Nws tau hais tias cov tuam txhab thev naus laus zis tab tom sib tw rau GPUs lossis ntawm txoj hauv kev kom tau txais lawv. Thaum lub Plaub Hlis, Tesla CEO Elon Musk yuav 10,000 GPUs thiab tau hais tias lub tuam txhab yuav txuas ntxiv mus yuav ntau tus GPUs los ntawm NVIDIA. Ntawm sab kev lag luam, IT cov neeg ua haujlwm tseem tab tom thawb nyuaj los xyuas kom meej tias GPUs tau siv tas li los ua kom rov qab los ntawm kev nqis peev ntau. Txawm li cas los xij, qee lub tuam txhab yuav pom tias thaum tus naj npawb ntawm GPUs nce, GPU idleness yuav hnyav dua.
Yog tias keeb kwm tau qhia peb txhua yam txog kev ua haujlwm siab (HPC), nws yog qhov chaw cia thiab kev sib txuas lus yuav tsum tsis txhob muab fij rau ntawm kev tsom mus rau kev suav ntau dhau. Yog tias kev khaws cia tsis tuaj yeem hloov cov ntaub ntawv mus rau cov khoos phis tawj, txawm tias koj muaj GPU ntau tshaj plaws hauv ntiaj teb, koj yuav tsis ua tiav qhov ua tau zoo.
Raws li Mike Matchett, tus kws tshuaj ntsuam ntawm Small World Big Data, cov qauv me tuaj yeem ua tiav hauv lub cim xeeb (RAM), tso cai rau kev tsom mus rau kev suav ntau dua. Txawm li cas los xij, cov qauv loj dua li ChatGPT nrog ntau txhiab tus ntawm cov nodes tsis tuaj yeem khaws cia hauv lub cim xeeb vim yog tus nqi siab.
Matchett hais tias "Koj tsis tuaj yeem haum ntau lab ntawm cov nodes hauv lub cim xeeb, yog li kev khaws cia tseem ceeb dua," Mattett hais. Hmoov tsis zoo, cov ntaub ntawv khaws cia feem ntau tsis saib xyuas thaum lub sijhawm npaj txheej txheem.
Feem ntau, tsis hais txog kev siv li cas, muaj plaub lub ntsiab lus hauv cov txheej txheem kev cob qhia tus qauv:
1. Kev cob qhia qauv
2. Daim ntawv thov kev xav
3. Cov ntaub ntawv khaws cia
4. Kev ua lej nrawm
Thaum tsim thiab siv cov qauv, feem ntau yuav tsum tau ua qhov tseem ceeb ntawm cov ntaub ntawv pov thawj ceev-ntawm-lub tswv yim (POC) lossis kev sim ib puag ncig los pib ua qauv kev cob qhia, nrog rau cov ntaub ntawv khaws cia yuav tsum tsis txhob xav txog sab saum toj.
Txawm li cas los xij, qhov kev sib tw nyob rau hauv qhov tseeb tias kev cob qhia lossis kev xa tawm tuaj yeem siv sijhawm ntev los yog ntau xyoo. Ntau lub tuam txhab tau nce lawv cov qauv ntau thiab tsawg thaum lub sijhawm no, thiab cov txheej txheem yuav tsum nthuav dav kom haum rau cov qauv loj hlob thiab cov ntaub ntawv.
Kev tshawb fawb los ntawm Google ntawm ntau lab ntawm ML kev cob qhia ua haujlwm qhia tau hais tias qhov nruab nrab ntawm 30% ntawm lub sijhawm cob qhia yog siv rau hauv cov ntaub ntawv xa tawm. Thaum cov kev tshawb fawb yav dhau los tau tsom mus rau kev ua kom zoo dua GPUs kom ceev kev cob qhia, ntau qhov kev sib tw tseem nyob hauv kev txhim kho ntau qhov chaw ntawm cov ntaub ntawv xa mus. Thaum koj muaj lub hwj chim tseem ceeb hauv kev suav, lub fwj tiag tiag yuav sai npaum li cas koj tuaj yeem pub cov ntaub ntawv rau hauv kev suav kom tau txais txiaj ntsig.
Tshwj xeeb, cov kev cov nyom hauv cov ntaub ntawv khaws cia thiab kev tswj hwm xav tau kev npaj rau kev loj hlob ntawm cov ntaub ntawv, tso cai rau koj txuas ntxiv rho tawm tus nqi ntawm cov ntaub ntawv thaum koj ua tiav, tshwj xeeb tshaj yog thaum koj nkag mus rau hauv kev siv ntau dua xws li kev kawm sib sib zog nqus thiab neural networks, uas ua rau muaj kev xav tau ntau dua. cia nyob rau hauv cov nqe lus ntawm muaj peev xwm, kev ua tau zoo, thiab scalability.
Tshwj xeeb:
Scalability
Kev kawm tshuab yuav tsum tuav cov ntaub ntawv ntau, thiab raws li qhov ntim ntawm cov ntaub ntawv nce, qhov tseeb ntawm cov qauv kuj tseem txhim kho. Qhov no txhais tau tias cov lag luam yuav tsum sau thiab khaws cov ntaub ntawv ntau dua txhua hnub. Thaum qhov chaw cia tsis tuaj yeem ntsuas, cov ntaub ntawv ua haujlwm hnyav ua rau muaj kev tsis sib haum xeeb, txwv kev ua haujlwm thiab ua rau lub sijhawm GPU kim heev.
Yooj yim
Kev txhawb nqa yooj yim rau ntau txoj cai (xws li NFS, SMB, HTTP, FTP, HDFS, thiab S3) yog qhov tsim nyog los ua kom tau raws li qhov xav tau ntawm cov tshuab sib txawv, tsis yog txwv rau ib hom ib puag ncig.
Latency
I/O latency yog qhov tseem ceeb rau kev tsim thiab siv cov qauv raws li cov ntaub ntawv tau nyeem thiab rov nyeem ntau zaus. Kev txo qis I / O latency tuaj yeem txo lub sijhawm kawm ntawm cov qauv los ntawm hnub lossis hli. Kev txhim kho tus qauv nrawm dua ncaj qha txhais rau kev lag luam zoo dua.
Tshaj tawm
Kev xa tawm ntawm cov tshuab khaws cia yog qhov tseem ceeb rau kev qhia ua qauv zoo. Cov txheej txheem kev cob qhia suav nrog ntau cov ntaub ntawv, feem ntau hauv terabytes ib teev.
Parallel Access
Txhawm rau kom ua tiav cov txiaj ntsig siab, cov qauv kev cob qhia tau faib cov haujlwm ua ntau yam haujlwm sib npaug. Qhov no feem ntau txhais tau hais tias tshuab kev kawm algorithms nkag mus rau tib cov ntaub ntawv los ntawm ntau cov txheej txheem (muaj peev xwm ntawm ntau lub cev servers) ib txhij. Lub kaw lus cia yuav tsum ua raws li cov kev xav tau sib thooj yam tsis muaj kev cuam tshuam kev ua haujlwm.
Nrog nws cov peev xwm zoo tshaj plaws hauv qis latency, high throughput, thiab loj-scale thaum uas tig mus I / O, Dell PowerScale yog ib qho zoo tagnrho cia ntxiv rau GPU-accelerated xam. PowerScale txo qis lub sijhawm xav tau rau cov qauv tshuaj ntsuam uas qhia thiab ntsuas cov ntaub ntawv ntau-terabyte. Hauv PowerScale tag nrho-flash cia, bandwidth nce 18 zaug, tshem tawm I / O cov fwj, thiab tuaj yeem muab ntxiv rau Isilon pawg uas twb muaj lawm kom nrawm thiab qhib tus nqi ntawm ntau cov ntaub ntawv tsis tsim nyog.
Ntxiv mus, PowerScale's multi-protocol access capabilities muab unlimited yooj rau khiav hauj lwm loads, tso cai rau cov ntaub ntawv khaws cia siv ib tug raws tu qauv thiab nkag mus siv lwm. Tshwj xeeb tshaj yog, cov yam ntxwv muaj zog, hloov tau yooj yim, scalability, thiab kev ua haujlwm qib kev ua haujlwm ntawm PowerScale platform pab daws cov teeb meem hauv qab no:
- Ua kom nrawm nrawm nrawm txog li 2.7 zaug, txo cov qauv kev cob qhia lub voj voog.
- Tshem tawm I / O lub fwj thiab muab cov qauv kev cob qhia sai thiab siv tau, txhim kho tus qauv raug, txhim kho cov ntaub ntawv kev tshawb fawb tsim tau, thiab ua kom rov qab tau zoo tshaj plaws ntawm kev nqis peev los ntawm kev ua lag luam-qib nta, kev ua haujlwm siab, sib npaug, thiab scalability. Txhim kho cov qauv raug nrog cov ntaub ntawv sib sib zog nqus, kev daws teeb meem siab dua los ntawm kev siv mus txog 119 PB ntawm qhov muaj peev xwm khaws tau zoo hauv ib pawg.
- Ua tiav kev xa tawm ntawm qhov ntsuas los ntawm kev pib me me thiab nws tus kheej ntsuas kev suav thiab khaws cia, xa cov ntaub ntawv muaj kev tiv thaiv thiab kev xaiv kev nyab xeeb.
- Txhim kho cov ntaub ntawv kev tshawb fawb kev tsim khoom nrog rau hauv qhov chaw analytics thiab pre-validated kev daws teeb meem kom sai dua, tsis tshua muaj kev xa mus.
- Leveraging proven designs based on best-of-breed technologies, suav nrog NVIDIA GPU acceleration thiab reference architectures with NVIDIA DGX systems. PowerScale qhov kev ua tau zoo thiab qhov sib txuam ua tau raws li qhov yuav tsum tau ua ntawm kev khaws cia ntawm txhua theem ntawm kev kawm tshuab, los ntawm kev nrhiav cov ntaub ntawv thiab kev npaj mus rau cov qauv kev cob qhia thiab kev xav. Ua ke nrog OneFS kev khiav hauj lwm qhov system, tag nrho cov nodes tuaj yeem ua haujlwm zoo nyob rau hauv tib lub OneFS-tsav pawg, nrog rau kev lag luam-theem nta xws li kev tswj hwm kev ua haujlwm, kev tswj hwm cov ntaub ntawv, kev ruaj ntseg, thiab kev tiv thaiv cov ntaub ntawv, ua kom tiav cov qauv kev cob qhia sai thiab siv tau rau kev lag luam.
Post lub sij hawm: Jul-03-2023
