Ukuthuthuka kwe-Supercomputing: kusetshenziswa amanethiwekhi optical neuromorphic

Emashumini ambalwa eminyaka adlule, inkambiso eyake yaziwa futhi inembayo, uMthetho kaMoore, owabikezelwa uGordon Moore we-IBM ngo-1965, manje isiba isilinganiso esingasebenzi sokusebenza kwekhompiyutha. Umthetho kaMoore wabikezela ukuthi cishe njalo eminyakeni emibili inani lama-transistors kusekethe edidiyelwe lizophindeka kabili, ukuthi kuzoba nama-transistors amaningi endaweni efanayo, okuholela ekwandeni kwezibalo kanjalo nokusebenza kwekhompyutha. Ngo-April 2005, engxoxweni, uGordon Moore ngokwakhe wathi ukubikezela kwakhe ngeke kusakwazi ukuqhubeka: “Ngobukhulu [bama-transistors] ungabona ukuthi sisondela kusayizi wama-athomu okuyisithiyo esiyisisekelo, kodwa kuyoba yizizukulwane ezimbili noma ezintathu ngaphambi kokuba sifike lapho—kodwa lokho kukude ngendlela esike sakwazi ukukubona ngayo. Sineminye iminyaka engu-10 kuya kwengu-20 ngaphambi kokuba sifinyelele umkhawulo obalulekile.”   

Yize umthetho kaMoore ulindeleke ukuthi ufe, ezinye izinkomba zekhompiyutha zibona ukukhula kokusebenza. Ngobuchwepheshe esibusebenzisa ezimpilweni zethu zansuku zonke, sonke singabona amathrendi amakhompuyutha aba mancane futhi abe mancane kodwa nokuthi amabhethri edivayisi ahlala isikhathi eside futhi isikhathi eside. Umkhuba wakamuva wamabhethri ubizwa ngokuthi uMthetho ka-Koomey, oqanjwe ngoprofesa wase-Stanford University u-Jonathan Koomey. Umthetho ka-Koomey ubikezela ukuthi "... ekulayisheni okugxilile kwekhompuyutha, inani lebhethri olidingayo lizokwehla ngokuphindwe kabili njalo ngonyaka nengxenye." Ngakho-ke, ukusetshenziswa kwamandla kagesi noma ukusebenza kahle kwamandla kwamakhompyutha kuphindaphindeka cishe njalo ezinyangeni eziyi-18. Ngakho-ke, lokho okukhomba khona futhi okuvezwa yizo zonke lezi zindlela kanye nezinguquko kuyikusasa lekhompyutha.

Ikusasa lekhompyutha

Sifike esikhathini emlandweni lapho kufanele sichaze kabusha ikhompuyutha njengoba amathrendi nemithetho eyabikezelwa emashumini ambalwa eminyaka adlule ingasasebenzi. Futhi, njengoba ikhompiyutha iqhubekela esikalini se-nano nese-quantum, kunemikhawulo esobala engokomzimba nezinselele ezisazofika. Mhlawumbe umzamo ophawuleka kakhulu wokusebenzisa i-supercomputing, i-quantum computing, unenselele esobala yokusebenzisa ngempela ukubopha kwe-quantum ukuze kubalwe ngokuhambisanayo, okungukuthi, ukwenza izibalo ngaphambi kokuhlukaniswa kwe-quantum. Nokho, naphezu kwezinselelo ze-quantum computing kube nenqubekelaphambili enkulu emashumini ambalwa eminyaka adlule. Umuntu angathola amamodeli okwakhiwa kwekhompyutha yendabuko kaJohn von Neumann asetshenziswa ku-quantum computing. Kodwa kunomunye umkhakha ongaziwa kangako we-(super)computing, obizwa nge-neuromorphic computing engalandeli i-architecture yendabuko ye-von Neumann. 

I-Neuromorphic computing yacatshangwa nguprofesa waseCaltech u-Carver Mead emuva ephepheni lakhe lokuqala ngo-1990.  Ngokuyisisekelo, imigomo yekhompuyutha ye-neuromorphic isekelwe ezimisweni zesenzo se-biological, njengalezo okucatshangwa ukuthi zisetshenziswa ubuchopho bomuntu ekubaleni. Umehluko omfishane phakathi kwethiyori yekhompyutha ye-neuromorphic ngokumelene nethiyori yekhompyutha ye-von Neumann yafingqiwe esihlokweni sikaDon Monroe encwadini ethi. Inhlangano Yezemishini Yekhompyutha iphephabhuku. Isitatimende sihamba kanje: “Esakhiweni sendabuko se-von Neumann, i-logic core (noma eminingana ngokufana) isebenza ngokulandelana kudatha elandwe enkumbulweni. Ngokuphambene, i-‘neuromorphic’ computing isakaza kokubili ukubala kanye nenkumbulo phakathi kwenani elikhulu ‘lama-neurons’ akudala uma kuqhathaniswa, ngalinye lixhumana namakhulu noma izinkulungwane zamanye ama-neurons ‘ngama-synapse.’”  

Ezinye izici ezibalulekile zekhompiyutha ye-neuromorphic zifaka ukungabekezelelani kwamaphutha, okuhlose ukumodela ikhono lobuchopho bomuntu lokulahlekelwa ama-neurons futhi likwazi ukusebenza. Ngokufanayo, kukhompyutha yendabuko ukulahlekelwa kwe-transistor eyodwa kuthinta ukusebenza kahle. Enye inzuzo ecatshangwayo nehlosiwe ye-neuromorphic computing asikho isidingo sokuhlelwa; le nhloso yokugcina iphinde ifanekise ikhono lobuchopho bomuntu lokufunda, ukuphendula nokujwayela amasignali. Ngakho-ke, i-neuromorphic computing njengamanje iyikhandidethi elingcono kakhulu lokufunda ngomshini nemisebenzi yobuhlakani bokwenziwa. 

Intuthuko ye-neuromorphic supercomputing

Okunye kwalesi sihloko kuzodingida intuthuko ye-neuromorphic supercomputing. Ngokuqondile, ucwaningo olusanda kushicilelwa ku-Arxiv oluvela ku-Alexander Tait et. al. ephuma eNyuvesi yasePrinceton ibonisa ukuthi imodeli yenethiwekhi ye-silicon-based photonic neural network idlula indlela evamile yokwenza ikhompuyutha cishe izikhathi ezingu-2000. Le nkundla ye-neuromorphic photonic yekhompuyutha ingaholela ekucutshungulweni kolwazi okushesha kakhulu. 

I-Tait et. al. iphepha elinesihloko I-Neuromorphic Silicon Photonics iqala ukuchaza okuhle nobubi bokusebenzisa uhlobo lokukhanya kwesithombe semisebe kagesi ukwenza ikhompuyutha. Amaphuzu ayinhloko okuqala ephepha ukuthi ukukhanya sekusetshenziswe kakhulu ekudluliseni ulwazi kodwa hhayi ekuguquleni ulwazi, okungukuthi i-digital optical computing. Ngokufanayo, ku-quantum computing, kunezinselelo zomzimba ezibalulekile ku-digital optical computing. Iphepha libe selingena emininingwaneni yesikhulumi sekhompiyutha esihlongoziwe se-neuromorphic photonic i-Tait et. al. ithimba elishicilelwe ngo-2014, elinesihloko Ukusakaza nesisindo: Inethiwekhi ehlanganisiwe yokucubungula i-photonic spike. Iphepha labo elisha lichaza imiphumela yokuboniswa kokuqala kokuhlolwa kwenethiwekhi ye-photonic neural ehlanganisiwe. 

Esakhiweni sekhompuyutha "sokusakaza nesisindo", "amanodi" anikezwa "isithwali se-wavelength" esiyingqayizivele okungukuthi "i-wavelength division multiplexed (WDM)" bese isakazwa kwamanye "amanodi". “Amanodi” kulo mdwebo ahloselwe ukulingisa ukuziphatha kwe-neuron ebuchosheni bomuntu. Bese amasiginali e-“WDM” acutshungulwa ngezihlungi ezinenani eliqhubekayo ezibizwa nge-“microring (MRR) weight banks” bese zifingqwa ngogesi enanini elilinganisiwe lokuthola amandla. Ukungalandeli komugqa kwalokhu kuguqulwa kwe-electro-optic yokugcina/ukubala ngekhompyutha kuwukungalandeli komugqa okudingekayo ukuze kulingise ukusebenza kwe-neuron, okubalulekile kukhompyutha ngaphansi kwezimiso ze-neuromorphic. 

Ephepheni, baxoxa ngokuthi lawa mandla okuguqulwa kwe-electro-optic aqinisekisiwe ngokwezibalo afana nemodeli "ye-2-node continuous-time recurrent neural network" (CTRNN). Le miphumela yobuphayona iphakamisa ukuthi amathuluzi okuhlela asetshenziselwe amamodeli e-CTRNN angasetshenziswa kumapulatifomu e-neuromorphic asekelwe ku-silicon. Lokhu kutholwa kuvula indlela yokujwayela indlela yokwenza ye-CTRNN ku-neuromorphic silicon photonics. Ephepheni labo, benza okufanayo okuyimodeli ehambisana nesakhiwo sabo "sokusakazwa nesisindo". Imiphumela ibonisa ukuthi imodeli ye-CTRNN efaniswe ekwakhiweni kwayo kwamanodi angu-49 ikhiqiza i-neuromorphic computing architecture ukuze idlulele amamodeli ekhompuyutha yakudala ngama-oda angu-3 wobukhulu.