Kua txias Boosts Performance thiab Efficiency ntawm Nvidia's Blackwell Platform

Oct 14, 2024

Tso lus

Kev loj hlob sai ntawm kev txawj ntse txawj ntse (AI), tshwj xeeb tshaj yog nyob rau hauv cov cheeb tsam xws li Generative AI (GenAI) thiab cov qauv lus loj (LLMs), tau ua rau muaj qhov tsis tau muaj dua qhov xav tau rau kev suav lub zog. Raws li AI qauv dhau los ua qhov nyuaj thiab xav tau cov ntaub ntawv ntau ntxiv, cov khoom siv xav tau los txhawb cov kev nce qib no ntsib cov teeb meem tseem ceeb, tshwj xeeb tshaj yog nrog cov cua sov. Cov txheej txheem cua txias, zoo li cua txias, tsis txaus los tswj lub zog noj ntawm cov khoom siv AI txiav, thiab qhov no yog qhov kev daws teeb meem ua kua txias.

Cov tshuab ua kua cua txias tau txais txiaj ntsig zoo raws li txoj hauv kev muaj txiaj ntsig zoo los tswj cov thermal tso tawm ntawm AI cov ntaub ntawv chaw. Nws yog ib qho tseem ceeb tshwj xeeb tshaj yog rau cov platforms xws li Nvidia's Blackwell architecture, qhov twg lub zog suav loj yuav tsum tau ua kom txias txias kom ntseeg tau tias kev ua tau zoo thiab siv tau ntev.

Nyob ntawmHotchip 2024lub rooj sib tham, Nvidia qhia txog nws txoj kev daws teeb meem, kev sib koom ua kedej sov Direct rau Chip kua txiasthev naus laus zis nrog nws cov cuab yeej Blackwell los daws cov teeb meem kev siv hluav taws xob thiab cua txias. Qhov kev txhim kho no qhia txog kev nce qib tseem ceeb hauv AI kho vajtse tsim, txo cov nqi khiav lag luam thiab txhim kho kev ua haujlwm zoo hauv kev siv AI loj.

Hot Chips Tutorial: Liquid Cooling Boosts Performance and Efficiency

▲ Kub Chips Tutorial: Ua Kua Txias Boosts Performance and Efficiency

Qhov nce ntawm Liquid Cooling Technology

Cov tshuab ua kua txias tau tshwm sim los ua ib qho tseem ceeb hauv AI cov ntaub ntawv chaw tsim khoom vim qhov kev xav tau ntawm AI chips loj zuj zus. Raws li AI daim ntawv thov loj hlob, kev siv hluav taws xob hauv cov chaw khaws ntaub ntawv xav tias yuav nce siab heev. Ntau AI processors, suav nrog Nvidia's H100 thiab nws cov tshiabBlackwell architecture, siv qhov twg los ntawm 700W txog 1200W ntawm lub zog. Qhov kev kos lub zog loj no yog sib xyaw ua ke thaum AI pawg, suav nrog ntau txhiab GPUs, ua haujlwm hauv kev sib koom ua ke.

Piv txwv li,Elon Musk's AI kev cob qhia pawg, qhov loj tshaj plaws nyob rau hauv lub ntiaj teb no nrog 100, 000 H100 GPUs, yog tag nrho cov kua-txias los tswj nws 31-megawatt zog thov. Xws li cov piv txwv qhia tau hais tias yog vim li cas kev ua kom zoo dua cov tshuab cua txias yog qhov tseem ceeb-tsis yog txo cov nqi ua haujlwm xwb tab sis kuj tseem txhim kho kev ua tau zoo ntawm AI systems. Raws li qhov xav tau rau kev ua tau zoo dua AI txuas ntxiv mus, cov kua txias yuav ua lub luag haujlwm tseem ceeb hauv kev txhawb nqa kev ua haujlwm siab.

Cov txiaj ntsig ntawm Cov Dej Txias Dej hauv AI Data Centers

Cov kua ua kua txias sawv tawm vim nws lub peev xwm los hloov hluav taws xob ncaj qha los ntawm cov khoom tseem ceeb xws li CPUs thiab GPUs, txhim kho cov cua sov dissipation piv rau cov cua txias. HauvDirect Liquid Cooling (DLC), coolant tuaj rau hauv kev sib cuag ncaj qha nrog cov nti, txhim kho thermal efficiency thiab txo qhov xav tau ntawm cov kiv cua loj thiab cov tshuab cua txias. Qhov no ua rau txo qis kev siv hluav taws xob rau kev ua kom txias, yog li txo cov nqi ua haujlwm tag nrho ntawm cov ntaub ntawv chaw.

Tsis tas li ntawd,Immersion Liquid Cooling, qhov twg tag nrho cov servers yog submerged nyob rau hauv cov kua dielectric, nthuav tawm ib qho kev ua kom txias dua. Txoj kev no tsis tsuas yog ua kom tag nrho cov khoom nyob twj ywm txias tab sis kuj txo qis cov khoom siv kho tshuab, uas ua rau lub neej ntev ntawm cov khoom siv, thiab txo cov suab nrov los ntawm cov khoom siv cua txav.

Kev Sib Tw Kev Siv Hluav Taws Xob hauv AI Hardware

Raws li AI systems ntsuas, lawv lub zog siv tau nthuav tawm qhov kev sib tw loj zuj zus. AI hardware, xws li covNvidia H100 GPUthiab covBlackwell architecture, paub tias yog lub zog tshaib plab, nrog lub zog xav tau txog li 1200W ib nti. Ib pawg AI ib txwm muaj 22, 000 H100 GPUs, piv txwv li, yuav xav tau txog li31 megawatts hluav taws xob- sib npaug rau kev siv hluav taws xob ntawm lub nroog me me.

Qhov kev thov lub zog loj no tsis yog tsuas yog ua rau cov nqi ua haujlwm ntawm cov chaw zov me nyuam cov ntaub ntawv xwb tab sis tseem ua rau muaj kev cuam tshuam loj rau ib puag ncig. Txhawm rau daws cov teeb meem no, cov chaw khaws ntaub ntawv yuav tsum tsom mus rau ob qho tib si txo kev siv hluav taws xob thiab txhim kho kev ua kom txias.

Nvidia's Blackwell Architecture thiab Sov-Dej Direct rau Chip Cooling

Nyob ntawmHotchip 2024 lub rooj sib tham, Nvidia nthuav tawm nws cov kev daws teeb meem rau kev sib xyaw ua kua txias nrog nws cov qauv tsim Blackwell, sivdej sov Direct rau Chip txias tshuab. Txoj hauv kev no siv dej sov (tsis yog dej txias) kom nqus tau thiab xa cov cua sov ncaj qha los ntawm nti. Los ntawm kev ua haujlwm sov-dej txias, Nvidia tuaj yeem txo qhov txias zog siv ntawm cov chaw khaws ntaub ntawv txog li 28%.

Qhov kev ua tau zoo ntawm cov tshuaj no yog ob npaug: nws tsis tsuas yog txo qis tag nrho lub zog kos rau qhov txias tab sis kuj tseem ua rau rov qab los ntawm cov cua sov pov tseg, uas tuaj yeem rov ua dua rau lwm yam kev siv, xws li cua sov cov tsev nyob ze. Tsis tas li ntawd, cov dej txias txias ua rau lub neej ua haujlwm ntev ntawm cov servers los ntawm kev ua kom cov chips nyob rau hauv qhov kub thiab txias, tiv thaiv overheating, thiab txo qhov hnav thiab tsim kua muag.

Warm-water Direct to Chip cooling solution for Nvidia's Blackwell architecture efficiently transfers heat and improves cooling performance in AI data centers.

▲Sov-dej Direct rau Chip Cooling Solution

Cov txheej txheem cua txias no yog qhov tseem ceeb tshwj xeeb tshaj yog cov ntawv thov AI xws li GenAI thiab LLMs txuas ntxiv ua kom lub zog siv hluav taws xob xav tau hauv cov chaw khaws ntaub ntawv. Kev muaj peev xwm tswj tau qhov ntsuas kub zoo ncaj qha cuam tshuam rau kev ua haujlwm thiab kev ua kom muaj peev xwm ntawm AI cov haujlwm ua haujlwm, kom ntseeg tau tias cov tshuab no tuaj yeem ua raws li cov kev xav tau hnyav rau lawv.

Immersion Liquid Cooling: Ib kauj ruam dhau

Ntxiv rau Direct Liquid Cooling,Immersion Liquid Coolingtseem tau txais traction raws li kev daws teeb meem txuas ntxiv rau cov tshuab AI loj. Txoj kev no submerges tag nrho cov servers nyob rau hauv ib tug uas tsis yog-conductive, dielectric kua, uas tag nrho absorbs thiab dissipates tshav kub los ntawm tag nrho cov Cheebtsam. Los ntawm kev ua kom txias tag nrho hauv qhov no, Immersion Liquid Cooling muaj cov txiaj ntsig hauv qab no:

Txhim khu kev ua kom txias txias: Los ntawm ib puag ncig tag nrho cov khoom siv nrog cov kua txias, txoj kev no muab cov cua sov kom zoo thiab tshem tawm.
Cov nqi kho mob qis dua: Tsis muaj qhov txav txav xws li cov kiv cua, muaj cov khoom siv tsawg dua, uas txo cov nqi tu thiab ncua lub neej ntawm cov khoom siv.
Txhim kho zog Efficiency: Kev tso dej txias tuaj yeem txo qis kev siv hluav taws xob ntau, vim nws tshem tawm qhov xav tau ntawm cov tshuab cua txias thiab lwm yam khoom siv cua txias.

Immersion Liquid Cooling system submerges servers in cooling liquid, offering efficient heat removal for AI hardware.

▲ Immersion Liquid Cooling System

Ntxiv mus,Immersion Liquid CoolingNws yog qhov muaj peev xwm loj heev, ua rau nws zoo tagnrho rau cov chaw khaws ntaub ntawv tuav AI workloads uas tsim kom muaj cua sov ntau, xws li cov uas siv los ntawm Nvidia's Blackwell architecture. Nrog rau cov qauv AI loj dua los ua ntau dua, immersion txias yuav dhau los ua kev daws teeb meem rau cov chaw zov me nyuam saib xyuas kom ntsuas lawv cov haujlwm thaum txo nqi hluav taws xob thiab kev cuam tshuam ib puag ncig.

Silicon Valley Kev nqis peev hauv Cov Cuab Yeej Cuab Yeej

Qhov kev thov nce ntxiv rau cov kev daws teeb meem ua kom txias dua hauv AI cov chaw zov me nyuam tau txais kev nyiam ntawm cov tuam txhab lag luam peev, tshwj xeeb hauvSilicon Valley. Cov tuam txhab no nquag nqis peev hauv kev pib ua lag luam uas tshwj xeeb hauvua kua txiasthev naus laus zis, lees paub tias kev tsim kho tshiab hauv daim teb no yog qhov tseem ceeb rau yav tom ntej ntawm AI kho vajtse.

Startups tsimadvanced cooling dawsTsis yog tsuas yog muab cov txiaj ntsig tam sim no rau cov tshuab AI tiam tam sim no tab sis kuj tseem ua haujlwm rau lub nthwv dej tom ntej ntawm AI kho vajtse, uas yuav xav tau ntau txoj kev ua kom txias dua. Cov thev naus laus zis no yuav tsum tau daws ob qhov kev xav tau ntawm lub zog siab thiab cov teeb meem thermal ntawm AI systems, ua rau lawv muaj peev xwm txaus nyiam rau cov uas tab tom nrhiav thawb cov ciam teb ntawm qhov ua tau zoo hauv kev ua haujlwm siab.

Silicon Valley venture capital firms are investing heavily in startups developing innovative AI cooling technologies.

▲ AI Data Center Investment Trends

Xaus

Raws li AI kho vajtse txuas ntxiv hloov zuj zus, qhov kev thov rau cov thev naus laus zis txias txias zuj zus. Nvidia covBlackwell architectureyog nyob rau hauv pem hauv ntej ntawm qhov kev hloov no, sivdej sov Direct rau Chipcua txias kom boost efficiency thiab txo cov nqi khiav lag luam.Ua kua txias, txawm hais tias los ntawm kev sib cuag ncaj qha lossis raus dej, yog qhov ua pov thawj tias yog txoj hauv kev zoo tshaj plaws los tswj cov kev siv hluav taws xob loj thiab cov khoom siv hluav taws xob ntawm cov tshuab AI niaj hnub no.

Cov tuam txhab lag luam peev tau txais kev ceeb toom, nrog ntau qhov kev nqis peev hauv kev pib ua lag luam uas tuaj yeem muab cov kev daws teeb meem rau lwm tiam. Raws li AI cov ntaub ntawv chaw ua haujlwm loj dua thiab nyuaj, qhov tseem ceeb ntawm cov tshuab ua kom txias thiab ua kom muaj zog tuaj yeem tsuas yog nce ntxiv, ua kom cov kua dej txias yog lub hauv paus ntawm kev ua tau zoo hauv kev suav yav tom ntej.