I-AI isebenzisa amanzi angakanani?

I-AI isebenzisa amanzi angakanani?

Ngibheja ukuthi uzwile yonke into kusukela “ku-AI iphuza ibhodlela lamanzi njalo ngemibuzo embalwa” kuya “kungamaconsi ambalwa,”. Iqiniso liyinkimbinkimbi kakhulu. Izigxivizo zamanzi ze-AI ziyashintshashintsha kakhulu ngokusekelwe lapho zisebenza khona, ukuthi ukwaziswa kwakho kude kangakanani, nokuthi isikhungo sedatha sipholisa kanjani amaseva aso. Ngakho-ke, inombolo yesihloko ikhona, kodwa ihlala ngaphakathi kwehlathi lemihume.

Ngezansi ngikhipha izinombolo ezicacile, ezilungele izinqumo, ngichaza ukuthi kungani izilinganiso zingavumelani, futhi ngibonise ukuthi abakhi nabasebenzisi bansuku zonke banganciphisa kanjani ithebhu yamanzi ngaphandle kokuphenduka izindela zokusimama.

Izihloko ongase uthande ukuzifunda ngemva kwalesi:

🔗 Iyini idathasethi ye-AI
Ichaza ukuthi amasethi edatha avumela kanjani ukuqeqeshwa komshini wokufunda nokuthuthukiswa kwemodeli.

🔗 I-AI ibikezela kanjani amathrendi
Ibonisa ukuthi i-AI ihlaziya kanjani amaphethini ukuze ibikezele izinguquko nemiphumela yesikhathi esizayo.

🔗 Ungakala kanjani ukusebenza kwe-AI
Ihlukanisa amamethrikhi abalulekile okuhlola ukunemba, isivinini, nokwethembeka.

🔗 Ungakhuluma kanjani ne-AI
Iqondisa amasu asebenzayo okwazisa ukuze kuthuthukiswe ukucaca, imiphumela, nokuvumelana.

I-AI isebenzisa amanzi angakanani? Izinombolo ezisheshayo ongazisebenzisa 📏

  • Isivinini ngasinye, ububanzi obujwayelekile namuhla: kusukela ku-sub-milliliter yesivinini sombhalo esimaphakathi ohlelweni olulodwa olujwayelekile, kufika ku -millilitha amashumi ukuze uthole impendulo ende, ebala kakhulu kolunye. Isibonelo, ukubalwa kwe-Google kokukhiqiza kubika isivinini sombhalo esimaphakathi ~0.26 mL (kufakwe i-overhead ephelele yokukhonza) [1]. Ukuhlolwa komjikelezo wokuphila kwe-Mistral kufaka impendulo yomsizi wamathokheni angu-400 ku-~45 mL (umbono ongaphansi) [2]. Umongo kanye nemodeli kubaluleke kakhulu.

  • Ukuqeqesha imodeli yesikali somngcele: ingafinyelela ezigidini zamalitha, ikakhulukazi kusukela ekupholiseni kanye namanzi ashunyekiwe ekuphehleni ugesi. Ukuhlaziywa kwezemfundo okucashunwe kabanzi kulinganiselwa kumalitha ayizigidi ezingu-5.4 ukuqeqesha imodeli yekilasi le-GPT, okuhlanganisa namalitha angu-~700,000 adliwe endaweni ukupholisa - futhi kwaphikiswana ngayo ngokushejula okuhlakaniphile ukuze kwehliswe amandla amanzi [3].

  • Izikhungo zedatha ngokujwayelekile: amasayithi amakhulu athatha amakhulu ezinkulungwane zamalitha ngosuku ngokwesilinganiso kuma-opharetha amakhulu, anenani eliphakeme kwamanye amakhampasi kuye ngesimo sezulu kanye nesakhiwo [5].

Masikhulume iqiniso: lezo zibalo zizwakala zingahambisani ekuqaleni. Kunjalo. Futhi kunezizathu ezinhle.

 

I-AI eyomile

Amamethrikhi okusebenzisa amanzi e-AI ✅

Impendulo enhle kokuthi Mangakanani amanzi asetshenziswa yi-AI? kufanele ukhethe amabhokisi ambalwa:

  1. Ukucaca komngcele
    Ingabe kufaka phakathi okupholisa asendaweni , noma namanzi angekho esizeni asetshenziswa izimboni eziphehla ugesi? Umkhuba ongcono kakhulu uhlukanisa ukuhoxiswa kwamanzi ngokumelene nokusetshenziswa kwamanzi kanye nezikophu 1-2-3, ezifana nokubala kwekhabhoni [3].

  2. Ukuzwela kwendawo
    Amanzi nge-kWh ayahlukahluka ngokwesifunda kanye nengxube yegridi, ngakho-ke isisho esifanayo singaba nemiphumela ehlukene yamanzi kuye ngokuthi asetshenziswa kuphi - isizathu esiyinhloko izincwadi ezitusa ukuhlela isikhathi nendawo [3].

  3. Amaqiniso omthwalo womsebenzi
    Ingabe inombolo ibonisa ukwaziswa kokukhiqiza okumaphakathi, okuhlanganisa umthamo wokungenzi lutho kanye nesihloko sesikhungo sedatha, noma isisheshisi kuphela esisezingeni eliphezulu? I-Google igcizelela ukubalwa kwesistimu egcwele (okungenzi lutho, ama-CPU/DRAM, kanye nesikhungo sedatha ngaphezulu) ukuze kucatshangwe, hhayi nje izibalo ze-TPU [1].

  4. Ubuchwepheshe bokupholisa
    Ukupholisa okuhwamukayo, ukupholisa uketshezi oluvala iluphu, ukupholisa komoya, nezindlela ezivelayo eziqonde ku-chip zishintsha ukushuba kwamanzi kakhulu. I-Microsoft yethula imiklamo ehloselwe ukuqeda ukusetshenziswa kwamanzi okupholisa kwezinye izingosi zesizukulwane esilandelayo [4].

  5. Isikhathi sosuku nesizini
    Ukushisa, umswakama, nezimo zegridi zishintsha ukusebenza kahle kokusetshenziswa kwamanzi empilweni yangempela; ucwaningo olulodwa olunomthelela luphakamisa ukuhlela imisebenzi emikhulu lapho futhi lapho amandla amanzi ephansi [3].

Ukukhipha amanzi uma kuqhathaniswa nokusetshenziswa kwamanzi, kuchazwe 💡

  • Ukuhoxiswa = amanzi athathwe emifuleni, emachibini, noma emanzini angaphansi komhlaba (amanye abuyiselwe).

  • Ukusetshenziswa = amanzi angabuyiswanga ngenxa yokuthi ayahwamuka noma ahlanganiswa nezinqubo/imikhiqizo.

Imibhoshongo yokupholisa idla amanzi ngokuhwamuka. Ukukhiqizwa kukagesi kungakhipha amanzi amaningi (ngezinye izikhathi kudle ingxenye yawo), kuye ngokuthi isitshalo nendlela yokupholisa. Amalebula ezinombolo zamanzi e-AI athembekile abikwayo [3].

Lapho amanzi eya khona nge-AI: amabhakede amathathu 🪣

  1. Ububanzi 1 - ukupholisa endaweni
    Ingxenye ebonakalayo: amanzi ahwamuka esikhungweni sedatha ngokwaso. Izinketho zedizayini ezifana ne -evaporative vs. air noma i-closed-loop liquid misa isisekelo [5].

  2. Ububanzi 2 - ukukhiqizwa kukagesi
    I-kWh ngayinye ingaphatha ithegi yamanzi efihliwe; ukuxuba nendawo kunquma ukuthi inani lamalitha nge-kWh lingakanani umthwalo wakho wokusebenza olizuza njengefa [3].

  3. Ububanzi 3 - uchungechunge lokuhlinzeka
    Ukukhiqiza ama-chip kuncike emanzini ahlanzekile kakhulu ekukhiqizweni. Ngeke ukubone kumethrikhi "ngesivinini ngasinye" ngaphandle kokuthi umngcele uhlanganisa ngokusobala imiphumela efakiwe (isb., i-LCA ephelele) [2][3].

Abahlinzeki ngezinombolo, nge-nuance 🧮

  • I-Google Gemini ikhuthaza
    indlela yokunikeza isitaki esigcwele (okuhlanganisa nokungenzi lutho kanye nendawo ephezulu). Ukwaziswa kombhalo omaphakathi ~0.26 mL wamanzi eduze kuka- ~0.24 Wh amandla; izibalo zibonisa ithrafikhi yokukhiqiza kanye nemingcele ebanzi [1].

  • Umjikelezo wokuphila we-Mistral Large 2
    I-LCA ezimele engavamile (ene-ADEME/Carbone 4) iveza ~281,000 m³ yokuqeqeshwa + ukusetshenziswa kwasekuqaleni kanye nomkhawulo wesilinganiso ~45 mL wempendulo 400-token [2].

  • Inhloso ye-Microsoft yokupholisa ngaphandle kwamanzi
    Izikhungo zedatha zesizukulwane esilandelayo zenzelwe ukusebenzisa amanzi angekho ukupholisa, zincike ezindleleni eziqondile kuya kuma-chip; ukusetshenziswa komlawuli kusadinga amanzi athile [4].

  • Isikali esijwayelekile sesikhungo sedatha
    Abasebenzi abakhulu babika esidlangalaleni amakhulu ezinkulungwane zamalitha ngosuku ngokwesilinganiso kumasayithi ngamanye; Isimo sezulu kanye nezinombolo zokuphusha phezulu noma phansi [5].

  • Isisekelo sezemfundo sangaphambilini
    Ukuhlaziywa okubalulekile kwe-“thirsty AI” kulinganisele izigidi zamalitha ukuqeqesha amamodeli ekilasi le-GPT, nokuthi izimpendulo eziphakathi eziyi-10-50 zingalingana cishe elingu-500 mL - kuncike kakhulu ekutheni zisebenza nini/kuphi [3].

Kungani izilinganiso zingavumelani kangaka 🤷

  • Imingcele ehlukene
    Ezinye izibalo zibala kuphela ukupholisa endaweni; ezinye zengeza amanzi kagesi; ama-LCA angase engeze ukukhiqizwa kwama-chip. Ama-apula, ama-orenji, kanye nesaladi yezithelo [2][3].

  • Imithwalo yemisebenzi eyahlukene
    I-prompt yombhalo omfushane ayiyona indlela ende yokusebenzisa ama-multimodal/code; ama-batching, concurrency, kanye nama-latency target ashintsha ukusetshenziswa [1][2].

  • Izimo zezulu ezahlukene kanye namagridi
    Ukupholisa okuhwamukayo endaweni eshisayo, eyomile ≠ ukupholisa komoya/uketshezi endaweni epholile, emanzi. Ukuqina kwamanzi egridi kuyahlukahluka kakhulu [3].

  • Izindlela zomthengisi
    I-Google ishicilele indlela yokukhonza ohlelweni lonke; i-Mistral ishicilele i-LCA esemthethweni. Ezinye zinikeza izilinganiso zamaphuzu ngezindlela ezimbalwa. Isimangalo "esiyingxenye yeshumi nanhlanu yethisipuni" esidumile ngesivinini ngasinye senze izihloko zezindaba - kodwa ngaphandle kwemininingwane yomngcele, asifani [1][3].

  • Ithagethi ehambayo
    Ukupholisa kuvela ngokushesha. I-Microsoft ihlola ukupholisa ngaphandle kwamanzi kumasayithi athile; ukukhipha lokhu kuzonciphisa amanzi esizeni ngisho noma ugesi ovela phezulu usaphethe isignali yamanzi [4].

Ongakwenza namuhla ukunciphisa unyawo lwe-AI lwamanzi 🌱

  1. Imodeli yosayizi wesokudla
    Amamodeli amancane, ahlelwe kahle imisebenzi avame ukufanisa ukunemba ngenkathi eshisa ukubala okuncane. Ukuhlolwa kwe-Mistral kugcizelela ukuhambisana okuqinile kosayizi nesinyathelo - futhi kushicilela izinombolo zokuphetha ezingaphansi ukuze ukwazi ukucabanga ngokuthengiselana [2].

  2. Khetha izindawo ezisebenzisa amanzi
    Khetha izindawo ezinesimo sezulu esipholile, ukupholisa okusebenzayo, kanye namagridi anomfutho ophansi wamanzi nge-kWh; umsebenzi "wokomisa i-AI" ubonisa ukuthi isikhathi nendawo kuyasiza [3].

  3. Shifta imithwalo yomsebenzi ngesikhathi
    Hlela ukuqeqeshwa/inqwaba yenqwaba yamahora akonga amanzi (ubusuku obupholile, izimo zegridi ezivumayo) [3].

  4. Buza umthengisi wakho ngamamethrikhi asobala
    Dinga amanzi ngokushesha ngakunye, izincazelo zemingcele, kanye nokuthi izinombolo zifaka umthamo wokungenzi-lutho kanye nesihloko esiphezulu sesikhungo. Amaqembu enqubomgomo afuna ukudalulwa okuyisibopho ukuze enze ukuqhathanisa kwama-apula nama-apula kwenzeke [3].

  5. Ubuchwepheshe bokupholisa bubalulekile
    Uma usebenzisa ihadiwe, hlola ukupholisa okuvaliwe/okuqondile kuya kuma-chip; uma usefwini, khetha izifunda/abahlinzeki abatshala imali ekwakhiweni kokukhanya kwamanzi [4][5].

  6. Sebenzisa amanzi ampunga bese uphinda usebenzise ongakhetha kukho
    Amakhampasi amaningi angashintsha imithombo engaphuzeki noma agaywe kabusha ngaphakathi kwamaluphu; ama-opharetha amakhulu achaza ukulinganisa imithombo yamanzi kanye nokukhetha kokupholisa ukuze kuncishiswe umthelela wenetha [5].

Isibonelo esisheshayo sokwenza kube ngokoqobo (hhayi umthetho ojwayelekile): ukuhambisa umsebenzi wokuqeqesha ubusuku bonke usuka endaweni eshisayo, eyomile phakathi nehlobo uye endaweni epholile, enomswakama kakhulu entwasahlobo - futhi ukuwusebenzisa ngesikhathi samahora angekho phezulu, apholile - kungashintsha kokubili endaweni kanye angaphandle kwendawo (igridi). Lolo uhlobo lokuhlela ukuwina okusebenzayo, okunedrama encane okungavula [3].

Ithebula lokuqhathanisa: ukukhetha okusheshayo ukuze wehlise inani lamanzi le-AI 🧰

ithuluzi izethameli intengo kungani isebenza
Amamodeli amancane, ahlelwe umsebenzi Amaqembu e-ML, abaholayo emikhiqizweni Okuphansi–okuphakathi Ukubala kancane ngethokheni ngayinye = ukupholisa okuncane + amanzi kagesi; kufakazelwe ekubikeni kwesitayela se-LCA [2].
Ukukhetha isifunda ngamanzi/kWh Abaklami befu, ukuthengwa kwempahla Okuphakathi nendawo Shiftla ezindaweni ezipholile kanye namagridi anamanzi aphansi; bhanqa nomzila oqaphela isidingo [3].
Amafasitela okuqeqeshwa kwesikhathi sosuku Ama-MLOp, abahleli bezinhlelo Phansi Ubusuku obupholile + izimo zegridi ezingcono zinciphisa amandla amanzi asebenzayo [3].
Ukupholisa okuqondile kuya ku-chip/iluphu evaliwe Ama-ops esikhungo sedatha Okuphakathi–okuphezulu Igwema imibhoshongo ehwamukayo lapho kungenzeka khona, inqamule ukusetshenziswa esizeni [4].
Ubude besikhathi kanye nezilawuli ze-batch Abathuthukisi bezinhlelo zokusebenza Phansi Amathokheni we-cap, batch ngobuhlakani, imiphumela yenqolobane; ama-millisecond ambalwa, amamililitha ambalwa [1][2].
Uhlu lokuhlola obala komthengisi CTOs, ukusimama kuholela Mahhala Iphoqelela ukucaca komngcele (kwisayithi uma kuqhathaniswa nangaphandle kwesayithi) kanye nokubika kwama-apula ukuya kuma-apula [3].
I-Greywater noma imithombo etholiwe kabusha Izinsiza, omasipala Okuphakathi nendawo Ukufaka amanzi angaphuzwayo kwehlisa ingcindezi ezintweni eziphuzwayo [5].
Ubambiswano lokusebenzisa kabusha ukushisa Abaqhubi, imikhandlu yendawo Okuphakathi nendawo Ukusebenza kahle kokushisa okungcono kunciphisa ngokungaqondile isidingo sokupholisa futhi kwakha ukuthakasela kwendawo [5].

(“Intengo” incane ngokwedizayini - ukuthunyelwa kuyahlukahluka.)

Ukujula okujulile: i-drumbeat yenqubomgomo iya ngokuya izwakala 🥁

Izinhlangano zobunjiniyela zifuna ukudalulwa okuphoqelekile kwamandla namanzi esikhungo sedatha ukuze abathengi nemiphakathi bakwazi ukwahlulela izindleko nezinzuzo. Izincomo zifaka izincazelo zobubanzi, ukubika kwezinga lesayithi, kanye nesiqondiso sokuthola indawo - ngoba ngaphandle kwezilinganiso ezifanayo, eziqaphela indawo, siphikisana ebumnyameni [3].

Ukujula okujulile: izikhungo zedatha aziphuzi zonke ngendlela efanayo 🚰

Kukhona inganekwane eqhubekayo yokuthi “ukupholisa umoya akusebenzisi manzi.” Akunjalo neze. Izinhlelo ezisebenzisa umoya omningi zivame ukudinga ugesi owengeziwe, othwala amanzi afihliwe kusuka kugridi; ngokuphambene nalokho, ukupholisa amanzi kunganciphisa ugesi kanye nokukhishwa kwegesi ngezindleko zamanzi asendaweni. Abaqhubi abakhulu balinganisa ngokucacile lokhu kushintshana kwezindawo nendawo [1][5].

I-Deep dive: isheke leqiniso elisheshayo mayelana nezimangalo ezibangelwa amagciwane 🧪

Kungenzeka ukuthi uke wabona izinkulumo ezinesibindi zokuthi isiphakamiso esisodwa silingana “nebhodlela lamanzi,” noma, ngakolunye uhlangothi, “amaconsi ambalwa nje.” Ukuma okungcono: ukuthobeka ngezibalo . Ama-bookend anamuhla athembekile angama- ~0.26 mL wesiphakamiso sokukhiqiza esimaphakathi esine-overhead ephelele [1] kanye nama- ~45 mL wempendulo yomsizi wamathokheni angu-400 (umbono ongaphansi) [2]. Isimangalo "esiyingxenye yeshumi nanhlanu yethisipuni" esabiwe kakhulu asinawo umngcele/indlela yomphakathi; siphathe njengesibikezelo sezulu ngaphandle kwedolobha [1][3].

I-Mini-FAQ: Mangakanani amanzi asetshenziswa yi-AI? futhi, ngesiNgisi esilula 🗣️

  • Ngakho-ke, kufanele ngithini emhlanganweni?
    “Ngokushesha, kuyahlukahluka kusukela ekuthoseni kuya ekuphuzeni okumbalwa, kuye ngokuthi imodeli, ubude, nokuthi igijima kuphi. Ukuqeqeshwa kuthatha amachibi, hhayi amadamu.” Bese usho isibonelo esisodwa noma ezimbili ngenhla.

  • Ingabe i-AI imbi ngendlela eyingqayizivele?
    Igxile ngendlela eyingqayizivele :ama-chip anamandla aphezulu ahlanganiswe ndawonye adala imithwalo emikhulu yokupholisa. Kodwa izikhungo zedatha nazo zilapho ubuchwepheshe obusebenza kahle kakhulu buvame ukufika khona kuqala [1][4].

  • Kuthiwani uma sihambisa yonke into ekupholiseni komoya? Ungasika amanzi esizeni kodwa wandise amanzi angekho endaweni ngogesi. Opharetha abayinkimbinkimbi banesisindo sobabili [1][5].

  • Kuthiwani ngobuchwepheshe besikhathi esizayo?
    Imiklamo egwema ukupholisa amanzi esikalini ingashintsha umdlalo ku-Scope 1. Abanye o-opharetha bahamba ngale ndlela; ugesi ovela phezulu usaphethe isignali yamanzi kuze kube yilapho amagridi eshintsha [4].

Amazwi Okugcina - Yinde Kakhulu, Angizange Ngiyifunde 🌊

  • Ngokuya kwakho: cabanga i-sub-milliliter ukuya kumashumi amamililitha, kuye ngemodeli, ubude bokwaziswa, nokuthi isebenza kuphi. Ukwaziswa okumaphakathi ~0.26 mL kusitaki esisodwa esikhulu; ~45 mL ngempendulo yamathokheni angu-400 kwenye [1][2].

  • Ukuqeqeshwa: izigidi zamalitha zamamodeli asemngceleni, okwenza ukuhlela, ukubeka indawo, kanye nokupholisa kube yinto ebalulekile [3].

  • Okufanele ukwenze: amamodeli anosayizi ofanele, khetha izindawo ezisebenzisa amanzi, shintsha imisebenzi enzima uye emahoreni apholile, khetha abathengisi abafakazela imiklamo yokukhanya kwamanzi, futhi bafune imingcele esobala [1][3][4][5].

Isingathekiso esinephutha kancane ekugcineni: I-AI iyi-orchestra eyomile - umculo ungowokubala, kodwa izigubhu ziyaphola futhi zigcwele amanzi. Lungisani ibhendi, izethameli zisawuthola umculo ngaphandle kokuthi ama-sprinkler aqhume. 🎻💦

Izinkomba

  1. I-Google Cloud Blog - Isebenzisa amandla angakanani i-AI ye-Google? Senze izibalo (indlela + ~0.26 mL i-median prompt, i-overhead yokukhonza okugcwele). Isixhumanisi
    (Iphepha lobuchwepheshe le-PDF: Ukulinganisa umthelela wezemvelo wokuletha i-AI esikalini se-Google.) Isixhumanisi

  2. I -Mistral AI - Umnikelo wethu endinganisweni yezemvelo yomhlaba wonke ye-AI (i-LCA ene-ADEME/Carbone 4; ~281,000 m³ ukuqeqeshwa + ukusetshenziswa kwasekuqaleni; ~45 mL ngempendulo yamathokheni angu-400 , ukuphetha okungaphansi).

  3. ULi nabanye - Ukwenza i-AI “Ingabi Nokoma” Kakhulu: Ukwembula Nokubhekana Nendlela Eyimfihlo Yamanzi Yamamodeli E-AI (ukuqeqesha izigidi zamalitha, isikhathi nendawo , ukuhoxisa vs. ukusetshenziswa). Isixhumanisi

  4. I-Microsoft - Izikhungo zedatha zesizukulwane esilandelayo zisebenzisa amanzi ayiziro ukupholisa (imiklamo eqonde ngqo ku-chip eqondise ukupholisa okungenamanzi kumasayithi athile). Isixhumanisi

  5. I-Google Data Centers - Isebenza ngokuzinzile (i-cooling-offs yesayithi ngesayithi; ukubika nokusebenzisa kabusha, okuhlanganisa ukubuyiswa/i-graywater; imiyalo yokusetshenziswa yansuku zonke yezinga lesayithi yobukhulu). Isixhumanisi

Thola i-AI Yakamuva Esitolo Esisemthethweni Somsizi we-AI

Mayelana NATHI

Buyela kubhulogi