Onxantathu Be-Iron: Amathuluzi Anamandla Okuhlaziya Ukuhwebelana Ekuthuthukisweni Komkhiqizo we-AI

kanye nokusebenza kwemikhiqizo ye-AI kubandakanya ukwenza uhwebo. Isibonelo, umkhiqizo wekhwalithi ephezulu ungathatha isikhathi esiningi nezisetshenziswa zokwakha, kuyilapho izingcingo eziyinkimbinkimbi zingase zihambe kancane futhi zibize kakhulu. Lokhu kuhwebelana kuwumphumela wemvelo wombono wezomnotho oyisisekelo wokushoda, ukuthi lokho esikufunayo okungenzeka ukuthi angenamkhawulo kunganeliseka kancane kuphela isethi elinganiselwe yezinsiza ezitholakalayo. Kulesi sihloko, sizoboleka uhlaka lukanxantathu olunembile kuthiyori yokuphatha iphrojekthi ukuze sihlole ukuhwebelana okubalulekile abakhi nabasebenzisi bemikhiqizo ye-AI okufanele bazulazule ngesikhathi sokuklama nesikhathi sokusebenza, ngokulandelana.

Qaphela: Zonke izibalo namafomula ezigabeni ezilandelayo zidalwe umbhali wale ndatshana.

I-Primer ku-Iron Triangles

Ukungezwani phakathi kobubanzi bephrojekthi, izindleko, nesikhathi kuye kwafundwa kakhulu izifundiswa nabasebenzi emkhakheni wokuphathwa kwamaphrojekthi kusukela okungenani ngeminyaka yawo-1950. Imizamo yokumela ngokubonakalayo ukungezwani (noma ukuhwebelana) phakathi kwalezi zilinganiso zekhwalithi ezintathu kubangele uhlaka lukanxantathu oluhamba ngamagama amaningi, okuhlanganisa “unxantathu wensimbi,” “i-triple constraint,” kanye “nonxantathu wokuphatha iphrojekthi.”

Uhlaka luveza amaphuzu ambalwa abalulekile:

• Kubalulekile ukuhlaziya ukuhwebelana phakathi kwephrojekthi ububanzi (iziphi izinzuzo, izici ezintsha, noma ukusebenza okuzolethwa iphrojekthi), izindleko (ngokwesabelomali sezimali, umzamo womuntu, izindleko ze-IT), kanye isikhathi (uhlelo lwephrojekthi, isikhathi sokulethwa).
• Izindleko zephrojekthi ziwumsebenzi wobubanzi nesikhathi (isb, amaphrojekthi amakhulu kanye nezinhlaka zesikhathi sokulethwa ezimfushane zizobiza ngaphezulu), futhi ngokwalokho okubizwa ngokuthi umthetho ojwayelekile webhalansi yebhizinisi“uthola lokho okukhokhelayo.”
• Esimweni lapho izinsiza ziyivelakancane, kungase kube nzima ukunciphisa izindleko kanye nesikhathi ngesikhathi esisodwa ngenkathi wandisa ububanzi. Lesi simo sithwetshulwe kahle inkulumo ethi “Kuhle, kuyashesha, kushibhile. Khetha okubili,” okuvame ukubizwa ngokuthi (nakuba ngaphandle kobufakazi obuqinile) kumhlaziyi wobuciko we-Victorian u-John Ruskin. Ngakho-ke abaphathi bephrojekthi bavame ukuphaphamela kakhulu ukuhamba kwe-scope (ukwengeza izici eziningi kububanzi bephrojekthi kunalokho okwakuvunyelwene ngakho ngaphambili ngaphandle kokuphatha okwanele), okungabangela ukubambezeleka kwephrojekthi kanye nokweqiwa kwesabelomali.
• Kunoma iyiphi iphrojekthi enikeziwe, kungase kube namazinga ahlukahlukene okuguquguquka kumazinga wobubanzi, izindleko, nesikhathi abhekwa njengesamukelekayo ababambiqhaza. Ngakho-ke kungase kwenzeke ukulungisa okukodwa noma ngaphezulu kwalezi zilinganiso ukuze kutholwe ukucushwa okuhlukile okwamukelekayo kwephrojekthi.

Ividiyo elandelayo ichaza ukusetshenziswa kohlaka lukanxantathu ekuphathweni kwephrojekthi ngokuningiliziwe:

Kumongo wokuthuthukiswa komkhiqizo we-AI, uhlaka lukanxantathu luzibophezela ekuhloleni ukuhwebelana kokubili ngesikhathi sokuklama (lapho umkhiqizo we-AI wakhiwe), nangesikhathi sokusebenza (lapho umkhiqizo we-AI usetshenziswa amakhasimende). Ezigabeni ezilandelayo, sizobheka kabanzi kulesimo ngasinye salezi zimo ezimbili ngokulandelana.

Ukuhwebelana Ngesikhathi Sokuklama

Umfanekiso 1 ubonisa okuhlukile kukanxantathu wensimbi othwebula ukuhwebelana okubhekene neqembu lomkhiqizo we-AI ngesikhathi sokuklama.

Izinhlangothi ezintathu zikanxantathu ziyi:

• Ububanzi besici (S) yomkhiqizo we-AI olinganiswa ngamaphoyinti endaba, amaphuzu okusebenza, noma amayunithi esici.
• Izindleko zokuthuthukisa (C) ngokwezinsuku zomuntu zomzamo womuntu (PM, engineering, UX, data science), kanye nezindleko zezimali zabasebenzi (abathuthukisi abanokuhlangenwe nakho bangase babe nezindleko ezigcwele ngokugcwele) kanye ne-IT (izinsiza zamafu, ama-GPU okuqeqesha amamodeli e-AI).
• Isikhathi sokumaketha (T), isb, emavikini noma ezinyangeni.

Singenza ithiyori ngale modeli encane elandelayo yesithiyo esiphindwe kathathu ngesikhathi sokuklama:

Izindleko zokuthuthukiswa zilingana nesilinganiso sobubanzi nesikhathi, kanye k isici esihle sesikali esimelela ukukhiqiza. Inani eliphezulu le k kusho izindleko eziphansi zesikhathi sokuklama ngobubanzi beyunithi ngayinye ngesikhathi seyunithi, kanjalo nokukhiqiza okukhulu kwesikhathi sokuklama. Imodeli ifana nomuzwa wethu oyisisekelo: as T ijwayele ukungapheli (noma S ifinyelela ku-zero), C ivame ukuba uziro (okungukuthi, ukwelula umugqa wesikhathi wephrojekthi noma ukunciphisa ububanzi kwenza iphrojekthi ishibhile).

Ngokwesibonelo, ake sithi iphrojekthi yethu iqukethe ukwakha umkhiqizo we-AI obiza amaphuzu ezindaba angu-300, esikhathini esiyizinsuku eziyi-100, nezici zokukhiqiza ezingu-0.012. Uma kubhekwa izindleko ezilayishwe ngokugcwele ezingu-$500 ngephuzu lendaba ngayinye, imodeli encane iphakamisa ukuthi kufanele senze isabelomali esilinganiselwa ku-$125k ukuze sithumele umkhiqizo:

Imodeli encane ihlanganisa umongo ofana nefiziksi wesikhathi esiphindwe kathathu sokudizayina. Ngempela, imodeli isikhumbuza isibalo esifundiswa esikoleni esixhumanisa ibanga (d), isivinini (v), kanye nesikhathi

Izinguqulo ezandisiwe zemodeli yesikhathi sokuklama zingacabangela:

• Izindleko ezingaguquki (isb, i-overhead eyisisekelo yokuhlela, ukuphatha, ukuhlinzekwa kwengqalasizinda), okusho ukuthi isibopho esiphansi sezindleko zesikhathi sokuklama sezizonke.
• Umthelela olinganiselwe wokwandisa abasebenzi ngaphezu kwephuzu elithile. Njengoba kuphawulwe nguFred Brooks encwadini yakhe yango-1975 Inyanga Yomuntu Wenganekwane“Ukwengeza abasebenzi kuphrojekthi yesofthiwe engemuva kwesikhathi kuyenza kamuva.”
• Ukukhiqiza okungewona umugqa (isb, ngenxa yokushesha noma ukwehla ezigabeni ezihlukene zephrojekthi), okungaba nomthelela kubudlelwano phakathi kwezindleko nesilinganiso sesikhathi sobubanzi.
• Ukubalwa kwezimali okusobala kwamazinga ekhwalithi ye-AI ukuze kuvunyelwe ukulandelwa okusobala kwamamethrikhi empumelelo (isb, ukubambelela ezidingweni zokulawula nezivumelwano zeleveli yesevisi namakhasimende). Okwamanje, ukubalwa kwezimali kwenzeka ngokungaqondile ngokuchazwa kwesici sokukhiqiza kanye nobubanzi.
• Ubudlelwano phakathi kokukhiqiza kanye nejika lokufunda leqembu lomkhiqizo we-AI, njengesipiliyoni, ukuphindaphinda kwenqubo, nokusetshenziswa kabusha kwekhodi kwenza ukuthuthukiswa kusebenze kahle ngokuhamba kwesikhathi.
• Ukubalwa kwezimali kwenani lilonke (okungukuthi, izinzuzo kususwe izindleko) noma imbuyiselo ku-investimenti (ROI) kunezindleko zokuthuthukisa zodwa.
• Ukugxila ekwabelaneni ngezinsiza ezishodayo kuyo yonke imikhiqizo ye-AI eyathuthukiswa ngokuhambisana. Lokhu kuzobandakanya ukuthatha umbono wephothifoliyo wemikhiqizo ye-AI esathuthukiswayo nganoma yisiphi isikhathi.

I-Trade-Offs ngesikhathi Sokuqalisa

Umfanekiso 2 ubonisa okuhlukile kukanxantathu wensimbi othwebula ukuhwebelana abhekene namakhasimende noma abasebenzisi bomkhiqizo we-AI ngesikhathi sokusebenza.

Ubukhulu obuthathu balo nxantathu yizi:

• Ikhwalithi yokuphendula (Q) yomkhiqizo we-AI ukalwe ngokuya ngokunemba kokubikezela, isikolo se-BLEU/ROUGE, noma enye imethrikhi yekhwalithi ethile yomsebenzi othile.
• Izindleko zokukhomba (C) ngokuya ngamadola noma amasenti ngekholi ngayinye ecatshangwayo, amasekhondi e-GPU aguqulelwe kumadola, noma izindleko zamandla.
• Ukubambezeleka kokuqagela (L) ngama-millisecond, imizuzwana, njll.

Singenza ithiyori ngale modeli encane elandelayo yesithiyo esiphindwe kathathu ngesikhathi sokusebenza:

Izindleko zokucatshangelwa zilingana nesilinganiso sekhwalithi yokuphendula nokubambezeleka, kanye k isici esihle se-scalar esimelela ukusebenza kahle kwesistimu. Inani eliphezulu le k kusho izindleko eziphansi zekhwalithi yokuphendula efanayo nokubambezeleka. Futhi, imodeli ihambisana nokuzwa kwethu okuyisisekelo: njenge L ijwayele ku-zero (noma Q ijwayele ukungapheli), C ivame ukungapheli (okungukuthi, umkhiqizo we-AI obuyisela ngesikhathi sangempela, izimpendulo zekhwalithi ephezulu zizobiza kakhulu kunomkhiqizo ofanayo oletha izimpendulo ezihamba kancane, eziphansi).

Isibonelo, ake sithi umkhiqizo we-AI uzuza ngokungaguquki ukunemba kokubikezela okungu-90% ngokubambezeleka okumaphakathi kwamasekhondi angu-0.5. Uma sithatha isici esisebenza kahle esingu-180, singalindela ukuthi izindleko zokucatshangelwa zibe cishe isenti elilodwa:

Izinguqulo ezinwetshiwe zemodeli yesikhathi sokusebenza zingacabangela:

• Izindleko ezigxilile zesisekelo (isb, zokulayisha imodeli, ukucutshungulwa kwangaphambili nangemuva kwezicelo zabasebenzisi).
• Izindleko zokukala eziguquguqukayo ngenxa yobudlelwano obungewona umugqa phakathi kwezindleko nekhwalithi (isb, ukusuka ku-80% ukuya ku-95% ukunemba kungase kube lula kunokusuka ku-95% kuya ku-99%). Lokhu kungase futhi kuthwebule uhlobo lwembuyiselo enciphayo ekulungiselelweni okulandelanayo komkhiqizo.
• Imvelo yekhwalithi ye-Stochastic, engahluka kuye ngokufaka (“udoti ungene, udoti uphume”). Lokhu kungenziwa ngokusebenzisa inani elilindelekile lekhwalithi, E(Q)esikhundleni senani eliphelele kumodeli yemikhawulo emithathu; bona lesi sihloko ukuze uthole ukujula okujulile ekuhlaziyweni kwenani elilindelwe ekuphathweni komkhiqizo we-AI.
• I-latency overheads engaguquki futhi eguquguqukayo. Izindleko zokukhomba zingamodelwa njengomsebenzi we ngempumelelo ukubambezeleka, ukubalwa kwemali ngokubambezeleka komugqa, ama-network hop, njll.
• Imithelela yokuphuma nokuvumelana. Izindleko zokucatshangelwa ngakunye zingaba ngaphansi kokucatshangelwayo okuqoqiwe (ngenxa yohlobo oluthile lokuncishiswa kwezindleko kukho konke okucatshangwayo ngenqwaba) noma ngaphezulu uma kunokuminyana kwenethiwekhi.
• Ukubala okusobala kokusebenza kahle kwengxenye ye-algorithm ye-AI (ngenxa yemodeli yezakhiwo ezithuthukisiwe, ukusetshenziswa kokuthena, noma ukulinganisa), ihadiwe (ukusebenza kwe-GPU/TPU), namandla (ukusetshenziswa kukagesi nge-FLOP ngayinye) ngokubola isici esihle k ngokufanele.
• Ukujwayela okunamandla kwesici sokusebenza kahle k ngokuphathelene nokulayisha, izingxenyekazi zekhompuyutha, noma uhlobo/izinga lokulungiselelwa. Isb, ukusebenza kahle kungase kuthuthuke nge-caching noma imodeli yokucwiliswa emanzini bese konakaliswa ngaphansi komthwalo osindayo ngenxa yokuphusha kwensiza noma ukuvinjwa.

Okokugcina, izinqumo ezenziwa ngesikhathi sokuklama zingalolonga isimo nezinhlobo zezinqumo ezingenziwa ngesikhathi sokusebenza. Isibonelo, ithimba lomkhiqizo lingase likhethe ukutshala izinsiza ezibalulekile ekuqeqesheni imodeli yesisekelo esibanzi, enganwetshwa ngokufunda okungaphakathi kokuqukethwe ngesikhathi sokusebenza; uma kuqhathaniswa ne-algorithm yokufunda yomshini evamile njengehlathi elingahleliwe, imodeli yesisekelo iwukukhetha kwesikhathi sokuklama okungase kuvumele ikhwalithi engcono yokuphendula ngesikhathi sokusebenza, nakuba ngezindleko eziphakeme ezingase zibe phezulu. Ukutshalwa kwezimali kwesikhathi sokuklama kukhodi ehlanzekile nengqalasizinda ephumelelayo kungakhuphula isici sokusebenza kahle kwesistimu yesikhathi sokusebenza. Ukukhethwa komhlinzeki wamafu kungase kunqume inani elincane lezindleko ezingafinyeleleka ngesikhathi sokusebenza. Ngakho-ke kubalulekile ukucabangela ukuklama kanye nokuhwebelana kwesikhathi sokusebenza ngokubambisana ngendlela ephelele.

Ukugoqa

Njengoba lesi sihloko sibonisa, unxantathu wensimbi ovela kuthiyori yokuphatha iphrojekthi ungasetshenziswa kabusha ukuze kukhiqizwe izinhlaka ezilula kodwa ezinamandla zokuhlaziya ukwakheka kanye nokuhwebelana kwesikhathi ekuthuthukisweni komkhiqizo we-AI. Unxantathu wensimbi wesikhathi sokuklama ungasetshenziswa amaqembu omkhiqizo ukwenza izinqumo mayelana nesabelomali, ukwabiwa kwezinsiza, nokuhlela ukulethwa. Unxantathu wensimbi ohambisanayo wesikhathi sokusebenza unikeza imininingwane eminingana yokuthi ubudlelwano phakathi kwezindleko zokucatshangelwa, ikhwalithi yokuphendula, nokubambezeleka kungathinta ukwamukelwa komkhiqizo nokwaneliseka kwekhasimende. Njengoba izinqumo zesikhathi sokuklama zingacindezela ukukhetha kwesikhathi sokusebenza, kubalulekile ukucabanga ngokuklama kanye nokuhwebelana kwesikhathi sokusebenza ngokubambisana kusukela ekuqaleni. Ngokubona ukuhwebelana kusenesikhathi nokusebenza eduze kwakho, amaqembu emikhiqizo namakhasimende awo angakha inani elengeziwe ekwakhiweni nasekusetshenzisweni kwe-AI.