台积电 x 英伟达：突破热壁：先进冷却技术如何驱动未来计算-2025.pdfVIP

TSMCxNvidia:BreakingtheThermalWall

HowAdvancedCoolingIsPoweringthe

FutureofComputing

台积电xNvidia：突破热壁：先进冷却技术如何驱

动未来计算

OriginalArticlebySemiVisionResearch(TSMC,NVIDIA,Coherent,Jentech,AVC,

Auras,CoolerMaster,MikrosTechnologies,Inspur,Invek,NingboJingda,

Fabric8Labs,Intel,xMEMS)

SEMIVISION

OCT05,2025·PAID

175Share

ThedevelopmentofAIchipsisfundamentallydrivenbythepursuitofhigher

performance,butthisalsobringsacriticalchallenge:thermalmanagement.

Effectivelydissipatingtheenormousheatgeneratedbyrisingpower

consumptionhasbecomeakeydesignconsiderationfornext-generationAI

chiparchitectures.

Onthesystemlevel,theintegrationofopticalenginesiswidelyviewedasthe

futuresolutiontoovercomehigh-bandwidthinterconnectbottlenecks.

However,opticalsignalsarehighlysensitivetotemperaturefluctuations—even

slightdeviationscanleadtotransmissionlossanddegradedperformance.This

iswhyTSMChasadoptedtheMicroringModulator(MRM)approach,butMRM

itselfhasstrictthermaloperatingrequirements.Asaresult,companieslike

TSMCandNVIDIAareactivelyexploringnext-generationcoolingtechnologies

tomeetthesethermaldemands.

Onthelogicprocessside,TSMC’sevolutionfromN3→N2→A16involves

morethangeometricscaling—itrepresentsatransistorarchitecturetransition

fromFinFET,toGAA(Gate-All-Around),andfinallytotheSuperPowerRail

architectureintheA16generation.Thisseriesofchangesisaimedatachieving

theoptimalbalanceofPPA(Power,Performance,andArea),layingthe

foundationforhigh-efficiencyAIcomputing.

However,fromapackagingandmaterialsperspective,newchallengesemerge.

TraditionalcoolingdependsonTIM(ThermalInterfaceMaterials),whose

materialselectionislimited,andthermalreliabilityisconstrainedwithinspecific