Zvinonzi makambani etekinoroji arikutsvagira maGPU kana munzira yekuatora. Muna Kubvumbi, CEO weTesla Elon Musk akatenga zviuru gumi zveGPU uye akataura kuti kambani icharamba ichitenga akawanda maGPU kubva kuNVIDIA. Kudivi rebhizinesi, vashandi veIT vari kusundidzirawo nesimba kuti vaone kuti maGPU anogara achishandiswa kuwedzera kudzoka pakudyara. Nekudaro, mamwe makambani anogona kuona kuti nepo huwandu hweGPU huchiwedzera, kusashanda kweGPU kunowedzera kuoma.
Kana nhoroondo yakatidzidzisa chero chinhu pamusoro pe-high-performance computing (HPC), ndezvekuti kuchengetedza uye networking haifaniri kubayirwa pamutengo wekutarisa zvakanyanya pamakomputa. Kana chengetedzo isingakwanise kuendesa data kumakomputa zvikamu, kunyangwe uine maGPU akawanda munyika, haugone kunyatso shanda.
Sekureva kwaMike Matchett, muongorori paDiki World Big Data, madiki mamodheru anogona kuuraiwa mundangariro (RAM), achibvumira zvakanyanya kutarisisa pamakomputa. Nekudaro, mamodheru akakura seChatGPT ane mabhiriyoni emanodhi haagone kuchengetwa mundangariro nekuda kwekudhura.
"Iwe haugone kukwana mabhiriyoni emanodhi mundangariro, saka kuchengetedza kunotonyanya kukosha," anodaro Matchett. Nehurombo, kuchengetedza data kunowanzo furatirwa panguva yekuronga.
Kazhinji, zvisinei nekesi yekushandiswa, pane zvina zvakajairika mapoinzi mumuenzaniso wekudzidzisa maitiro:
1. Kudzidzisa Muenzaniso
2. Inference Application
3. Kuchengeta Data
4. Yakakurumidza Computing
Paunenge uchigadzira uye nekuisa modhi, zvakawanda zvinodikanwa zvinotungamira nekukurumidza humbowo-hwe-pfungwa (POC) kana nharaunda dzekuyedza kuti utange kudzidzisa modhi, nekuchengetedza data kusingatariswe zvakanyanya.
Nekudaro, dambudziko riri pakuti kudzidziswa kana kuendesa kuendesa kunogona kugara kwemwedzi kana makore. Makambani mazhinji anokurumidza kukwidza saizi yavo yemhando panguva ino, uye zvivakwa zvinofanirwa kuwedzera kuti zvigadzirire mamodheru ari kukura uye dataset.
Tsvagiridzo kubva kuGoogle pamamirioni eML yekudzidzira basa remedu rinoratidza kuti avhareji ye30% yenguva yekudzidziswa inopedzerwa pane yekuisa data pombi. Nepo tsvakiridzo yapfuura yakatarisana nekugadzirisa maGPU kuti akurumidze kudzidziswa, matambudziko mazhinji achiri mukugadzirisa zvikamu zvakasiyana zvepombi yedata. Kana iwe uine yakakura computational simba, iyo chaiyo bhodhoro inova kukurumidza kwaungadyisa data mumakomputa kuti uwane mibairo.
Kunyanya, zvinonetsa mukuchengetedza data uye manejimendi zvinoda kuronga kukura kwedata, zvichikubvumidza kuti urambe uchibvisa kukosha kwedata paunenge uchifambira mberi, kunyanya kana iwe uchinge wapinda mune epamusoro-soro kesi dzekushandisa dzakadai sekudzidza kwakadzama uye neural network, iyo inoisa zvichemo zvepamusoro. kuchengetedza maererano nekugona, kuita, uye scalability.
Zvikuru sei:
Scalability
Kudzidza kwemichina kunoda kubata huwandu hwakawanda hwe data, uye sezvo huwandu hwe data hunowedzera, huchokwadi hwemhando zvakare hunovandudza. Izvi zvinoreva kuti mabhizinesi anofanirwa kuunganidza uye kuchengetedza data rakawanda mazuva ese. Kana chengetedzo isingakwanise kuyera, data-yakanyanya kuwanda basa rinogadzira mabhodhoro, kudzikisira kuita uye zvichikonzera inodhura GPU nguva isina basa.
Kuchinja-chinja
Flexible tsigiro yemaprotocol akawanda (kusanganisira NFS, SMB, HTTP, FTP, HDFS, uye S3) inodiwa kuti isangane nezvinodiwa zvemasisitimu akasiyana, pane kungogumira kune imwe mhando yenzvimbo.
Kunonoka
I/O latency yakakosha pakuvaka uye kushandisa mamodheru sezvo data ichiverengwa uye kuverengwazve kakawanda. Kuderedza I/O latency kunogona kupfupisa nguva yekudzidziswa yemamodhi nemazuva kana mwedzi. Inokurumidza modhi yekuvandudza inoshandura zvakananga kune yakakura mabhenefiti ebhizinesi.
throughput
Kubuda kwemasisitimu ekuchengetedza kwakakosha pakudzidziswa kwemuenzaniso. Maitiro ekudzidzisa anosanganisira huwandu hukuru hwe data, kazhinji mumaterabytes paawa.
Parallel Access
Kuti uwane huwandu hwepamusoro, mamodheru ekudzidzisa anopatsanura zviitiko kuita akawanda akafanana mabasa. Izvi zvinowanzoreva kuti muchina kudzidza algorithms inowana iwo mafaera akafanana kubva kune akawanda maitiro (zvichida pane akawanda emuviri maseva) panguva imwe chete. Iyo yekuchengetedza system inofanirwa kubata zvinodiwa panguva imwe chete pasina kukanganisa kuita.
Nekugona kwayo kwakatanhamara mu low latency, yakakwira throughput, uye hukuru-hukuru hwakafanana I/O, Dell PowerScale ndiyo yakanaka yekuchengetedza inoenderana neGPU-inomhanyisa komputa. PowerScale inonyatso kudzikisa nguva inodiwa yekuongorora mhando dzinodzidzisa uye kuyedza akawanda-terabyte dataset. MuPowerScale yose-flash kuchengetedza, bandwidth inowedzera ne18 nguva, kubvisa mabhodhoro eI / O, uye inogona kuwedzerwa kumasumbu eIsilon aripo kuti ikurumidze uye kusunungura kukosha kwehuwandu hwemashoko asina kurongeka.
Uyezve, PowerScale's multi-protocol yekuwana kugona inopa kuchinjika kusingagumi kwekumhanyisa basa, ichibvumira data kuchengetwa uchishandisa imwe protocol uye kuwanikwa uchishandisa imwe. Kunyanya, iwo ane simba maficha, kuchinjika, scalability, uye bhizinesi-giredhi mashandiro ePowerScale papuratifomu anobatsira kugadzirisa anotevera matambudziko:
-Kurumidza kuvandudza nekusvika 2.7 nguva, kuderedza modhi yekudzidzira kutenderera.
-Bvisa mabhodhoro eI / O uye upe nekukurumidza modhi yekudzidziswa uye kusimbiswa, yakagadziridzwa modhi yechokwadi, yakawedzera data sainzi chigadzirwa, uye yakawedzera kudzoka pakudyara kwekombuta nekusimudzira mabhizinesi-giredhi maficha, kuita kwepamusoro, concurrency, uye scalability. Wedzera huchokwadi hwemodhi neakadzama, epamusoro-resolution datasets nekusimudzira anosvika 119 PB yeanoshanda kuchengetedza kugona musumbu rimwe chete.
-Kuwana kutumirwa pachiyero nekutanga diki uye yakazvimirira kuyera komputa uye kuchengetedza, kuendesa yakasimba kuchengetedza data uye kuchengetedza sarudzo.
-Kuvandudza kugadzirwa kwesainzi yedata ne-in-place analytics uye pre-yakatenderwa mhinduro dzekukurumidza, dzakaderera-njodzi yekutumirwa.
-Kushandisa madhizaini akasimbiswa akavakirwa pamhando yepamusoro-ye-tekinoroji, kusanganisira NVIDIA GPU kukwidziridza uye mareferenzi ekuvaka ane NVIDIA DGX masisitimu. Kuita kwePowerScale kwepamusoro uye kuwirirana kunosangana nezvinodiwa zvekuchengetedza padanho rega rega rekudzidza muchina, kubva pakutora data uye kugadzirira kusvika kumuenzaniso kudzidziswa uye inference. Pamwe chete neOneFS operating system, node dzose dzinogona kushanda zvisina mutsetse mukati meboka rimwe chete rinofambiswa neOneFS, riine zvinhu zvemabhizinesi-pamwero senge manejimendi ekuita, manejimendi edata, chengetedzo, uye kuchengetedzwa kwedata, zvichiita kuti kukurumidza kupera kwekudzidziswa kwemuenzaniso uye kusimbiswa kwebhizinesi.
Nguva yekutumira: Jul-03-2023
