comparing energy and latency of asynchronous and synchronous

NoC issues

2010 Fourth ACM/IEEE International Symposium on Networks-on-Chip

ComparingEnergyandLatencyofAsynchronousandSynchronousNoCsforEmbeddedSoCs

DanielGebhardt

SchoolofComputingUniversityofUtahgebhardt@cs.utah.edu

JunbokYouKennethS.Stevens

Dept.ofElectricalandComputerEngineering

UniversityofUtah

{jyou,kstevens}@ece.utah.edu

Abstract—Powerconsumptionofon-chipinterconnectsisaprimaryconcernformanyembeddedsystem-on-chip(SoC)appli-cations.Inthispaper,wecompareenergyandperformancechar-acteristicsofasynchronous(clockless)andsynchronousnetwork-on-chipimplementations,optimizedforanumberofSoCdesigns.WeadaptedtheCOSI-2.0frameworkwithORION2.0routerandwiremodelsforsynchronousnetworkgeneration.Ourowntool,ANetGen,speci estheasynchronousnetworkbydeterminingthetopologywithsimulated-annealingandrouterlocationswithforce-directedplacement.Itusesenergyanddelaymodelsfromour65nmbundled-datarouterdesign.SystemCsimulationsvariedtraf cburstinessusingtheself-similarb-model.Resultsshowthattheasynchronousnetworkprovidedlowermedianandmaximummessagelatency,especiallyunderburstytraf c,andusedfarlessrouterenergywithaslightoverheadfortheinter-routerwires.

I.INTRODUCTION

Embedded,energy-constrainedSoCdesignscanberoughlyseparatedintotwoclasses:platform-basedand xed-function(alsocalledapplication-speci c).Theformerisconcernedwithbeingabletoperformawidevarietyoftasks,manyofwhichcannotbeforeseenatdesigntime.Thelatteristargetedtowardsaparticularfunction,orafewfunctions,thathaveknownproperties.A xed-functiondesignmightconsistofanumberofhighlyspecializedcoresandmemories,andfewergeneral-purposeprocessors.Thenetwork-on-chip(NoC)ofboththeseclassesshouldbeoptimizedforminimalenergyusagewhilemeetingthepredictedperformancerequirements;however,theapplication-speci cNoCmaybemorespecializedasithasaprioriknowledgeofthecommunicationpatternsbetweencores.Thisisincontrasttogeneral-purposeinterconnectsthatareoftenevaluatedwithtraf cpatternssuchasspatially-uniform,bit-transpose,etc..Thedomainofthisworkisthe xed-function,ratherthantheplatform-basedSoC.

Someglobally-asynchronouslocally-synchronous(GALS)interconnectsolutionsrelyonaclock,eitherwithstandardsynchronousclockdistribution,oramesochronousmethod.However,anasynchronous(alsocalledclockless)networkhasanumberofpotentialadvantagesoveraclockednetworkinaGALSenvironment.Standardargumentsforasynchronouscir-cuitdesignincluderobustnesstoprocess/voltage/temperaturevariation,average-caseinsteadofworst-caseperformance,andothersuchpoints.However,therearealsomanyNoC-speci carguments.InasynchronousNoC,theclocktreeforallroutersandpipelinebufferscanconsumesigni cantpowerasshownin

aheterogeneousnetwork[1],andinalargeCMP(chipmulti-processor)33%ofrouterpower[2].ManySoCdesignshavequiteburstyand“reactive”traf c.Inthiscase,asynchronousmethodsarebene cialinthattheyconsumelittledynamicpowerduringperiodsoflowtraf cwithoutrelyingonclockgatingtechniques.

Availablebandwidthoneachasynchronouslinkcanbeindependentlyset,tosomeextent,bywirelengthbetweenrouters,linkpipelinedepth,orbyvaryingthephysicalwireproperties(metallayer,width,andspacing).Thisispotentiallyusefulwhenbandwidthrequirementsoncore-to-corepathsvaryconsiderably.Thisisincontrasttoclockednetworkswhichcommonlyuseasinglefrequencyforallroutersandiswastefultothosepathsnotrequiringhighbandwidth.AclockedNoCcanusediscrete“islands”ofdifferingclockspeedstoachieveasimilareffect,butinamuchcoarser-grainedfashion.

Designautomationtechniquesarecommonlyusedtogen-erateaNoCforaspeci cSoCdesign.Thesemethodscandecreasetimeofdevelopmentincommercialproductsorallowaresearchertoexplorealargerdesignspace.TheNoCsolutionischosenbasedonsomemetric,usuallyafunctionofenergyandperformance.Intheoptimizationprocess,potentialsolutionsmustbeevaluatedforquality,andthisoftenrequiresanabstractedmodeloftheSoCcharacteristics.

ThisabstractioncanbedoneatavarietyoflevelsdependinguponcompletenessoravailabilityoftheSoCdesignandNoCcomponents.Ideally,onecouldsimulatetheexactfunctionalityofthevariouscorescomposingthedesign,andtheNoCwouldbefullyimplementedtomodelthecommunication.Unfortu-nately,thismethodislaborandsimulation-timeintensive,andnotagoodchoiceforearly-explorationoftheNoCdesignspace.Asusual,tradeoffsmustbemadeasfunctionbecomesmoreabstracted.

Acommonlyusedabstractionusedintheliteraturehasbeentitledacommunicationtracegraph[3](CTG)oracoregraph.Apathdescribespairsofsourceanddestinationcores,andtheparticularlinksandroutersapackettraverses.TheCTGhasan-tupleofvaluesperpath,butoftenincludesaverageexpectedtraf crateperpathandsometimesalatencyrequirementofapacket.AnexampleCTGisshowninFigure1thatweuseinourevaluation.

Toourknowledge,theredoesnotexistpublishedmethodstoaidinautomatinghigh-levelasynchronousNoCoptimization

NoC issues

Figure1:ExampleCTGgraph.EdgeweightsareinMBytes/s.for xed-functionSoCs.Thisisincontrasttotechniquesthatutilizesynchronoustoolsforimplementingaspeci cnetwork[4].Inthiswork,wegiveanoverviewofourcircuitsanddesignautomationtechniques,andcomparetheresultingasynchronousNoCtoasynchronousonegeneratedbyanexistingtool.Wealsoshowthataddingameasureofburstytraf ctoaCTGdesignabstractionleadstoamoreconclusiveNoCevaluation.AlsouniqueisourSystemCsimulatorthatmodelsasynchronousrouters,andimportantly,thelinkdelayasafunctionofwirelengthbetweenrouters.

Thispaperisorganizedasfollows.SectionIIgivesanoverviewofrelatedwork.Thesynchronousnetworkgenera-tionframeworkisdiscussedinSectionIII.OurasynchronousrouterisdiscussedinSectionIVatacircuitlevel.SectionVdescribesourmethodologyforasynchronousNoCgenerationandsimulation.OurevaluationmethodologyandsetupisgiveninSectionVI,withtheresultsdiscussedinSectionVII.WeconcludeinSectionVIII.

II.RELATEDWORK

TheCOSIframework[5]generatesanapplication-speci cNoCand oorplan,takingasinputsuchconstraintsascoreareasandaveragebandwidthbetweencores.Whileitisexten-siblewithnewalgorithmsandcomponents,itdoesnotconsiderasynchro …… 此处隐藏：35852字，全部文档内容请下载后查看。喜欢就下载吧 ……