Impendulo emfushane: Isikhungo sedatha se-AI siyindawo enabantu abaningi lapho amandla, ukupholisa, indwangu, kanye nendawo yokugcina izinto kugxile khona ekuqeqesheni nasekukhonzeni amamodeli, hhayi igumbi leseva elinama-GPU engeziwe. Ama-chip athola udumo; isakhiwo sinquma ukuthi ayasebenza yini. Uma izithombe zingakwazi ukuphuma, faka kabusha iseli elincane; amaqembu amaningi kufanele aqashe.
Izinto ezibalulekile okufanele uzicabangele:
Ama-chips vs ukwakha: Amandla, ukupholisa, indwangu, kanye nendawo yokugcina izinto kunquma ukuthi ama-accelerator ayasebenza yini.
Izindawo hhayi izigodlo: Faka kabusha ama-rack amabili uma idatha ingakwazi ukuphuma; ungathengi ikhampasi.
Isilinganiso vs i-median: Ekugijimeni okungu-8, i-median hid iyaqala kabusha; sebenzisa isilinganiso samahora angu-18.
Ukumelana nokusetshenziswa kabi: Ungasho i-PUE yama-racks amabili ehholo elixubile.
Imiphumela echazayo: Ukugijima okungu-8 kuyimephu encane, hhayi ukonga umkhiqizo okungu-50%.

Izihloko ongase uthande ukuzifunda ngemva kwalesi:
🔗 Ingabe i-AI ithembekile? Ividiyo nemibuzo
Hlola ukuthembeka kwe-AI ngevidiyo ehehayo kanye nemibuzo esebenzisana nayo.
🔗 Indlela yokusebenzisa i-AI empilweni yansuku zonke
Thola izindlela ezisebenzayo i-AI engenza ngazo imisebenzi yansuku zonke kanye nemikhuba ibe lula.
🔗 Indlela yokusebenzisa i-AI emsebenzini
Funda izindlela ezisebenzayo zokusebenzisa i-AI ukuze uthole umkhiqizo ohlakaniphile endaweni yokusebenza.
🔗 Ingabe i-AI ingazicabangela yona?
Qonda ukuthi ubuhlakani bokwenziwa bungacabanga ngokuzimela noma bucabange.
Indlela ehluke ngayo esikhungweni sedatha "esivamile"
Amahholo endabuko asebenza kahle kakhulu ukuze kube nemisebenzi exubile kanye nesikhathi sokusebenza ezinsizeni eziningi ezincane. Uyakhathalela ukuphindaphinda, yebo, futhi nge-PUE njengomqondo - amandla engeziwe isakhiwo esiwasebenzisayo ukuze sinikeze i-watt yokubala. Ngokuvamile, awuklami yonke imigwaqo ezungeze i-rack esebenza njengepulazi lesifudumezi esiphathwayo.
Amasayithi e-AI aguqula izilinganiso. Ubuningi buyanda. Inethiwekhi iba yindwangu umsebenzi wokuqeqesha ongeke ukwazi ukuhamba ngaphandle kwayo. Indawo yokugcina izinto kufanele igcine izindawo zokuhlola zihamba noma ama-accelerator ahlale engenzi lutho, njengamahhashi omjaho esiminyaminyeni sezimoto.
Kukhona nomehluko wamasiko. Ama-opharetha ebhizinisi acabanga ngamathikithi futhi ashintshe amafasitela. Ama-opharetha e-AI acabanga emigqeni yemisebenzi kanye nomuzwa wokugula lapho i-node ifa sekwephuzile ngokuhamba kwesikhathi. Ngicabanga ukuthi usengabiza zombili ngokuthi "izikhungo zedatha" ngoba ziyizo. Ilebula lifihla nje amapayipi.
| Uhlobo | Yakhelwe ini | Ihadiwe evelele | Uhlamvu lwamandla/lokupholisa | Kufanela bani | Kungani ikhona |
|---|---|---|---|---|---|
| Isikhungo sedatha sebhizinisi lendabuko | I-IT Exubile: izizindalwazi, ama-VM, i-imeyili, amafayela | Ama-CPU, amaseva ajwayelekile, isitoreji esijwayelekile | Iholwa ngomoya; ukuminyana okulinganiselwe; i-PUE njengendawo yokukhuluma | Izinkampani ezisebenzisa izinhlelo zansuku zonke | Gcina izinhlelo zokusebenza zebhizinisi zisebenza |
| Iqembu lokuqeqeshwa kwe-AI | Imisebenzi yokuqeqesha emide, ehlangene kahle | Ama-racks aqinile okusheshisa; Ukuxhumeka kwe-GPU | Ukuminyana okuphezulu; ukupholisa uketshezi noma [ama-heat exchanger angemuva komnyango](https://datacenters.lbl.gov/sites/default/files/rdhx-doe-femp.pdf) | Amalebhu nabakhi bemodeli abahlala emigqeni yemisebenzi | Qedela ukugijima ngaphandle kokulamba ama-chips |
| Isikhungo sokuphetha se-AI | Ukuphakelwa kwemodeli; izimpendulo zesikhathi sangempela kanye ne-batch | Ama-accelerator; ama-load balancers abalulekile | Kusashisa, nje... akuseyona idrama; ukubambezeleka ngaphezu kobuningi obukhulu | Imikhiqizo okumele iphendule manje | Beka imodeli eduze komsebenzisi |
| Ihholo le-AI elihlanganisiwe | Ukuqeqesha nokukhonza ngaphansi kophahla olulodwa; uhlobo oluthile | Ama-rack axubile; amachibi abiyelwe ngocingo; indwangu ehlanganyelwe | Abantu ababili abapholile ekamelweni elilodwa lesitshalo; kuba nzima | Amaqembu angenakukwazi ukukhokhela amakhampasi amabili | Imali inqunyelwe; impilo ayihlelekile |
Akuyona i-ranking yokuziphatha. Imishini ehlukene, isibongo esifanayo.
Ama-GPU, ama-accelerator, kanye ne-rack edla amandla
Hamba phansi okuphakanyisiwe ngokwesiko bese ama-rack ebukeka ehloniphekile. Hamba emgqeni we-AI bese kubukeka sengathi bafuna ukugwinya isakhiwo.
Ihadiwe evelele akuyona i-CPU ehlakaniphile. Yithreyi ye-accelerator: ama-GPU noma amanye ama-chip e-AI, afakwe kumaseva, bese kuba ama-racks, bese kuba imigqa ehlanganyela indwangu ye-bandwidth ephezulu. I-GPU ixhumanisa ama-chip abe yinto engenza sengathi iyi-accelerator eyodwa enkulu; indwangu ye-cluster yenza icebo elifanayo esikalini somugqa. Ukuqeqeshwa okuhambisanayo kusebenza kuphela uma lezo zixhumanisi zihlala ziqinile futhi zibikezelwa. Uma ulahlekelwa yilokho, "i-cluster" yakho iyinqwaba yezindawo zokusebenza ezibizayo ezabelana ngekhodi yeposi.
Amandla alandela ama-chip. Akuyona i-PDU enkulu yehhovisi. Ama-Megawatts omthwalo we-IT uma ihholo seligcwalisiwe - ngeke ngisungule inombolo. Ubuningi be-rack ngayinye yinguquko yendaba. Ama-racks ambalwa. Ngayinye isithando esincane. Isici esivimbelayo ngokuvamile yisiteshi esincane, hhayi uhlu lwezifiso zama-GPU. Uyazi ukuthi kunjani: ukuthengwa kwempahla kufuna ama-accelerator amaningi; umbuso ufuna ingxoxo ende kanye nesheke elikhulu kakhulu.
Isingathekiso esisodwa esilula kancane engingakwazi ukusishukumisa: i-rack iyisilwane esilambile. Ungazala esisheshayo. Kusadingeka uyiphe ukudla, futhi kusadingeka uyikhiphe ukushisa. Yeqa noma yikuphi bese kuba yiphepha elibizayo kakhulu.
Amandla, ukushisa, kanye nenkinga yokupholisa
Ugesi uyangena. Ukushisa kuyaphuma. Yileyo inkolo yonke.
Ama-rack aqinile akhipha ukushisa ngendlela umoya ongakaze ucelwe ukuba uphathe ngayo. Ungawucindezela umoya kakhulu - ama-heat exchanger angemuva, ama-aisles ashisayo, ukuvimba okuhlakaniphile - futhi lokho kusebenza kuze kufike eqophelweni elithile. Bese kuvela uketshezi:
-
Izihibe zokupholisa ezenza igumbi lesitshalo libukeke njengezinto zamakhemikhali
-
Ukucwiliswa emiklamweni embalwa, okusathusa abantu abacabanga ukuthi amanzi namaseva akufanele babelane ngomusho owodwa
Akukho nokukodwa kwalokhu okukhazimulayo. Konke lokhu kungumkhiqizo, ngoba i-accelerator e-throttles yileyo osuvele uyikhokhele futhi ongeke ukwazi ukuyisebenzisa ngokugcwele.
I-PUE isabalulekile. Kuyisilinganiso, hhayi ubuntu. Abaqhubi bayayilandela ngoba yonke i-watt echithwa kumafeni nasemaphampini iyi-watt engayanga ku-GPU. Ngeke ngicaphune isithombe "esijwayelekile"; isimo sezulu nendlela odweba ngayo umngcele iyawuhambisa, futhi ukunemba okungelona iqiniso kubi kakhulu kunokungabikho. Amanzi nawo akhona endabeni. Ezinye izitshalo ziyaphuza. Ezinye ziyagwinya. Ukupholisa okuhwamukayo kusebenza kahle kuze kube yilapho umfula noma isomiso senza kube yipolitiki.
Ukuxhumana: kungani indwangu ibaluleke kakhulu njengama-chips
Abantu bathatha izithombe zama-GPU. Kufanele bathathe izithombe zamaswishi.
Ukuqeqeshwa kuyingxoxo. Izinkulungwane zama-accelerator zishintshana ngama-gradient, amapharamitha, izingcezu zemodeli, ngokuvumelana okuqinile. Uma indwangu inyakaza, wonke umsebenzi ulinda kancane kakhulu. Yingakho ama-AI halls egxile kakhulu kunethiwekhi ye-bandwidth ephezulu, ukubambezeleka okuphansi, kanye ne-topology engasongeki phakathi lapho isixhumanisi sehluleka. Izindwangu ze-InfiniBand-ish, i-Ethernet endimeni efanayo, i-GPU ixhuma ngaphakathi kwebhokisi, izintambo okumele zibe zilungile okokuqala.
Ukucabanga kuyingxoxo ehlukile. Ukukhonza imodeli kukhathalela ukubambezeleka komsila - impendulo ehamba kancane, hhayi evamile. Iqembu vs isikhathi sangempela lihlukanisa ubuntu. Iqembu lokuqeqesha lifuna ukunyakaza okukhulu, okuhlangene, okuphelele. Iqembu lokukhonza lifuna izicelo eziningi ezincane kanye nokuhlukaniswa, ngaphandle kokuminyana kwezimoto endaweni yokulinganisela umthwalo.
I-Miss, njengoba ngilapha: abantu baphatha inethiwekhi njengamapayipi kanye nama-chips njengendawo yokudlela. Kulesi sakhiwo amapayipi yindawo yokudlela. Uma uphuthelwa yilokho, uthenga ikhishi elingenakukwazi ukulethwa.
Ukuqeqeshwa vs ukuqagela: izakhiwo ezimbili, ngezinye izikhathi ngokoqobo
Ukuqeqeshwa kuwumkhankaso. Uhlanganisa iqembu, ulinike idatha, uhlole ngokwenkolo, ugijime amahora noma amasonto, bese uthandaza ukuthi indwangu ihlale ingenazo izinkinga. Igcwele inqwaba, ilambele i-bandwidth, inesineke ngokuthula - kuze kube yilapho i-node ehlulekile iqeda ukugijima. I-accelerator eyodwa efile emsebenzini ohlangene kahle ingavimba yonke i-chorus.
Ukucabanga kuyindlela yokuthenga. Amamodeli asevele eqeqeshwe, manje aphendula imibuzo, ahlela izithombe, akhiqiza umbhalo. Ukubambezeleka kubalulekile. I-accelerator endala kancane eduze kwekhasimende inganqoba enye enhle kakhulu ekude nezwekazi.
Ngakho uthola ukuhlukana. Amakhampasi okuqeqesha axosha amandla, umhlaba, kanye nobuningi. Izindawo zokuqagela zixosha ukubambezeleka nokuba khona. Amahholo ahlanganisiwe nawo akhona, ngoba imali ayinamkhawulo. Hhayi-ke - akuzona njalo izakhiwo ezimbili. Ngezinye izikhathi ihholo elilodwa elinentambo ye-velvet kanye nezihibe ezimbili zokupholisa.
Yilapho futhi kulapho kutholakala khona indawo yokuhlala, amakhampasi e-hyperscale, kanye nefoloko elise-prem. Ama-hyperscalers akha ngezikali ezenza sonke sibukeke sengathi sihlela ifenisha. Ama-Colo athengisa ubuningi eceleni kwe-rack. I-On-prem isaqhubeka lapho idatha ingakwazi ukuphuma.
Ukufinyelela kwesitoreji, izindawo zokuhlola, kanye nama-chips alambile
Akekho obeka uhlelo lwamafayela oluhambisanayo esembozweni sencwajana.
Idatha yokuqeqesha kufanele ifike ngokushesha ngokwanele ukuze ama-GPU angashayi izinyawo zawo. Izindawo zokuhlola kufanele zifike ukuze ingozi ingalahli isonto lonke. Izisindo zamamodeli kufanele zilayishwe ngaphambi kokuba ikhophi ekhonzayo ivele bukhoma. Umthamo wokugcina - hhayi nje umthamo - uyinkinga ethule. Ungachitha imali eningi kuma-accelerator bese uwabulala nge-array ehloniphekile eyenzelwe imishini ebonakalayo.
Iphethini ijwayelekile: iqembu elicwebezelayo, ulayini womsebenzi, kanye ne-I/O elinde emuva njenge-invoyisi engcolile. Ukuhlanganisa ama-computing ngaphandle kokuhlanganisa indlela yedatha yindlela othola ngayo i-idle ebiza kakhulu. Gcina ama-chips ephakelwa noma uvume ukuthi uthenge isithombe esiqoshiwe.
Isofthiwe, ukuhlelwa kabusha, kanye nesendlalelo se-ops esingathandeki
Ihadiwe iyona esaziwayo. Abahleli bezinhlelo benza umsebenzi.
Isikhungo sedatha se-AI asisebenzi uma imisebenzi ingakwazi ukuthola ama-GPU, uma amaqembu amabili engakwazi ukwabelana ngeqoqo ngaphandle kokulwa ngenqindi, uma izinga elihlulekile lingenakushintshwa, uma i-firmware inyakaza kuze kube yilapho indwangu yehluleka ngokuhamba kancane. Ukuhlelwa, ukubonwa, ukuvala amandla - ungqimba lwe-ops olungathandeki, okuyiyona ndlela owazi ngayo ukuthi lubalulekile.
Ngithanda kakhulu lolu ngqimba. Futhi lapho i-NIC engalungiselelwanga kahle izenza inkinga yokupholisa intambama yonke... uthola lokho ngendlela enzima.
Uma i-node ifa phakathi nokugijima
I-node iyafa. Amafasitela okushintsha asashayisana nokuqeqeshwa okungakhathali ngekhalenda lakho. Nihlela imisebenzi emikhulu ngamaqembu, nivala amachibi okuhlola, nibhala izincwadi zokwehluleka ezibukeka sengathi "umsebenzi uhamba kancane" kuze kube yilapho kubukeka sengathi "umsebenzi ufile." Ukuphindaphinda kusabalulekile - amandla, ukupholisa, izindlela, indawo yokugcina - kodwa imodi yokwehluleka ayihlelekile kahle kuneslayidi esidala sesikhathi sokusebenza. Ukuqonda: ukukopisha, ukukhipha i-node yokugula, qhubeka uphendula. Ukuqeqeshwa kufuna izindawo zokuhlola, hhayi ithemba.
Indawo, amanzi, igridi, kanye nomakhelwane
Awuyilahli enye yalezi eduze kwendlu encane ukuze ubone umbono.
Ukuxhumeka kwegridi kuvame ukuba inqubo yangempela yokukhetha indawo. Umhlaba ulula uma uqhathaniswa nesiteshi esincane kanye nohlelo lokusebenza olunamakhasimende amanye. Amanzi okupholisa, uma uwasebenzisa, aba yinkinga yomakhelwane ngokushesha nje lapho imvula ingabi mnandi. Umsindo. Ubuningi obubonakalayo. Ukushisa ocingweni lomngcele. Amakomidi okuhlela athola imibono mayelana "nefu" ngesikhathi lapho edinga khona insimu nomfula.
Ukubambezeleka kudonsela eceleni. Ukucabanga kuthanda ukuba seduze kwabasebenzisi kanye nezixhumanisi. Ukuqeqeshwa kungacasha ezimakethe zamandla ezishibhile kanye nezimo ezipholile. Imboni ikhuluma sengathi kukhona indawo eyodwa ephelele. Ayikho. Kukhona isivumelwano nesitatimende sabezindaba.
Ukushisa okusele kanye nesithando somlilo esingamenywanga
Izincwajana ziyayithanda le nto. Faka ukufudumala kwemfucuza emakhaya, emachibini, nasezindlini zokugcina izithombo; umusho omuhle. Ngezinye izikhathi umjikelezo wesifunda ungokoqobo. Ngezinye izikhathi ikhampasi isendaweni engafanele futhi ukushisa kusaya emoyeni. Ngingabaza inguqulo yencwajana; angingabazi ngefiziksi.
Ubani oyidingayo (futhi ubani okufanele ayiqashe esikhundleni sayo)
Iningi labantu akudingeki libe nelinye lala mahholo.
Uhlu olucacile:
-
Ama-Hyperscalers, ngoba umkhiqizo ungowemikhumbi
-
Amalebhu, lapho isikhathi somugqa siyinkinga, noma idatha ingakwazi ukuphuma
-
Ibhange, iqembu lesibhedlela, uhulumeni, noma umenzi onesethi yedatha eyimfihlo - i-on-prem noma ikheji le-colo eliyimfihlo kungaba nengqondo, noma ngabe kuyinto engenangqondo
Bonke abanye kufanele baqashe. Indawo yokuhlala enobuningi obulungele i-AI. Ukubhukha ngamafu. Iqembu eliphethwe. Uthola ama-accelerator ngaphandle kokuba ngumsebenzisi wesitshalo sikagesi. Uthando luyaphela okokuqala lapho othile ebuza ukuthi ubani ocela i-coolant loop ngo-3 ekuseni.
Ngizovuma ukuthi kukhona into yokuziqhenya. Ukuba neqembu kuzwakala sengathi ukuba nezindlela zokubikezela. Bese kufika i-invoyisi kagesi, bese i-pride ihlala phansi.
Isakhiwo sihloselwe ini?
Ngakho-ke, Iyini i-AI Data Center? Ikhampasi ekhethekile, enabantu abaningi lapho izisheshisi, ugesi, ukupholisa, indwangu, kanye nendawo yokugcina izinto zihlelwe khona ngokuhambisana nokuqeqeshwa kanye nokuphakelwa kwamamodeli - hhayi i-IT ejwayelekile ene-GPU ekhoneni. Ibukeka njengendlu yokugcina impahla. Iziphatha njengesiteshi sikagesi esenza izibalo.
Uma ungakhumbuli lutho olunye: ama-chip athola udumo; isiteshi esincane, umshini wokupholisa, kanye nenethiwekhi banquma ukuthi lawo ma-chip ayengumqondo omuhle yini. Ukuqeqeshwa kanye nokucabanga kungabelana ngophahla olulodwa; basafuna izindlela ezahlukene zokuziphatha. Izinhlangano eziningi kufanele ziqashe. Ezimbalwa kufanele zakhe. Omakhelwane bazobona noma yikuphi.
Ifu lalilokhu linesakhiwo. Kulezi zinsuku isakhiwo sinemibono.
Isibonelo sangempela: Iseli lokuqeqesha elinama-rack amabili lapho izithombe zingakwazi ukuphuma
Isimo
UTomos ungumholi wengqalasizinda eKestrel Precision, umkhiqizi wabantu abangu-400 eWest Midlands. Bavele banehholo elincane elise-prem: i-ERP, ukwabelana ngamafayela, imishini ebonakalayo, indawo exubile lesi sihloko esiqale ngayo. Ukupholisa umoya. I-Ethernet evamile. I-SAN ehloniphekile ngokuphelele kumadiski ehhovisi.
Ithimba elibona ngomshini lidinga ukuqeqesha imodeli yokuhlola kuma-still asefektri. Ama-still awakwazi ukuphuma endaweni. Abonisa inqubo inkampani engeke iyibeke kudiski yamafu, ngisho neyangasese. Isixazululo sokuthengwa kwamasevisi ama-server amane ane-dual-accelerator kanye nesilayidi esinesihloko esithi "isikhungo sethu sedatha se-AI". Ama-server angena kuma-rack amabili akhona, ngoba kukhona isikhala, futhi isikhala sizwakala njengesivimbelo.
Akunjalo. Phakathi namasonto amabili ama-GPU azwakala ematasa bese ephenduka kancane kancane. Imisebenzi iyakhasa lapho indawo yokuhlola ifika ku-SAN. I-node iyafa ngehora leshumi nanye futhi ukugijima kuphelile. UTomos akahlulekanga ukuthenga ama-chips. Uthenge inqwaba yezindawo zokusebenza ezibizayo ezihlanganyela ikhodi yeposi. Isakhiwo sasiseyihholo lebhizinisi.
Abadingi ikhampasi, isiteshi esisha, noma umfula. Badinga iseli elincane elisebenza njengesikhungo sedatha se-AI esincane: amandla ama-rack angawabamba ngempela, ukushisa okuphumayo ngaphandle kokupheka ama-chips, indwangu umsebenzi wokuqeqesha ongayikhuluma ngayo, indawo yokugcina engathatha indawo yokuhlola, kanye nencwadi yokugijima lapho i-node ifa. Ngenxa yokuthi izithombe azikwazi ukuhamba, ukuqasha ikheji le-colo imizuzu engamashumi amane ukusuka lapha akuyona impendulo. Ukulungisa ama-rack amabili kuyisisombululo.
Lokho okudingwa yiseli
-
Isabelomali samandla esilinganisiwe kulowo mugqa, kusukela kuma-PDU, hhayi ohlwini lwezifiso ze-GPU. Uma amandla asele engakwazi ukondla amaseva amane ngesikhathi sokuqeqeshwa, ingxoxo iyaphela lapho bese bebheka umbala oqinile, hhayi isimangaliso
-
Ukupholisa umoya akuzange kucelwe ukuthi kuphathwe ngale ndlela: ama-heat exchanger angemuva uma ihholo lingakwazi ukuwathatha, noma i-liquid loop encane uma ingakwazi ukuwathatha esikhundleni salokho. Uma kungekho noyedwa, amaseva awangeni
-
Indwangu ezinikele ye-bandwidth ephezulu phakathi kwama-node amane, hhayi i-Ethernet yehhovisi. Uma umthengisi eneswishi engu-10 Gb kuphela kukhathalogi, lokho akulona iqembu
-
Indawo yokugcina izinto ngokushesha yendawo yezindawo zokuhlola kanye nezindawo zokuqeqesha, hhayi i-VM SAN
-
Umhleli wesikhathi, isikhawu sokuhlola, kanye nendlela yokuqalisa kabusha ebhaliwe. Ihadiwe iyona edumile. Lo msebenzi ungowesigaba
-
Ibhokisi lokucabanga elibiyelwe lomugqa wefektri, elihlukile engxoxweni yokuqeqesha, ngoba ukukhonza kufuna ukubambezeleka komsila kanye nokuhlukaniswa, hhayi ukuhlangana
-
Imvume yokufaka amandla e-log, amawashi e-GPU, imicimbi ye-throttle, kanye newashi lodonga lomsebenzi. Uma bengakwazi ukuwahlola, bazothatha izithombe zama-GPU bese bephuthelwa amaswishi
Isibonelo semiyalelo
UTomos ubeka lokhu ngesifinyezo sezikhungo, ngolimi olujwayelekile:
Ungakubizi lokhu ngokuthi ikhampasi ye-AI. Yakha iseli lokuqeqesha elinama-rack amabili ehholo elikhona lamaseva amane ane-dual-accelerator. Izithombe zasefektri zihlala endaweni. Impumelelo iwukuthi: ama-node amane aqedela iresiphi yokuqeqeshwa kokuhlola kwezikhathi eziyi-12 ngaphandle kwe-thermal throttle, abhale indawo yokuhlola ngemizuzu hhayi amashumi emizuzu, bese eqala kabusha kusukela kuleyo ndawo yokuhlola lapho sibulala i-rank ngamabomu. Ukupholisa kumele kugcine ama-GPU emawashini awo okuqeqesha. Ukuxhumana kumele kube yindwangu ehlanganyelwa yilawo mabhokisi amane, hhayi indlela edlula ehhovisi. Indawo yokugcina kumele yondle ama-chips. Uma ama-heat exchanger angemuva engangeni, yisho bese uyeka. Ungacaphuni i-PUE yama-rack amabili ehholo elixubile. Lelo nani lizoba yi-theatre.
Ube esebeka lokhu emsebenzini wokuqeqesha uqobo:
Indawo yokuhlola njalo ngemizuzu engama-30 eya echibini elisheshayo lendawo. Uma irenki ifa, qala kabusha kusukela endaweni yokugcina yokuhlola ephelele. Ungalindi ku-SAN. Ungaqhubeki nokuziqeqesha ngenkathi amawashi ehlile ekushiseni. Yibhale phansi bese uyama ukuze sibone indawo yokugcina.
Ihora elihle libukeka kanje: wonke ama-GPU ayisishiyagalombili ewashi lokuqeqesha, ifayela lokuhlola lifikile, umsebenzi usavaliwe. Ihora elibi libukeka kanje: abalandeli bememeza kakhulu, amawashi ehla, indawo yokuhlola ibhalwe ngo-2% ngemva kwemizuzu eyishumi, kanye nomuntu ehhovisi ethi "iqembu liphezulu". Phezulu akusikho ukuqeqeshwa.
Indlela yokuyihlola
Babhala isivivinyo ngaphambi kokulungiswa kabusha, okuyiwona mgomo wokungathembi idemo.
-
Iresiphi efanayo yezikhathi ezingu-12, izithombe ezifanayo zokuhlola ezingu-120,000, ama-run ayisishiyagalombili
-
Isikhathi sihambisana newashi lodonga leresiphi eqediwe, kufaka phakathi noma yikuphi ukuqala kabusha, okulinganiswe kusukela ekuthunyelweni komsebenzi kuya endaweni yokugcina yokuhlola yesikhathi se-12
-
Ukudlulisa ukushisa: Amawashi e-GPU ahlala endaweni yokuqeqeshwa ukuze agijime. Umcimbi we-throttle awuphumeleli ngisho noma umsebenzi uphela ekugcineni
-
Isitoreji esidlulisiwe: isikhathi sokubhala sendawo yokuhlola, esimaphakathi nesibi kakhulu, kusukela kulogi yomsebenzi
-
Indwangu yokudlulisa: umsebenzi awuhlali endaweni yokulinda ndawonye ngenkathi isikhundla siphilile. Uma bengakwazi ukubona lokho kulinda, izinsimbi aziqedwa
-
Ama-run amabili kwangu-8 abulawa ngesinyathelo esinqunyiwe, ngenhloso, ukuze kuhlolwe indlela yokuqalisa kabusha
-
Ukuphetha kuwukuhlolwa okuhlukile kwezithombe ezingu-200 kusukela emugqeni wefektri kuya ebhokisini lokuphakela elibiyelwe. Impumelelo yokuqeqesha ayibalwa njengempumelelo yokuphakela
-
Baqopha amandla e-rack avela kuma-PDU ngamasampula emizuzu eyi-15 ngakho-ke akekho okufanele asungule i-megawatt
Ukwamukelwa kokubiza iseli ngokuthi "lilungele i-AI": izindlela zokupheka ezingu-8 kwezingu-8 ziphelile; u-0 kwabangu-8 ubonisa i-thermal throttle; kokubili ukugijima okubulewe kuyaqala; izindawo zokuhlola zihlala ngemizuzu embalwa. Uma bephuthelwa yilokho, basenegumbi leseva elinamakhadi ehluzo amahle.
Umphumela
Umphumela obonisayo, ovela ekuhlolweni okwenziwe ngama-run ayisishiyagalombili, hhayi isibalo se-Kestrel esishicilelwe.
Okucatshangwayo: amaseva amane ane-dual-accelerator kuma-rack amabili; imodeli eyodwa yokuhlola; izithombe ezingu-120,000; ukugijima okungu-8 kweresiphi ye-12-epoch ehleliwe; iwashi lodonga lihlanganisa ukuqala kabusha kuze kube yilapho iresiphi iphelile; i-throttle isho ukulahla okugciniwe ewashini lokuqeqesha; isikhathi sokuhlola siwukubhala, hhayi isifiso; amandla e-rack amasampula e-PDU, hhayi i-PUE yekhampasi.
Ngaphambi kokulungiswa kabusha, ama-GPU kuma-rack ajwayelekile apholile emoyeni ku-Ethernet yehhovisi kanye ne-VM SAN:
-
Ama-run ama-5 kwangu-8 aqediwe ngomzamo wokuqala. Iwashi lodonga eliphakathi nendawo phakathi kwalawo mahlanu: amahora ayi-14
-
Abathathu kwabayisishiyagalombili kwadingeka baqale futhi ngemva kokumiswa cishe ngehora le-11 (ababili ngemva kokuba i-node ife nendawo yokuhlolwa endala, eyodwa ngemva kwendawo yokuhlolwa yagcwalisa i-SAN). Zama futhi ngokushesha, akukho ukulinda ubusuku bonke ewashini: amahora ayi-11 achithwa kanye nomzamo wesibili wamahora ayi-14, noma amahora angama-25 ukuze uthole iresiphi ephelele kulawo mathathu
-
Isikhathi esimaphakathi kuya esiqediwe seresiphi kuzo zonke eziyisishiyagalombili, ukuqalisa kabusha kufakiwe: amahora ayi-18. (Amahlanu ku-14, amathathu ku-25. Isilinganiso sazo zonke eziyisishiyagalombili siseyi-14, okungafihla ukuqalisa kabusha. Yingakho isilinganiso siyisisekelo lapha.)
-
I-throttle yokushisa ifinyelele ku-7 kwangu-8. Isikhathi esimaphakathi ngewashi elincishisiwe: amahora ama-3 emzamweni wamahora ayi-14
-
Ukubhala indawo yokuhlola: isilinganiso semizuzu engama-22
-
Ukubulala ngengozi kwakungeyona indlela ababengaphumelela ngayo. Babengenayo incwadi yokubhalisa. Ukufa okubili ngengozi kwaba yikona okwatholakala
-
Umthwalo we-IT ophezulu kuma-rack amabili ngesikhathi sokuqeqeshwa: cishe u-18 kW. Umugqa wawunama-watts. Wawungenawo ukupholisa noma indwangu
Ngemva kokushintshana kokushisa okusemnyango wangemuva kulawo ma-rack amabili, indwangu ezinikele phakathi kwama-node amane, ichibi elincane le-NVMe lezindawo zokuhlola, indawo yokuhlola yemizuzu engama-30, kanye ne-runbook yokuqalisa kabusha:
-
8 kwangu-8 aqede ngomzamo wokuqala. Iwashi eliphakathi nendawo lodonga: amahora angu-9
-
I-throttle yokushisa: 0 kwangu-8
-
Ukubhala kwe-Checkpoint: okumaphakathi imizuzwana engama-90. Okubi kakhulu kusethi: imizuzu emi-3
-
Ukubulalana okubili ngamabomu kokubili kuqale kabusha kusukela endaweni yokuhlola yokugcina yemizuzu engama-30. Iwashi elengeziwe lodonga kulawo ma-run amabili: cishe imizuzu engama-40 lilinye, kufakwe nokuxilongwa, ngakho labo ababili bahlala cishe amahora ayi-9 nemizuzu engama-40. Isilinganiso sisukela kuma-round ayisishiyagalombili kuya kumahora ayi-9
-
Umthwalo we-IT ophakeme usacishe ube ngu-18 kW. Ama-chips afanayo. Ukuziphatha okuhlukile kwesakhiwo
Kuyo yonke le sampula, iresiphi ephakathi nesikhathi esiqediwe yehle kusukela emahoreni ayi-18 kuya emahoreni ayi-9, noma amahora ayi-9 ngalinye, amahora angama-72 phakathi kokugijima okuyisishiyagalombili. Oqede okokuqala kusuke ku-5 kwangu-8 kuya ku-8 kwangu-8. I-Throttle isuke ku-7 kwangu-8 yaya ku-0 kwangu-8. Leyo nombolo yokugcina yileyo esho ukuthi bathenge yini ama-accelerator noma ama-paperweights. Ngeke babize ukushintsha kwesikhathi njengokonga okungu-50% ekukhiqizeni. Ukugijima okuyisishiyagalombili kuyisethi encane, elula.
Lezi zibalo ziyisibonelo sokulinganisa esisekelwe esivivinyweni esishiwo, isampula encane, imodeli eyodwa, kanye nehholo elilodwa elixubile. Akuyona i-PUE, akuyona i-megawatt yekhampasi, futhi akusibo ubufakazi bokuthi i-Kestrel kufanele yakhe indawo yokuqeqesha ensimini. Ukubuyekezwa kwezingodo kuhlale ngaphakathi kwamahora angu-9; abazange bakufihle. Isibalo esingu-18 kW siwukufundwa kwe-PDU okuzungezayo, akuyona i-invoyisi yesevisi. Abazange baguqule amahora angu-72 abe yindleko, ngoba izinga likagesi elixubile lefektri lalizobangela icala lebhizinisi elingamanga.
Isheke lokuphakelwa lalihlukile futhi lincane: izithombe ezingu-200 eziya ebhokisini elibiyelwe, zonke zisendaweni. Lokho kungukucabanga, hhayi ukuqeqeshwa. Ukuxuba lokhu okubili bekuyoba yiphutha le-hybrid-hall nge-miniature.
Yini engase ihambe kabi
-
Ukuthengwa kwempahla kuqhubeka nokubiza amaseva amane ngokuthi “isikhungo sedatha se-AI”. Ilebula lifihla amapayipi, futhi ukuthengwa okulandelayo ama-GPU amaningi e-aisle efanayo yokugula
-
Ama-exchange omnyango wangemuva ayangena futhi akekho ovula amanzi. Abalandeli basamemeza. Amawashi asawa
-
Indwangu iyiswishi eyodwa engenandlela eyisipele. Isixhumanisi esisodwa esihlulekile futhi "iqembu" liphinde libe yizindawo zokusebenza ezine
-
Izindawo zokuhlola zihlala e-SAN ngoba ichibi le-NVMe "laliyisigaba sesibili". Isigaba sesibili asifiki ngaphambi kwesigaba esilandelayo esifile
-
Bacaphuna i-PUE yama-rack amabili ehholo elixubile. Isimo sezulu, umngcele, kanye nomugqa we-ERP wonke kwenza leso silinganiso sibe yisambatho
-
Ukuqeqeshwa nokuphakelwa kwabelana ngendwangu. Ukukhukhuleka kwendawo yokuhlola kwenza umugqa wefektri ulinde. Ukubambezeleka komsila kungumkhiqizo lapho, hhayi ukuminyana
-
I-node iyafa bese othile eyiphatha njengethikithi lesikhathi sokusebenza esikhundleni sokuqala kabusha indawo yokuhlola. Ama-op ebhizinisi kanye nama-op e-AI bayaxoxa bedlulana ntambama yonke
-
Babengaqasha ikheji, ngaphandle kokuthi izithombe azikwazi ukuphuma. Ukukhohlwa ukuthi umkhawulo ubathumela engxoxweni yamafu okuzodingeka baphumule kuwo
Ukudla okuwusizo
Kuyini i-AI Data Center, kulesi simo, akuyona indawo yokugcina impahla futhi akuyona i-invoyisi ye-GPU. Yiseli elivumela ama-accelerator ukuthi aqedele umsebenzi: amandla angawabamba, ukushisa okuphumayo, indwangu, indawo yokuhlola, kanye nomuntu obambekayo lapho isikhundla sifa. I-Kestrel yayingadingi ikhampasi. Babedinga ama-rack amabili ukuze bayeke ukuzenzisa. Amaqembu amaningi kufanele aqashe lokho kuziphatha. Abambalwa, abanesethi yedatha eyimfihlo kanye nehholo elingathwala ukushisa, kufanele bakhe iseli bese benqaba isilayidi.
Imibuzo Evame Ukubuzwa
Iyini i-AI Data Center?
Indawo yokubala enabantu abaningi lapho amandla, ukupholisa, ukuxhumana, kanye nesitoreji kugobile khona ekuqeqeshweni okuhambisanayo kanye nokukhonza imodeli, hhayi imigqa ehlelekile yamaseva enhloso ejwayelekile. Yakhelwe ukuze inani elikhulu lama-accelerator likwazi ukuqeqesha nokusebenzela amamodeli ngaphandle kokuncibilika, ukuma kunethiwekhi, noma ukulinda kudiski. Ihholo lebhizinisi elijwayelekile liyindawo exubile. Ihholo le-AI liyi-monoculture enenani layo eliyi-accelerator, njenge-GPU noma ama-chip esitayela se-TPU. Libukeka njengendlu yokugcina impahla futhi liziphatha njengesiteshi sikagesi esenza izibalo.
Isikhungo sedatha se-AI sihluke kanjani esikhungweni sedatha esivamile?
Amahholo endabuko asebenza kahle kakhulu ukuze kube nemithwalo yemisebenzi exubile kanye nesikhathi sokusebenza ezinsizeni eziningi ezincane. Amasayithi e-AI aguqula izilinganiso: ubuningi buyakhuphuka, inethiwekhi iba yindwangu umsebenzi wokuqeqesha ongeke ukwazi ukuyikhipha, futhi indawo yokugcina kufanele igcine izindawo zokuhlola zihamba noma ama-accelerator ahlale engenzi lutho. Ama-opharetha ebhizinisi acabanga ngamathikithi futhi ashintshe amafasitela. Ama-opharetha e-AI acabanga emigqeni yemisebenzi kanye nomuzwa wokugula lapho i-node ifa sekwephuzile ngokuhamba kwesikhathi. Usengashayela zombili izikhungo zedatha. Ilebula ifihla nje amapayipi.
Kungani izikhungo zedatha ze-AI zisebenzisa amandla amaningi kangaka?
Ihadiwe evelele akuyona i-CPU ehlakaniphile. Yithreyi yokusheshisa: ama-GPU noma amanye ama-AI chips, afakwe kumaseva, bese kuba ama-racks, bese kuba imigqa ehlanganyela indwangu ye-bandwidth ephezulu. Ubuningi nge-rack yinguquko yendaba: ama-racks ambalwa, ngalinye libe isithando esincane. Ama-Megawatts omthwalo we-IT avela uma ihholo seligcwalisiwe, yize ukusungula isithombe esijwayelekile kungafanele. Isici esikhawulelayo ngokuvamile yisiteshi esincane, hhayi uhlu lwezifiso zama-GPU.
Izikhungo zedatha ze-AI zipholiswa kanjani?
Ama-rack aphezulu alahla ukushisa ngendlela umoya ongakaze ucelwe ukuba uphathe ngayo. Ungacindezela umoya kakhulu ngezishintshi zokushisa ezingemuva, izindawo ezishisayo, kanye nokugcina okuhlakaniphile. Bese kuvela uketshezi: ukupholisa okuqondile kuya kuma-chip, izihibe zokupholisa, kanye nokucwiliswa emiklamo embalwa. I-accelerator ephothayo yileyo osuvele uyikhokhele futhi ongeke ukwazi ukuyisebenzisa ngokugcwele. I-PUE isabalulekile ngoba i-watt ngayinye echithwe kumafeni nasemaphampini iyi-watt engayanga ku-GPU. Amanzi nawo akhona endabeni: ezinye izitshalo ziyaphuza, ezinye ziyagwinya.
Kungani ukuxhumana kwenethiwekhi kubaluleke kakhulu njengama-GPU?
Ukuqeqeshwa kuyingxoxo: izinkulungwane zama-accelerator zishintshana ngama-gradient, amapharamitha, kanye nezingcezu zemodeli ngokuvumelanisa okuqinile. Uma indwangu inyakaza, wonke umsebenzi ulinda kancane kakhulu. Yingakho ama-AI halls egxile kakhulu kunethiwekhi ye-bandwidth ephezulu, ukubambezeleka okuphansi, izindwangu ze-InfiniBand-ish noma i-Ethernet endimeni efanayo, kanye ne-topology engasongeki phakathi lapho isixhumanisi sehluleka. I-Inference ikhathalela ukubambezeleka komsila, izicelo eziningi ezincane, kanye nokuhlukaniswa. Kulesi sakhiwo amapayipi yindawo yokudlela.
I-AI Data Center isetshenziselwa ini: ukuqeqeshwa noma ukuphetha?
Ukuqeqeshwa kuwumkhankaso: ukuhlanganisa iqembu, ukulipha idatha, ukuhlola ngokwenkolo, bese kuqhubeka amahora noma amasonto. Ukucabanga kuyindlela yokuthenga: amamodeli asevele eqeqeshwe, manje aphendula, ngokubambezeleka okubalulekile. Amakhampasi okuqeqesha axosha amandla, umhlaba, kanye nobuningi. Izindawo zokucabanga zixosha ukubambezeleka nokuba khona. Amahholo ahlanganisiwe akhona ngoba imali ayipheli, ngezinye izikhathi ihholo elilodwa elinama-loops amabili okupholisa. I-accelerator endala kancane eduze kwekhasimende inganqoba elikhazimulayo elikude nezwekazi.
Kungani isitoreji sibalulekile esikhungweni sedatha se-AI?
Idatha yokuqeqesha kufanele ifike ngokushesha ngokwanele ukuze ama-GPU angashayi izinyawo zawo. Izindawo zokuhlola kufanele zifike ukuze ingozi ingalahli isonto lonke. Izisindo zamamodeli kufanele zilayishwe ngaphambi kokuba ikhophi ekhonzayo ivele bukhoma. Umthamo wokugcina, hhayi nje umthamo, uyinkinga ethule. Ungachitha imali eningi kuma-accelerator bese uwabulala nge-array ehloniphekile eyenzelwe imishini ebonakalayo.
Kwenzekani uma i-node ifa phakathi nesikhathi sokugijima?
I-accelerator eyodwa efile emsebenzini wokuqeqesha ohlangene kahle ingavimba yonke i-chorus. Ukuqeqeshwa kufuna izindawo zokuhlola, hhayi ithemba. Ukucabanga kuqeda i-node yokugula, kugcina impendulo ephindaphindayo, futhi kuhlala kubukhoma. Amafasitela okushintsha asashayisana nokugijima kokuqeqeshwa okungakhathaleli ikhalenda lakho. Ukuphindaphinda kusabalulekile ngamandla, ukupholisa, izindlela, kanye nokugcina, kodwa imodi yokwehluleka ayihlelekile kahle kune-slide yakudala ye-nines-of-uptime.
Iyini ithonya le-AI Data Center kugridi, amanzi, kanye nomakhelwane?
Ukuxhumeka kwegridi kuvame ukuba inqubo yangempela yokukhetha indawo. Umhlaba ulula uma uqhathaniswa nesiteshi esincane kanye nohlelo lokusebenza olunamakhasimende amanye. Amanzi okupholisa, uma uwasebenzisa, aba yinkinga yomakhelwane ngokushesha nje lapho imvula ingabi mnandi, kanye nomsindo, ubukhulu bokubona, kanye nokushisa ocingweni lomngcele. Ukuqeqeshwa kungacasha ezimakethe zikagesi ezishibhile kanye nezimo zezulu ezipholile. Ukucabanga kuthanda ukuba seduze nabasebenzisi. Ayikho indawo ephelele. Kukhona isivumelwano nesitatimende sabezindaba.
Ingabe ngidinga ukuba nesikhungo sedatha se-AI, noma kufanele ngiqashe?
Abantu abaningi akudingeki babe nelinye lala mahholo. Ama-Hyperscaler akha ngoba umkhiqizo ungowezimoto eziningi. Ama-lab akha uma isikhathi somugqa sinzima noma idatha ingakwazi ukuphuma. Ibhange, isibhedlela, uhulumeni, noma umenzi onesethi yedatha eyimfihlo angenza kube nengqondo ukuthi i-on-prem noma i-colo cage yangasese. Bonke abanye kufanele baqashe: i-AI-ready colocation, ukubhukha ifu, noma iqembu eliphethwe. Uthola ama-accelerator ngaphandle kokuba ngumsebenzisi wesitshalo sikagesi.
Izinkomba
-
I-IEA - www.iea.org
-
I-LBNL - izikhungo zedatha.lbl.gov
-
Igridi Eluhlaza - www.thegreengrid.org
-
I-LBNL - izikhungo zedatha.lbl.gov
-
I-LBNL - izikhungo zedatha.lbl.gov
-
Isikhungo se-Uptime - journal.uptimeinstitute.com
-
I-NVIDIA - developer.nvidia.com
-
I-NVIDIA - docs.nvidia.com
-
I-NVIDIA - docs.nvidia.com
-
I-NVIDIA - docs.nvidia.com
-
I-Google Cloud - docs.cloud.google.com