Akunakwenzeka ukuthi noma ubani angaphikisana neqiniso lokuthi ukuklama ama-LSI angokwezifiso (ASICs) kude nenqubo elula nesheshayo. Kodwa ngifuna futhi ngidinga ukuthi kusheshe: namuhla ngikhiphe i-algorithm, futhi ngemva kwesonto ngathatha iphrojekthi yedijithali eqediwe. Iqiniso liwukuthi ama-LSI akhethekile cishe angumkhiqizo owodwa. Lokhu akuvamile ukudingeka ngamaqoqo ezigidi, ekuthuthukisweni ongasebenzisa imali eningi kanye nezinsiza zabantu ngendlela othanda ngayo, uma lokhu kudinga ukwenziwa ngesikhathi esifushane kakhulu. Ama-ASIC akhethekile, ngakho-ke asebenza kahle kakhulu ekuxazululeni imisebenzi yawo, kufanele ashibhile ukuwathuthukisa, osekubalulekile kakhulu kulesi sigaba samanje sokuthuthukiswa kokufunda komshini. Ngaphambili, imithwalo eqoqwe imakethe yekhompyutha futhi, ikakhulukazi, impumelelo ye-GPU emkhakheni wokufunda ngomshini (ML) ngeke isagwenywa.
Ukuze kusheshiswe ukwakhiwa kwama-ASIC wemisebenzi ye-ML, i-DARPA isungula uhlelo olusha - Ukufunda Ngomshini Wesikhathi Sangempela (RTML). Uhlelo lokufunda umshini lwesikhathi sangempela lubandakanya ukwakha umdidiyeli noma inkundla yesofthiwe engaklama ngokuzenzakalela ukwakheka kwe-chip yohlaka oluthile lwe-ML. Inkundla kufanele ihlaziye ngokuzenzakalelayo i-algorithm yokufunda yomshini ehlongozwayo kanye nedatha esethelwe ukuqeqesha le-algorithm, ngemva kwalokho kufanele ikhiqize ikhodi ku-Verilog ukuze idale i-ASIC ekhethekile. Abathuthukisi be-algorithm ye-ML abanalo ulwazi lwabaklami bama-chip, futhi abaklami abavamile ukujwayelana nezimiso zokufunda komshini. Uhlelo lwe-RTML kufanele lusize ukuqinisekisa ukuthi izinzuzo zakho kokubili zihlanganiswa kunkundla yokuthuthukisa ye-ASIC ezenzakalelayo yokufunda komshini.
Ngesikhathi somjikelezo wempilo wohlelo lwe-RTML, izixazululo ezitholakele zizodinga ukuhlolwa ezindaweni ezimbili eziyinhloko zohlelo: amanethiwekhi e-5G nokucubungula izithombe. Futhi, uhlelo lwe-RTML kanye nezinkundla zesofthiwe ezidaliwe zokuklama okuzenzakalelayo kwezisheshisi ze-ML zizosetshenziselwa ukuthuthukisa nokuhlola ama-algorithms amasha namasethi edatha e-ML. Ngakho, ngisho nangaphambi kokuklama i-silicon, kuzokwazi ukuhlola amathuba ezinhlaka ezintsha. Uzakwethu we-DARPA ohlelweni lwe-RTML kuzoba yiNational Science Foundation (NSF), ephinde ebambe iqhaza ezinkingeni zokufunda ngomshini nasekuthuthukisweni kwama-algorithms e-ML. Umhlanganisi othuthukisiwe uzodluliselwa ku-NSF, futhi emuva i-DARPA ilindele ukuthola umdidiyeli nenkundla yokuklama ama-algorithms e-ML. Esikhathini esizayo, ukuklama kwe-hardware nokudala ama-algorithms kuzoba yisisombululo esihlanganisiwe, esizoholela ekuveleni kwezinhlelo zemishini ezizifundelayo ngesikhathi sangempela.
Source: 3dnews.ru