Tabbatar da Amfani da Makamashi na AI: Tabbatar da CodeCarbon da Ma'aunai na Waje

Teburin Abubuwan Ciki

Kuskuren Kiyasi

Har zuwa 40%

Matsakaicin bambanci daga ma'aunai na gaskiya

Gwaje-gwaje

Daruruwa

An gudanar da gwaje-gwajen AI don tabbatarwa

Amfani da Kayan Aiki

2M+

Zazzagewar CodeCarbon akan PyPI

1 Gabatarwa

Hankalin dan Adam na gabatar da manyan kalubale na muhalli duk da yuwuwar kirkire-kirkire sa. Ci gaban saurin samfuran ML ya haifar da matukar damuwa game da amfani da makamashi, tare da kayan aikin kiyasi na yanzu suna yin zato mai ma'ana wanda zai iya lalata daidaito. Wannan binciken yana tabbatar da tsarin kiyasin makamashi na tsaye da na motsi da ma'aunai na gaskiya.

2 Hanyar Bincike

2.1 Tsarin Gwaji

Tsarin tabbatarwa ya ƙunshi ɗaruruwan gwaje-gwajen AI a cikin ayyukan hangen nesa da sarrafa harshe. An gudanar da gwaje-gwaje ta amfani da girman samfura daban-daban daga sigogi miliyan 10 zuwa biliyan 10 don ɗaukar tasirin sikelin.

2.2 Tsarin Ma'auni

An sami ma'aunin makamashi na gaskiya ta amfani da ma'aunin wutar lantarki na kayan aiki da kayan aikin sa ido na tsarin. An gudanar da nazarin kwatancen tsakanin hanyoyin kiyasi na tsaye (ML Emissions Calculator) da na motsi (CodeCarbon).

3 Sakamako da Bincike

3.1 Daidaiton Kiyasi

Dukansu kayan aikin kiyasi sun nuna bambance-bambance masu mahimmanci daga ma'aunai na gaskiya. ML Emissions Calculator ya nuna tsarin ƙima da wuce gona da iri daga -40% zuwa +60% a cikin nau'ikan samfura da girmansu daban-daban.

3.2 Tsarin Kurakurai

Samfuran hangen nesa sun nuna tsarin kurakurai daban-daban idan aka kwatanta da samfuran harshe. CodeCarbon gabaɗaya yana ba da ƙima madaidaiciya amma har yanzu yana nuna kurakurai na yau da kullun har zuwa 40% a wasu saituna.

Mahimman Fahimta

Hanyoyin kiyasi na tsaye sun fi saurin yin manyan kurakurai tare da samfura masu sarƙaƙiya
Bibiyar motsi tana ba da mafi kyawun daidaito amma har yanzu tana da son rai na yau da kullun
Gine-ginen samfur yana tasiri sosai ga daidaiton kiyasi
Bambance-bambancen saitin kayan aiki suna ba da gudummawa sosai ga kurakuran kiyasi

4 Aiwarta Fasaha

4.1 Tsarin Lissafi

Za a iya ƙirƙirar amfani da makamashi na samfuran AI ta amfani da ma'auni mai zuwa:

$E_{total} = \sum_{i=1}^{n} P_i \times t_i + E_{static}$

Inda $P_i$ ke wakiltar amfani da wutar lantarki na bangaren i, $t_i$ shine lokacin aiwatarwa, kuma $E_{static}$ yana lissafin amfani da makamashin tsarin tushe.

4.2 Aiwarta Code

Aiwatar da tushen bibiyar amfani da makamashi ta amfani da CodeCarbon:

from codecarbon import track_emissions

@track_emissions(project_name="ai_energy_validation")
def train_model(model, dataset, epochs):
    # Code na horar da samfur
    for epoch in range(epochs):
        for batch in dataset:
            loss = model.train_step(batch)
    return model

# Bibiyar amfani da makamashi
with EmissionsTracker(output_dir="./emissions/") as tracker:
    trained_model = train_model(resnet_model, imagenet_data, 100)
    emissions = tracker.flush()

5 Aiwatar da Gaba

Za a iya faɗaɗa tsarin tabbatarwa zuwa wasu fagage ciki har da koyo mai ƙarfi da samfuran ƙirƙira. Aikin gaba ya kamata ya mayar da hankali kan ingantaccen amfani da makamashi na ainihi da ƙirar samfur mai sane da kayan aiki. Haɗawa da tsarin koyo na tarayya zai iya ba da damar sa ido kan amfani da makamashi a cikin na'urori masu gefe.

Bincike na Asali: Kalubalen Kiyasin Makamashin AI da Damarmomi

Binciken wannan binciken yana nuna muhimman kalubale a cikin kiyasin makamashin AI waɗanda suka yi daidai da batutuwa a wasu fagagen lissafi. Kurakuran kiyasi na 40% da aka lura suna da damuwa musamman idan aka yi la'akari da haɓakar buƙatun lissafin AI da masu bincike irin su Amodei da Hernandez (2018) suka rubuta, waɗanda suka gano buƙatun lissafin AI suna ninka kowane wata 3.4. Kamar yadda CycleGAN (Zhu et al., 2017) ta kawo juyin juya hali a fassarar hoto ta hanyar cibiyoyin sadarwa masu juyi, muna buƙatar sabbin abubuwa na asali a hanyoyin auna makamashi.

Kurakuran da aka gano a cikin hanyoyin kiyasi na tsaye da na motsi suna nuna cewa kayan aikin na yanzu sun kasa ɗaukar muhimman hulɗar kayan aiki-software. Kamar yadda aka lura a cikin Rahoton Amincin AI na Duniya (2023), dole ne dorewar muhalli ta zama babban abin la'akari a cikin ci gaban AI. Tsarin da aka lura a cikin wannan binciken yana kama da kalubalen farko a cikin hasashen aikin gine-ginen kwamfuta, inda samfurai masu sauƙi sau da yawa suka kasa yin lissafin halayen cache masu sarƙaƙi da matakan ƙwaƙwalwar ajiya.

Idan aka duba binciken dorewar lissafi mai faɗi, Ƙungiyar Aiki mai Tasirin Kuzarin Lissafi ta Koli ta kafa ma'auni don auna ingancin lissafi wanda zai iya ba da labari game da bibiyar makamashin AI. $E_{total} = \sum P_i \times t_i + E_{static}$ da aka yi amfani da shi a cikin wannan binciken yana ba da tushe mai ƙarfi, amma aikin gaba ya kamata ya haɗa da ƙarin samfura masu rikitarwa waɗanda ke lissafin sikelin ƙarfin lantarki da mitar motsi, matsi na zafi, da matsanancin bandeji na ƙwaƙwalwar ajiwa.

Tsarin tabbatarwa na binciken yana wakiltar wani muhimmin mataki zuwa ga daidaitaccen kimanta makamashin AI, kamar yadda ImageNet ta daidaita ma'aunin hangen nesa na kwamfuta. Yayin da samfuran AI ke ci gaba da haɓaka—tare da tsarin kwanan nan kamar GPT-4 wanda aka kiyasta yana amfani da makamashi daidai da ɗaruruwan gidaje—daidaitaccen kiyasin makamashi ya zama mahimmanci ga ci gaba mai dorewa. Kayan aikin gaba ya kamata su koyi daga ƙirar ƙarfi a cikin lissafi mai ƙarfi yayin da suke daidaitawa ga halayen musamman na tunani da horar da hanyoyin sadarwar jijiyoyi.

6 Nassoshi

Amodei, D., & Hernandez, D. (2018). AI da Lissafi. OpenAI Blog.
Zhu, J. Y., Park, T., Isola, P., & Efros, A. A. (2017). Fassarar Hoton-da-Ba-a Haɗa ba ta amfani da Cibiyoyin Sadarwa masu Juyi. ICCV.
Rahoton Amincin AI na Duniya (2023). Hadurori na Tsari da Dorewar Muhalli.
Lacoste, A., et al. (2019). Ƙididdige Hayakin Carbon na Koyon Injin. arXiv:1910.09700.
Schwartz, R., et al. (2020). Green AI. Sadarwar ACM.
Ƙungiyar Aiki mai Tasirin Kuzarin Lissafi ta Koli (2022). Ma'auni don Ma'aunin Ingancin Lissafi.
Anthony, L. F. W., et al. (2020). Carbontracker: Bibiya da Hasashen Sawun Carbon na Horar da Samfuran Koyo mai zurfi. Taron ICML.

Ƙarshe

Wannan binciken ya kafa muhimmin shaida na zahiri don ingancin kiyasin makamashin AI, yana tabbatar da kayan aikin da ake amfani da su yayin da yake gano manyan iyakokin daidaito. Tsarin tabbatarwa da jagororin da aka gabatar suna ba da gudummawa sosai ga koyo mai sane da albarkatu da ci gaban AI mai dorewa.