Sekit agbakwunyere Photonic, ma ọ bụ ibe ngwa anya, nwere ike inye ọtụtụ uru karịa ndị otu eletrọnịkị ha, dị ka mbelata oriri ike yana mbelata latency na mgbakọ. Ọ bụ ya mere ọtụtụ ndị na-eme nchọpụta ji kwenye na ha nwere ike ịdị irè nke ukwuu na mmụta igwe na ọgụgụ isi (AI). Intel na-ahụkwa nnukwu atụmanya maka iji silicon photonics na ntụziaka a. Ndị otu nyocha ya na usoro ọhụrụ zuru ezu nke nwere ike iweta netwọkụ akwara anya ka ọ bịaruo nso n'eziokwu.
Na nso nso a , raara onwe ya nye mmụta igwe, na-akọwa otú nyocha na ngalaba nke netwọkụ akwara anya si malite. Nnyocha nke David AB Miller na Michael Reck egosila na enwere ike ịhazi ụdị sekit photonic nke a maara dị ka interferometer Mach-Zehnder (MZI) iji mee 2 × 2 matrix multiplication mgbe etinyere MZI na ntupu triangular maka ịba ụba nnukwu matrices, onye nwere ike. nweta sekit nke na-emejuputa algọridim matrix-vector multiplication algọridim, ngụkọ bụ isi eji na mmụta igwe.
Nchọpụta Intel ọhụrụ lekwasịrị anya n'ihe na-eme mgbe ntụpọ dị iche iche nke ibe anya na-enwe ike ịnweta n'oge nrụpụta (ebe ọ bụ na photonics computational bụ analog na okike) na-ebute ọdịiche dị na mgbako n'etiti ibe dị iche iche nke otu ụdị. Ọ bụ ezie na a na-eduzi ọmụmụ ihe ndị yiri ya, n'oge gara aga, ha lekwasịrị anya karịa n'ịkwalite mmepụta ihe iji kpochapụ ihe na-ezighị ezi. Mana usoro a nwere scalability na-adịghị mma ka netwọk na-ebuwanye ibu, na-ebute mmụba nke ike mgbakọ na mwepụ chọrọ iji melite netwọk anya. Kama nkwalite nrụpụta ihe, Intel tụlere ibe ọzụzụ otu oge tupu imepụta ya site na iji ihe owuwu na-anabata mkpọtụ. A zụrụ netwọkụ anya akwara anya otu ugboro, mgbe nke ahụ gasịrị, a na-ekesa usoro ọzụzụ n'ofe ọtụtụ ebe netwọk e kere eke nwere ọdịiche dị na akụkụ ha.
Ndị otu Intel tụlere ụlọ ọrụ abụọ maka iwulite sistemụ ọgụgụ isi dabere na MZI: GridNet na FFTNet. GridNet buru amụma na-edobe MZI na grid, ebe FFTNet na-etinye ha na urukurubụba. Mgbe ha zụchara ha abụọ na ịme anwansị na njirimara akara akara mmụta miri emi (MNIST), ndị nyocha ahụ chọpụtara na GridNet nwetara izi ezi dị elu karịa FFTNet (98% vs. 95%), mana ụlọ ọrụ FFTNet “siri ike nke ukwuu.” N'ezie, arụmọrụ GridNet gbadara n'okpuru 50% yana mgbakwunye nke mkpọtụ arụrụ arụ (nnyonye anya na-egosipụta ntụpọ enwere ike na nrụpụta mgbawa ngwa anya), ebe FFTNet ọ nọgidere na-adịgide adịgide.
Ndị ọkà mmụta sayensị ahụ kwuru na nyocha ha na-atọ ntọala maka usoro ọzụzụ ọgụgụ isi nke nwere ike iwepụ mkpa ọ dị imezi ngwa anya ngwa anya mgbe emechara ha, na-echekwa oge na akụrụngwa bara uru.
"Dị ka ọ dị n'usoro mmepụta ihe ọ bụla, ụfọdụ ntụpọ ga-eme nke pụtara na a ga-enwe obere ọdịiche dị n'etiti ibe nke ga-emetụta izi ezi nke ngụkọta oge," ka Casimir Wierzynski, onye isi nchịkwa nke Intel AI Product Group na-ede. "Ọ bụrụ na ihe ndị na-ahụ maka akwara anya ga-abụ akụkụ kwesịrị ekwesị nke gburugburu ebe obibi ngwaike AI, ha ga-achọ ịkwaga na nnukwu ibe na teknụzụ mmepụta ihe. Nnyocha anyị na-egosi na ịhọrọ ụlọ kwesịrị ekwesị n'ihu nwere ike ịbawanye ohere na ibe ndị na-esi na ya ga-arụ ọrụ a chọrọ, ọbụlagodi n'ichepụta mmepụta ihe. "
N'otu oge ahụ na Intel bụ isi na-eduzi nyocha, onye ndoro-ndoro ochichi MIT PhD Yichen Shen tọrọ ntọala Lightelligence nke dabeere na Boston, nke welitere $ 10,7 nde na ntinye ego na ntinye ego. mgbawa ngwa anya prototype maka mmụta igwe nke dị 100 ugboro ngwa ngwa karịa ibe eletrọnịkị ọgbara ọhụrụ ma na-ebelatakwa ike ike site n'usoro ịdị ukwuu, nke gosipụtara n'ụzọ doro anya ọzọ nkwa nke teknụzụ photonic.
isi: 3dnews.ru