibe yimveliso enye. Oku kunqabile ukuba kufuneke kwiibhetshi zezigidi, ekuphuhliseni onokuthi uchithe imali eninzi kunye nezibonelelo zabantu njengoko uthanda, ukuba oku kufuneka kwenziwe ngexesha elifutshane. Ii-ASIC ezikhethekileyo, kwaye ke ngoko ezona zisebenzayo ekusombululeni imisebenzi yazo, kufuneka zifikelele kwixabiso eliphantsi ukuphuhlisa, nto leyo eya iba nkulu kakhulu kwinqanaba langoku lophuhliso lokufunda koomatshini. Ngaphambili, imithwalo eqokelelwe yimarike yekhompyutheni kwaye, ngakumbi, impumelelo ye-GPU kwintsimi yokufunda koomatshini (ML) ayinakuphetshwa.
Ukukhawulezisa uyilo lwe-ASICs kwimisebenzi ye-ML, i-DARPA iseka inkqubo entsha - i-Real Time Machine Learning (RTML). Inkqubo yokufunda yomatshini yexesha langempela ibandakanya ukuphuhlisa umqokeleli okanye iqonga lesoftware elinokuthi liyilwe ngokuzenzekelayo uyilo lwe-chip lwesakhelo esithile seML. Iqonga kufuneka lihlalutye ngokuzenzekelayo i-algorithm yokufunda yomatshini ecetywayo kunye nedatha esetyenzisiweyo yokuqeqesha le algorithm, emva koko kufuneka ivelise ikhowudi kwi-Verilog ukudala i-ASIC ekhethekileyo. Abaphuhlisi be-algorithm ye-ML abanalo ulwazi lwabayili be-chip, kwaye abayili abafane baqhelane nemigaqo yokufunda koomatshini. Inkqubo ye-RTML kufuneka incede ukuqinisekisa ukuba iingenelo zombini zidibene kwi-automated ASIC iqonga lophuhliso lokufunda ngomatshini.
Ngethuba lokuphila kweprogram ye-RTML, izisombululo ezifunyenweyo ziya kufuneka zivavanywe kwiindawo ezimbini eziphambili zesicelo: amanethiwekhi e-5G kunye nokulungiswa komfanekiso. Kwakhona, inkqubo ye-RTML kunye neqonga lesoftware elenziweyo loyilo oluzenzekelayo lwe-ML accelerators luya kusetyenziswa ukuphuhlisa nokuvavanya ii-algorithms zeML ezintsha kunye neeseti zedatha. Ngaloo ndlela, nangaphambi kokuyila i-silicon, kuya kukwazi ukuvavanya amathuba ezicwangciso ezintsha. Iqabane le-DARPA kwiprogram ye-RTML iya kuba yi-National Science Foundation (NSF), ekwabandakanyeka kwiingxaki zokufunda ngomatshini kunye nophuhliso lwe-algorithms ye-ML. Umqambi ophuhlisiwe uya kudluliselwa kwi-NSF, kwaye umva i-DARPA ilindele ukufumana i-compiler kunye neqonga lokuyila i-algorithms ye-ML. Kwixesha elizayo, i-hardware yoyilo kunye nokudalwa kwe-algorithms iya kuba sisisombululo esidibeneyo, esiza kukhokelela ekuveleni kweenkqubo zoomatshini abazifundela ngokwabo ngexesha langempela.
umthombo: 3dnews.ru