Ko All AI Computing Centers Adopt Yakavharwa-Loop Liquid Cooling Systems?

Nekuwedzera kuri kuita kukurumidza kudzidziswa kwemhando yemutauro, kugadzirwa kweAI uye zvikwata zveGPU zvakakwirira-zvikuru, kutonhodza mhepo kwechinyakare uye shongwe dzekutonhorera dzakavhurika hadzichakwanisi kuzadzisa zvinodiwa zvekutonga kwemafuta enzvimbo dzemazuva ano dzeAI.

 

Parizvino, makuru-akakura eAI supercomputing hubs, hyperscale cloud vendor AI bases uyebhizinesi -vakavaka maGPU masumbuzvose torayakavharwa-loop liquid cooling seyakajairika system inotsigira, ukuwo madiki-scale lightweight AI computing rooms anogona kuramba achitonhodzwa nemhepo sechirongwa chekuchinja. Yakavharwa-loop cooling ndiyo yave musimboti unosungirwa kugadzirwa kwe-load load AI computing scenarios, uye mafambiro ayo akazara-asingadzoreki muindasitiri.

 

 

 

 

Chikonzero chikuru nei-loop masisitimu akavharika achitongaAI inotonhorera nhemamukukwira kwakapinza kwechip heat load. New-chizvarwa cheAI machipi akaita seNVIDIA Blackwell, AMD MI300X uye maGPU epamba akakwira-performance GPUs ane simba rimwe chete-chip thermal power rinopfuura 700W, uye kabhatiti imwe chete inokwanisa kutakura zvinopisa zvinosvika 80–150kW, zvinodarika 10–20kW{1 kabati yemhepo.

 

Kutonhora kwemhepo kunovimba nemhepo inotenderedza kupisa kupisa, ine yakaderera kupisa kwekufambisa kushanda uye hombe fan simba rekushandisa, iyo inosundira iyo yakazara PUE yedata data kusvika 1.3-1.5.

 

Yakavharwa-loop fluid kutonhodza inoshandisa mapaipi akavharwa kutenderedza chinotonhodza zvakananga kumapuleti anotonhora akanamirwa kumachipisi, ane kupisa kwakanyanya kudarika 85%, kuderedza kushandiswa kwemagetsi ekutonhodza ne30%–40% uye kusunda PUE pasi pe1.1, uye 45℃inodziya yakavharwa{6}}loop scheme kubva kuNVIDIA inogona kutodzikisa PUE.08.

 

Iri simba-rekuchengetedza mukana rinoderedza zvakananga mutengo wekushandisa wemabhesi makuru eAI:50MW yepamusoro-chiyero cheAI computingcentre inogona kuchengetedza mari inodarika mamiriyoni mana emadhora ekuAmerica pagore rega rega mushure mekuchinja-loop cooling yakavharwa.

 

 

 

 

 

Dzvinyiriro yezviwanikwa zvemvura ndiyo imwe mutyairi mukuru wekutora kuzere-loop masisitimu munzvimbo dzeAI. Vhura shongwe dzekutonhodza dzinovimba nekupfava kwemvura pakubvisa kupisa, zvichipedza mamiriyoni emarita emvura yakachena gore rega rega pa megawati yemagetsi ekombuyuta, zvichikonzera zvirambidzo zvakasimba munzvimbo dzisina mvura-.

 

Yakavharwa-loop cooling inoumba maviri akazvimirira akavharirwa ma cycle: the internal circuit kubatanidza server cold plates uyeyakavharwa shongwe dzekutonhodzayakaparadzaniswa zvachose kubva kunharaunda yekunze, ine inotonhorera inozadzwa kamwe chete uye inodzokororwa kwehupenyu hwese kutenderera, inenge zero evaporation kurasikirwa panguva yekushanda kwakajairika.

 

Makesi eindasitiri anosanganisira OpenAI Stargate AI training base uye Microsoft itsva-generation computing park all adopt full closed-loop architecture, kuderedza kushandiswa kwemvura pagore kusvika pamwero weresitorendi diki, kugadzirisa zvakanyanya kurwadza kwekushandisa mvura kwe "AI water hogs".

 

Yakavharwa shongwe dzekutonhodza dzakafananidzwa nehurongwa hwakapatsanura mukati megadziriro mvura uye yekunze yekupfapfaidza mvura kuburikidza nesimbi isina tsvina, saka mvura yakachena inotenderera haimbobata guruva, algae uye tsvina mumhepo, kudzivirira kuyera uye kuparara kwe server inotonhora mahwendefa, uye kuwedzera hupenyu hwebasa remvura inotonhorera maseva nemakore anopfuura 5 zvichienzaniswa neakavhurika masisitimu.

 

 

 

Yakavharwa-loop cooling zvakare inounza kugadzikana kwemamiriro ekushanda akakosha kuAI computing. Kudzidzira hombe modhi kunoda 7 × 24h isingapindike yakagadzikana tembiricha kutonga; kuchinja kwekushisa kunokonzeresa kuderedzwa kweGPU frequency, zvichitungamira mukukanganisika kwekudzidziswa uye kurasikirwa kukuru kwekombuta.

 

Iyo loop yakavharwa yakavharwa{0}}loop masisitimu ane smart frequency conversion pumps, tembiricha sensing modules uye otomatiki bypass control, izvo zvinogona kuchengetedza inotonhorera inodziya tembiricha mukati me ±0.5 madhigirii gore rose.

 

Muchirimo, matsutso nehusiku-kunodzikira, sisitimu inochinja ichiendanatural free cooling mode, kuvhara chillers kuti iwedzere kuderedza kushandiswa kwesimba; muchirimo-kunopisa zvikuru, inoshanda pamwe chete nefiriji kuchengetedza kupisa kwakagadzikana pasina kukanganiswa neguruva,jecha uye hunyoro hwepamusoromunzvimbo yekunze.

 

Kusiyana neizvi, shongwe dzekutonhodza dzakavhurika dzinowanzovhara tsvina, zvinoda kugadzwa kemikari pamwe nekutsiviwa kwemvura, uye kugadzikana kwetembiricha hakuna kunaka, izvo zvisingakwanise kutsigira mabasa ekudzidzisa eAI kwenguva refu.

 

 

 

Zvakadaro, hadzisi dzese AI-dzimba dzemakombuyuta dzinechekuita dzinoshandisa zvakavharwa{1}}masisitimu ekutonhodza. Diki-kumucheto AI computing nodes, offline lightweight inference equipment uyemakamuri ekuyedza-akadereraine rack simba pazasi 20kW ichiri kushandisa kutonhora kwemhepo kana shongwe dzakavhurika dzekutonhodza kudzora kwekutanga kudyara.

 

Mamiriro akadai ane kupisa kwakadzika, nguva pfupi inoramba ichishanda uye kusanzwa kwemvura pakushandisa mvura, saka kuvhara-loop masisitimu achaunza mari yakawandisa yekuvaka nekugadzirisa.

 

Asi kunzvimbo dzepakati nepakati-dzakakura dzeAI computing dzine makabati anodarika gumi eGPU, mitemo yeindasitiri yenyika yenzvimbo dzakangwara dzemakompuyuta zvave kuvharwa -kutonhorera kwemvura kutonhora sechirongwa chinodiwa chinotonhora, uye mapurojekiti matsva epamusoro anodiwa kugadzirisa.yakavharwa-loop coolingshongwe semidziyo yakajairika inotsigira.

 

 

 

 

 

Kubva pamaonero ekudzokorora kwemaindasitiri, kuvhara-loop kutonhora kwave kuita kwakabatana dhizaini yepasi rose AI infrastructure. NVIDIA's DSX AI fekitori referensi dhizaini, dhizaini huru yevatengesi vemakore, inotsanangura 100% yakazara yakavharwa-loop liquid kutonhora pamapuratifomu ese eGPU computing, kukanzura kwakavakirwa{4}}mukutonhodza mafeni emaseva zvachose.

 

Zvemumbakumabvazuva-data-kumadokero-computing intelligent computing hubs, new energy AI training bases and pharmaceutical material computing platforms zvese zvinotora yakavharwa cooling tower inoenderana yakavharwa-loop liquid cooling, uchizadzisa zvinangwa zviviri zvekuchengetedza simba, kuchengetedza mvura uye high-density computing.

 

Nekuenderera mberi kukwidziridzwa kweAI chip simba, chikumbaridzo chekusungirirwa kuendesayakavharwa-loop masisitimuichadzikiswazve, uye zvirongwa zvekutonhodza mhepo zvichabviswa zvishoma nezvishoma mumamiriro epakati uye makuru emakomputa.

 

 

 

Muchidimbu, madiki ekumucheto AI michina inogona kutora kutonhora kwemhepo kwenguva, asi ese ari epakati uye makuru AI komputa nzvimbo, super kudzidzisa masumbu uye hyperscale intelligent computing hubs dzinofanira kushandisa.yakavharwa-loop liquid coolingmasisitimu anotsigirwa neshongwe dzekutonhodza dzakavharwa.

 

 

Mhinduro iyi inogadzirisa mapoinzi matatu ekurwadziwa kwekupisa kukuru, kushandiswa kwemvura yakawandisa uye kusagadzikana kwekudzora tembiricha yeAI computing, uye yave iyo yega nzira yekudzora kupisa kwemafuta kune yakakura-kukura kweindasitiri yehungwaru hwekugadzira.