Umonakalo kwimozulu obangelwa ukukhutshwa kwegesi ye-greenhouse kubonakala ngakumbi kwaye ukunceda uluntu lophando ukuqonda igalelo lobukrelekrele bokwenziwa kutshintsho lwemozulu kunye nokwamkela iiparadigms zophando olutsha apho ukuncitshiswa kokukhutshwa kwezinto ezikhutshwayo Ukuphathwa njengomlinganiselo obalulekileyo wokusebenza, iqela labaphandi be-AI bamazwe ngamazwe kunye noososayensi bedatha baye basebenzisana ukuyila isofthiwe ekwazi ukuqikelela i-carbon footprint yemisebenzi ye-IT.
CodeCarbon Yisoftware yomthombo ovulekileyo yenzelwe ukunceda amashishini abeke iliso kwi-AI carbon footprint.
I-Comet, umboneleli wezisombululo ze-MLOps, usebenzisane ne-consortium yesayensi yedatha kunye neenkampani ze-AI ezivela kwihlabathi jikelele: i-MILA, ilebhu yophando ye-AI ekhokelwa nguYoshua Bengio eMontreal, BCG GAMMA, icandelo lohlalutyo kunye nesayensi yedatha evela eBoston Consulting. Iqela kunye neKholeji yaseHaverford ePennsylvania, ukudala isoftware yomthombo ovulekileyo.
Malunga CodeCarbon
ICodeCarbon yisoftware ngokusekelwe kwipython que iya kuvumela abadwelisi benkqubo ukuba benze ikhowudi yabo isebenze ngakumbi kwaye banciphise inani le-CO2 eyenziwe kusetyenziso lwezixhobo zekhompyuter kwaye iya kubakhuthaza ukuba benze njalo.
Isoftware ayiqikeleli kuphela isixa se-CO2 esiveliswayo ukusetyenziswa kwezixhobo ze-IT, kodwa ikwabonelela abaphuhlisi ngeengcebiso malunga nendlela yokunciphisa izinto ezikhutshwayo ukukhetha isiseko sakho selifu kwimimandla esebenzisa imithombo yamandla aphantsi.
UYoshua Bengio, umseki we-MILA kunye nophumelele iMbasa yeTuring, uthe:
"I-AI yitekhnoloji enamandla kwaye inamandla okwenza okulungileyo, kodwa kubalulekile ukuqaphela impembelelo yayo ekhulayo yokusingqongileyo. Iprojekthi yeCodeCarbon ijolise ngokuchanekileyo ekufezekiseni le njongo kwaye ndiyathemba ukuba ikhuthaza uluntu lwe-AI ukuba lubale, luveze kwaye lunciphise indawo yekhabhoni. ”
USylvain Duranton, uMlawuli oLawulayo kunye noMlingane oPhezulu weBoston Consulting Group (BCG) kunye noMlawuli weHlabathi we-BCG GAMMA, uthe:
Ngokusekwe kwimbali yamva nje, ukusetyenziswa kwe-IT ngokubanzi, kunye ne-AI ngakumbi, kuya kuqhubeka nokukhula ngokukhawuleza kwihlabathi liphela. Kulo mxholo, iCodeCarbon inokunceda imibutho iqinisekise ukuba i-carbon footprint yabo edibeneyo yonyuka kancinane kangangoko.
Kwimeko yophando egxile kufundo olunzulu, ukuqhubela phambili kubukrelekrele bokwenziwa ubukhulu becala kuphunyezwa ngokwakha iimodeli ezinkulu, ukudibanisa iiseti zedatha ezinkulu, kunye nokusebenzisa amandla amakhulu ekhompyuter.
Ukuqeqesha i-algorithm yokufunda enamandla kunokufuna ukusetyenziswa kweekhompyuter ezininzi ngeentsuku okanye iiveki.
Ulwakhiwo olunje ngeVGG, BERT, GPT-2 kunye neGPT-3, ezinezigidi zohlengahlengiso kwaye ziqeqeshwe kwii-GPU ezininzi iiveki ezininzi, oku kunokuba ngumahluko wamakhulu aliqela eekhilogram zeCO -eq.
I-OpenAI's GPT-2 yaqalwa ngo-2019 isekwe kwi-1.5 yeebhiliyoni zeeparamitha, ngelixa i-GPT-3 eyalandelayo yasungulwa kulo nyaka uphelileyo, iiparamitha zayo ezingama-175 zeebhiliyoni ziyenza ibe ngaphezulu kwamaxesha ali-100 makhulu kunowandulelayo. Njengoko iimodeli ezinkulu ziqhubeka nokuqhubela phambili ebaleni, inani lamandla asetyenzisiweyo ukuwaqeqesha nalo liya kwanda.
CodeCarbon inemodyuli yendlela yokulandelela erekhoda ubungakanani bamandla asetyenzisiweyo ngababoneleli abakhulu becomputing yelifu kunye namaziko edatha agcinwe ngasese.
Emva koko, Inkqubo isebenzisa idatha evela kwimithombo yoluntu ukuqikelela umthamo we-CO2 owenziweyo, ukuqinisekisa izibalo zombane womnatha apho izixhobo zixhunyiwe.
Umkhondo uqikelela i-CO2 eveliswe kuvavanyo ngalunye usebenzisa imodyuli ethile ye-AI, ukugcina idatha yokukhutshwa kweeprojekthi kunye nombutho wonke.
Umbono kukuba iCodeCarbon iya kunceda i-IT kunye neenkampani ze-AI ukunciphisa i-carbon footprint njengoko zikhula. I-CodeCarbon iya kuvelisa ideshibhodi eya kuvumela iinkampani ukuba zibone ngokulula ubungakanani bezinto ezikhutshwayo eziveliswa ngokuqeqesha imifuziselo yokufunda koomatshini.
Ukukwazi ukulandelela ukukhutshwa kwe-CO2 kubonisa ukuqhubela phambili okubonakalayo kubuchule babaphuhlisi bokusebenzisa izixhobo zamandla ngobulumko kwaye ngoko ke ukunciphisa impembelelo yomsebenzi wabo kwimeko ethe tye.
Umthombo: https://www.comet.ml/