Ukulimala kwesimo sezulu okudalwe ukukhishwa kwegesi ebamba ukushisa kungaphezu kokusobala futhi ukusiza umphakathi ocwaningayo ukuqonda umnikelo wobuhlakani bokufakelwa ekuguqukeni kwesimo sezulu futhi samukele izindlela ezintsha zocwaningo lapho kunciphisa ukungcola Iphathwe njengesilinganiso sokusebenza esibucayi, iqembu labaphenyi be-AI bamazwe omhlaba kanye nososayensi be-data bahlanganyele ekwakheni isoftware ekwazi ukulinganisela i-carbon footprint yemisebenzi ye-IT.
ICodeCarbon isoftware yomthombo ovulekile yakhelwe ukusiza izinkampani ukuqapha i-AI carbon footprint yazo.
UComet, ongumhlinzeki wezixazululo ze-MLOps, uhlanganyele ne-Consortium ye-AI nezinkampani zesayensi yedatha ezivela emhlabeni wonke: I-MILA, ilebhu yocwaningo ye-AI eholwa ngu-Yoshua Bengio eMontreal, BCG GAMMA, uphiko lwe-analytics kanye nesayensi yedatha evela eBoston Ukubonisana neqembu neHaverford College ePennsylvania, ukudala isoftware yomthombo ovulekile.
Mayelana neCodeCarbon
ICodeCarbon iyisoftware python isuselwe lokho kuzovumela abenzi bohlelo ukwenza ikhodi yabo isebenze kahle futhi banciphise inani le-CO2 ekhiqizwayo ukusetshenziswa kwezinsizakusebenza zekhompyutha futhi kuzobagqugquzela ukuthi benze njalo.
Isoftware akugcini nje ngokulinganisa inani le-CO2 elikhiqiziwe ukusetshenziswa kwezinsizakusebenza ze-IT, ibuye inikeze onjiniyela iseluleko sendlela yokunciphisa ukungcola ukukhetha ingqalasizinda yakho yamafu ezifundeni ezisebenzisa imithombo yamandla aphansi.
U-Yoshua Bengio, umsunguli we-MILA nowine umklomelo we-Turing Prize, uthe:
“I-AI ingubuchwepheshe obunamandla futhi inamandla okwenza okuhle, kodwa kubalulekile ukuthi wazi ngomthelela wayo okhulayo kwezemvelo. Iphrojekthi yeCodeCarbon ihlose ngqo ukufeza le njongo futhi ngiyethemba izogqugquzela umphakathi we-AI ukuthi ubale, udalule futhi wehlise umkhondo wekhabhoni. ”
USylvain Duranton, uMqondisi Ophethe kanye no-Senior Partner eBoston Consulting Group (BCG) kanye noMqondisi Womhlaba ku-BCG GAMMA, uthe:
“Ngokuya ngomlando wakamuva, ukusetshenziswa kwe-IT ngokujwayelekile, futhi ikakhulukazi i-AI, kuzoqhubeka nokukhula kakhulu emhlabeni jikelele. Kulesi simo, iCodeCarbon ingasiza izinhlangano ukuthi ziqinisekise ukuthi umkhondo wekhabhoni ohlanganayo wenyuka kancane ngangokunokwenzeka ”.
Endaweni yocwaningo egxile ekufundeni, ukuthuthuka kobuhlakani bokufakelwa kutholakala kakhulu ngokwakha amamodeli amakhulu, ukuhlanganisa amasethi wedatha amakhulu, nokusebenzisa amandla amakhulu wekhompyutha.
Ukuqeqesha i-algorithm yokufunda enamandla kungadinga ukusetshenziswa kwamakhompyutha amaningi ezinsukwini noma emasontweni.
Okwezakhiwo ezifana ne-VGG, BERT, GPT-2 ne-GPT-3, ezinezigidi zokulungiselela futhi eziqeqeshwe kuma-GPU amaningi amasonto ambalwa, lokhu kungaba umehluko wamakhilogremu angamakhulu amaningana we-CO-eq.
I-GPT-2 ye-OpenAI ekhishwe ngo-2019 isuselwa kumapharamitha ayi-1.5 billion, kuyilapho owalandela i-GPT-3 akhishwa ngonyaka owedlule, onamapharamitha ayi-175 billion awenza abe ngaphezu kokuphindwe kalikhulu. Njengoba amamodeli amakhulu aqhubeka ethuthuka emkhakheni, inani lamandla asetshenzisiwe ukuwaqeqesha nawo azokwanda.
ICodeCarbon inemodyuli yokulandela ngomkhondo eqopha inani lamandla asetshenzisiwe ngabahlinzeki abakhulu bekhompiyutha bamafu nezikhungo zedatha ezisendaweni eyimfihlo.
Khona-ke, uhlelo lusebenzisa idatha evela emithonjeni yomphakathi ukulinganisa ivolumu ye-CO2 eyenziwe, ukuqinisekisa izibalo zenethiwekhi kagesi lapho okokusebenza kuxhunywe khona.
I-tracker ilinganisela i-CO2 ekhiqizelwe ukuhlolwa ngakunye kusetshenziswa imodyuli ethile ye-AI, igcina idatha yokukhishwa kwamaphrojekthi kanye nenhlangano yonke.
Umqondo ukuthi iCodeCarbon izosiza izinkampani ze-IT ne-AI ukunciphisa umkhawulo wazo wekhabhoni njengoba zikhula. ICodeCarbon izokhiqiza ideshibhodi ezovumela izinkampani ukuthi zibone kalula inani lokukhishwa okukhiqizwa ukuqeqesha amamodeli wazo wokufunda ngomshini.
Amandla okulandelela ukukhishwa kwe-CO2 amele inqubekela phambili ebalulekile ekhonweni lonjiniyela lokusebenzisa izinsiza zamandla ngokuhlakanipha futhi ngenxa yalokho anciphise umthelela womsebenzi wabo endaweni ebucayi.
Umthombo: https://www.comet.ml/