Chii chaitika
Meta yakaburitsa mutsananguri muna Nyamavhuvhu 27 ichitsanangura matonhodzero anotonhodza ma-high density AI maseva. Iyo kambani inoti nzvimbo dzayo nyowani dzeAI-yakagadziridzwa data dzinonyanya kushandisa yakavharwa-loop mvura kutonhora, ukuwo kudzidzira kwekusimbisa kuri kushandiswa mukuenzanisa uye vatyairi kukwenenzvera kutonhora mashandiro.
Meta inoti kukura kwesimba uye kupisa kuburitsa kweAI Hardware kuri kuita kuti kutonhora kwemhepo kwakajairika kusashande kune dense server deployments. Iyo kambani inosiyanisa nyowani AI zvivakwa neyakamboita data-center muenzaniso muAltoona, Iowa, uko maraki ane gumi nematanhatu Nvidia H100 GPU akatonhodzwa zvachose nemhepo uye akashandisa mvura shoma. Zvinoenderana neMeta, dhizaini nyowani dzakawedzera kudiwa kweimwe nzira. Chirevo chepakati chechinyorwa ndechekuti kutonhodza chikamu chakakosha cheAI zvivakwa, kwete kungoita zvekuvaka-sevhisi.
Iyo system Meta inotsanangura inotenderera musanganiswa wemvura uye glycol zvakananga kuburikidza ne server hardware kutakura kupisa kure nema racks. Zvinopisa zvinopisa zvinobva zvatamisa kupisa kubva muinotonhodza musanganiswa mumwechete iwoyo usati wadzokera kumasevhisi mune inoenderera loop. Meta inoti mvura iyi haibviswe panzvimbo iyi uye inotarisira kuti inotonhorera irambe ichishanda kwemakore gumi isina kuisirwa. Iko kune chivakwa chisina kusanganiswa kwemvura-kutonhora zvivakwa, kambani inoti inoshandisa mhepo-inobatsirwa kutonhora kwemvura: akagovaniswa racks ane mapombi uye kupisa kunopisa izvo zvinopa imwecheteyo yakavharwa-loop maitiro pachiyero chidiki.
Meta zvakare inopokana kuti kutonhora kwemvura kunogona kuvandudza kusimba kwemuviri. Mutsananguri wake anoti mhepo inotonhodza maseva mamwe chete ingada kaviri sevha-tireyi nzvimbo nekuda kweimwe yekutonhodza michina. Yakananga-kune-chip kutonhora kwemvura, nekusiyana, inogona kubvumira maGPU akawanda kuti akwane mune imwechete rack, kuderedza huwandu hwema racks anodiwa kune imwe nzvimbo ine yakapihwa komputa. Meta inoenderera mberi ichiti yakajairika AI-yakagadziridzwa data centre inoshandisa yakavharwa-loop kutonhora uye yakaoma inotonhorera inoshandisa mvura shoma pagore pane akati wandei-sevhisi maresitorendi, asi haipe iwo epagore mavhoriyamu kana hwaro chaihwo hwekufananidza ikoko.
Chinyorwa chinowedzera chikamu chesoftware. Meta inoti mainjiniya akavaka fizikisi-based data-center simulator inoenzanisira mamiriro ekunze, sevha inotakura uye inotonhorera-zvishandiso maitiro. Kusimbisa kudzidza kunogona kuyedza sarudzo dzekutonhodza munzvimbo iyoyo yakateedzerwa pane kutanga kuyedza pane inorarama nzvimbo. Meta inoti nzira iyi yakatambanudzwa kubva pakuyedza kuenda kunzvimbo dzedhata dzakatonhodzwa nemhepo pazvikepe zvayo uye kuti mutyairi wendege pane imwe saiti akaderedza simba rinoshandiswa nemhepo inotonhorera mafeni neavhareji ye20% uku achideredza kushandiswa kwemvura ne4% mumamiriro ekunze akasiyana. Nzvimbo yacho hairatidze nzvimbo yekuyedza, nguva yayo, mashandisiro ayo ekutanga kana huwandu hwezvivakwa zvinobatanidzwa mukuburitswa kwakafara.
Kwakabva mashoko: about.fb.com ↗
Nei zvichikosha
Kutonhodza kuri kuita chinomanikidza sezvo AI hardware inogadzira kupisa kwakawanda uye inoda denser deployments. Meta's account inoratidza kuti kutonhora kwemvura uye software-based optimization inogona kukanganisa simba, mvura, nzvimbo uye zvivakwa zvinodiwa zvenzvimbo hombe dzeAI komputa, kunyangwe mhedzisiro yakataurwa ichibva kuMeta uye mutyairi mumwe chete.
AI zvivakwa zvinowanzo kurukurwa maererano nemachipisi nemagetsi, asi sosi inotaridza bhodhoro risingaonekwe remuviri: kubvisa kupisa kubva pamichina yekombuta yakasimba. Kana kutonhora kwemvura kuchibvumira mamwe maGPU mune imwechete rack, vashandisi vanogona kukwanisa kuwedzera komputa mukana pasina kuwedzera chivakwa chetsoka pamwero wakafanana. Izvo zvinogona kukanganisa dhizaini uye mutengo wezvivakwa zvinoshandiswa kudzidzisa nekumhanyisa maAI masisitimu.
Kushandiswa kwemvura chinhu chinonetsa veruzhinji chakatenderedza nzvimbo dzedata, asi sosi yacho inoiratidza sei zvinoenderana nedhizaini yekutonhodza kwete chinhu otomatiki cheAI komputa. Meta's yakavharwa-loop tsananguro inoratidza kuti mvura inogona kushanda zvakanyanya seinodzoreredza inotonhorera pane kugara ichidyiwa kana kuburitswa. Musiyano uyu wakanyanya kukosha kunharaunda dzinoongorora zvivakwa zvakarongwa, kunyangwe kuenzanisa kwechinyorwa kweresitorendi kusina kukwana kuratidza huwandu hwemvura hweMeta's data-center fleet.
Iyo yakashumwa yekusimbisa-kudzidza mutyairi yakakosha nekuti inobatanidza nzira dzeAI nekushanda kweAI zvivakwa. Iyo modhi haina kurondedzerwa sekutonga hurongwa hwese-chinangwa kana kuita sarudzo dzisina kutarisirwa nezvesevhisi yeruzhinji; inoshandiswa kuyedza nzira dzekutonhora mune simulator inomiririra mamiriro emuviri. Meta inoti mutemo wakabuda wakaderedza simba re fan ne 20% paavhareji uye kushandiswa kwemvura ne 4% mukutyaira. Kana ikadzokororwa, optimization yakadai inogona kudzikisa mutengo wekushandisa uye kudiwa kwezviwanikwa pasina kuda shanduko kune iri pasi peAI basa rekuita.
Iyo sosi zvakare inotsanangura nzira yekugovana indasitiri. Meta inoti Open Project yayo, yakavambwa muna 2011, inotarisirwa kuita kuti data-center infrastructure ishande, iwedzere uye irambe yakasimba, uye inocherekedza kuti kambani yakazivisa IcePack liquid-yakatonhorera network-rack platform muna 2025 yekugovana mahara kuburikidza neprojekti iyoyo. Icho chiziviso chepakutanga chiri kumashure kwete nhau itsva muchinyorwa chino. Mupiro mutsva, unoenderana nenguva ndeye Meta tsananguro yakadzama yezvivakwa zvinotonhorera uye nhoroondo yayo yekusimbisa-yekudzidza mhedzisiro, kwete kuburitswa kwechigadzirwa chitsva kana yakazvimiririra yakasimbiswa indasitiri-yakafara mwero.
Interactive Mechanism: Iyo Inonyatsoshanda
Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.
Which component of an AI application is the machine-learning model itself?
Zvekutarisa zvinotevera
Mibvunzo yakakosha ndeyekuti Meta yakachengetwa yakachengetwa inobata munzvimbo dzakawanda, mamiriro ekunze uye mabasa, uye kuti imvura yakawanda sei uye simba rinoshandiswa nemasisitimu zvachose. Tarisa uone yakazvimiririra kusimbiswa, ruzivo nezve chiyero chekutumira, uye humbowo hwekuti madhizaini akavhurika uye nzira dzekusimbisa-yekudzidza dzinobatsira kupfuura Meta zvivakwa zvayo.
Chinonyanya kukosha ndechekusimbisa. Huwandu hwese hwekuita mune sosi ndeye Meta's zvichemo, uye chinyorwa hachitaure yakazvimirira odhiyo, ongororo-yakaongororwa nevezera kana zviyero kubva kunze kwekambani. Iyo 20% fan-simba kudzikisira uye 4% kudzikiswa kwemvura nekudaro kunofanirwa kubatwa seyakashumwa mhinduro dzemutyairi, kwete kusimbiswa mhedzisiro yeAI data nzvimbo kazhinji. Kuzivisa kweramangwana kunofanirwa kutsvaga saiti yemutyairi, nguva, mamiriro ekunze, chimiro chebasa, zviyero zvekutanga uye kusiyanisa kwehuwandu.
Mhedziso yekushandisa zviwanikwa haisati yazivikanwa. Meta inopa kudzikisira uye kuenzanisa nemaresitorendi maviri akazara-sevhisi, asi haitauri mashandisirwo emvura epagore enzvimbo yedata, kushandiswa kwemagetsi, kutonhora mutoro kana huwandu hwakazara hweGPU. Kuenzanisa zvakare hakutsananguri kana kunosanganisira mvura inoshandiswa kumwe kunhu munzvimbo kana mukugadzira magetsi. Pasina iwo manhamba, vaverengi havagone kuverenga chiyero chezvakachengetwa kana kuenzanisa sisitimu nemamwe madhizaini ekutonhodza.
Deployment scope mumwe mubvunzo wakavhurika. Meta inoti ruzhinji rwenzvimbo dzayo nyowani dzeAI-yakagadziridzwa data dzinoshandisa yakavharwa-loop kutonhora kwemvura uye kuti nzira yekusimbisa-yekudzidza yakamisikidzwa kumhepo-yakatonhorera nzvimbo dzedata mungarava dzayo, asi hazviratidze kuti zvingani zvivakwa zvakavharwa kana masisitimu ari kushanda achienderera. Iyo kambani zvakare inoti inotonhorera inogona kugara kusvika kumakore gumi, tarisiro inogona kuenderana nekugadzirisa, kusvibiswa kwekudzora, dhizaini yemidziyo uye mamiriro ekushanda emunharaunda ayo asingakurukurwe nekwakabva.
Muedzo wakakura uchave wekuti nzira yacho inofamba kupfuura Meta. Yakavharwa-loop performance inogona kusiyana nemamiriro ekunze, dhizaini yekuvaka, kugadzirwa kwehardware, rack kumisikidzwa uye mutoro webasa. Masisitimu anobatsirwa nemhepo anogona kubatsira kune retrofits, asi sosi yacho haipe inofananidzwa mutengo, kuvimbika kana kuchengetedza data. Tarisa zviyero zvakazvimirira, zvinoburitswa kubva kumasangano anoshandisa IcePack kana magadzirirwo akafanana, uye humbowo hwekuti simulation-akadzidziswa kutonhora marongero anoramba akachengeteka kana mamiriro awira kunze kwezviratidziro. Izvo hazvina kujeka zvakare kana ramangwana, rakawanda-dense-dense AI hardware ichachengetedza yakashumwa kubudirira kubudirira kana kugadzira zvitsva zvekutonhora zvinodiwa.