文献翻译——微机发展简史

第1页外文文献资料WirelesssensornetworksThefirststoredprogramcomputersbegantoworkaround1950.TheonewebuiltinCambridge,theEDSACwasfirstusedinthesummerof1949.Theseearlyexperimentalcomputerswerebuiltbypeoplelikemyselfwithvaryingbackgrounds.Weallhadextensiveexperienceinelectronicengineeringandwereconfidentthatthatexperiencewouldstandusingoodstead.Thisprovedtrue,althoughwehadsomenewthingstolearn.Themostimportantofthesewasthattransientsmustbetreatedcorrectly;whatwouldcauseaharmlessflashonthescreenofatelevisionsetcouldleadtoaseriouserrorinacomputer.Asfarascomputingcircuitswereconcerned,wefoundourselveswithanembarassderichess.Forexample,wecouldusevacuumtubediodesforgatesaswedidintheEDSACorpentodeswithcontrolsignalsonbothgrids,asystemwidelyusedelsewhere.Thissortofchoicepersistedandthetermfamiliesoflogiccameintouse.ThosewhohaveworkedinthecomputerfieldwillrememberTTL,ECLandCMOS.Ofthese,CMOShasnowbecomedominant.Inthoseearlyyears,theIEEwasstilldominatedbypowerengineeringandwehadtofightanumberofmajorbattlesinordertogetradioengineeringalongwiththerapidlydevelopingsubjectofelectronics.dubbedintheIEEperlyrecognisedasanactivityinitsownright.Irememberthatwehadsomedifficultyinorganisingaconferencebecausethepowerengineerswaysofdoingthingswerenotourways.AminorsourceofirritationwasthatallIEEpublishedpaperswereexpectedtostartwithalengthystatementofearlierpractice,somethingdifficulttodowhentherewasnoearlierpracticeConsolidationinthe1960sBythelate50sorearly1960s,theheroicpioneeringstagewasoverandthecomputerfieldwasstartingupinrealearnest.Thenumberofcomputers第2页intheworldhadincreasedandtheyweremuchmorereliablethantheveryearlyones.Tothoseyearswecanascribethefirststepsinhighlevellanguagesandthefirstoperatingsystems.Experimentaltime-sharingwasbeginning,andultimatelycomputergraphicswastocomealong.Aboveall,transistorsbegantoreplacevacuumtubes.Thischangepresentedaformidablechallengetotheengineersoftheday.Theyhadtoforgetwhattheyknewaboutcircuitsandstartagain.Itcanonlybesaidthattheymeasuredupsuperblywelltothechallengeandthatthechangecouldnothavegonemoresmoothly.Soonitwasfoundpossibletoputmorethanonetransistoronthesamebitofsilicon,andthiswasthebeginningofintegratedcircuits.Astimewenton,asufficientlevelofintegrationwasreachedforonechiptoaccommodateenoughtransistorsforasmallnumberofgatesorflipflops.Thisledtoarangeofchipsknownasthe7400series.Thegatesandflipflopswereindependentofoneanotherandeachhaditsownpins.Theycouldbeconnectedbyoff-chipwiringtomakeacomputeroranythingelse.Thesechipsmadeanewkindofcomputerpossible.Itwascalledaminicomputer.Itwassomethinglessthatamainframe,butstillverypowerful,andmuchmoreaffordable.Insteadofhavingoneexpensivemainframeforthewholeorganisation,abusinessorauniversitywasabletohaveaminicomputerforeachmajordepartment.Beforelongminicomputersbegantospreadandbecomemorepowerful.Theworldwashungryforcomputingpowerandithadbeenveryfrustratingforindustrynottobeabletosupplyitonthescalerequiredandatareasonablecost.Minicomputerstransformedthesituation.Thefallinthecostofcomputingdidnotstartwiththeminicomputer;ithadalwaysbeenthatway.ThiswaswhatImeantwhenIreferredinmyabstracttoinflationinthecomputerindustrygoingtheotherway.Astimegoesonpeoplegetmorefortheirmoney,notless.ResearchinComputerHardware.ThetimethatIamdescribingwasawonderfuloneforresearchincomputerhardware.Theuserofthe7400seriescouldworkatthegateandflip-floplevelandyettheoveralllevelofintegrationwassufficientto第3页giveadegreeofreliabilityfarabovethatofdiscreettransistors.Theresearcher,inauniversityorelsewhere,couldbuildanydigitaldevicethatafertileimaginationcouldconjureup.IntheComputerLaboratorywebuilttheCambridgeCAP,afull-scaleminicomputerwithfancycapabilitylogic.The7400serieswasstillgoingstronginthemid1970sandwasusedfortheCambridgeRing,apioneeringwide-bandlocalareanetwork.PublicationofthedesignstudyfortheRingcamejustbeforetheannouncementoftheEthernet.Untilthesetwosystemsappeared,usershadmostlybeencontentwithteletype-basedlocalareanetworks.Ringsneedhighreliabilitybecause,asthepulsesgorepeatedlyroundthering,theymustbecontinuallyamplifiedandregenerated.Itwasthehighreliabilityprovidedbythe7400seriesofchipsthatgaveusthecourageneededtoembarkontheprojectfortheCambridgeRing.TheRISCMovementandItsAftermathEarlycomputershadsimpleinstructionsets.Astimewentondesignersofcommerciallyavailablemachinesaddedadditionalfeatureswhichtheythoughtwouldimproveperformance.Fewcomparativemeasurementsweredoneandonthewholethechoiceoffeaturesdependeduponthedesignersintuition.In1980,theRISCmovementthatwastochangeallthisbrokeontheworld.ThemovementopenedwithapaperbyPattersonandDitzelentitledTheCasefortheReducedInstructionsSetComputer.Apartfromleadingtoastrikingacronym,thistitleconveyslittleoftheinsightsintoinstructionsetdesignwhichwentwiththeRISCmovement,inparticularthewayitfacilitatedpipelining,asystemwherebyseveralinstructionsmaybeindifferentstagesofexecutionwithintheprocessoratthesametime.Pipeliningwasnotnew,butitwasnewforsmallcomputersTheRISCmovementbenefitedgreatlyfrommethodswhichhadrecentlybecomeavailableforestimatingtheperformancetobeexpectedfromacomputerdesignwithoutactuallyimplementingit.Irefertotheuseofapowerfulexistingcomputertosimulatethenewdesign.Bytheuseofsimulation,RISCadvocateswereabletopredictwithsomeconfidencethatagoodRISCdesignwouldbeabletoout-performthebestconventionalcomputersusingthesamecircuittechnology.Thispredictionwasultimatelybornoutin第4页practice.Simulationmaderapidprogressandsooncameintouniversalusebycomputerdesigners.Inconsequence,computerdesignhasbecomemoreofascienceandlessofanart.Today,designersexpecttohavearoomfulof,computersavailabletodotheirsimulations,notjustone.Theyrefertosucharoomfulbytheattractivenameofcomputerfarm.Thex86InstructionSetLittleisnowheardofpre-RISCinstructionsetswithonemajorexception,namelythatoftheIntel8086anditsprogeny,collectivelyreferredtoasx86.ThishasbecomethedominantinstructionsetandtheRISCinstructionsetsthatoriginallyhadaconsiderablemeasureofsuccessarehavingtoputupahardfightforsurvival.Thisdominanceofx86disappointspeoplelikemyselfwhocomefromtheresearchwings.bothacademicandindustrial.ofthecomputerfield.Nodoubt,businessconsiderationshavealottodowiththesurvivalofx86,butthereareotherreasonsaswell.Howevermuchweresearchorientedpeoplewouldliketothinkotherwise.highlevellanguageshavenotyeteliminatedtheuseofmachinecodealtogether.Weneedtokeepremindingourselvesthatthereismuchtobesaidforstrictbinarycompatibilitywithprevioususagewhenthatcanbeattained.Nevertheless,thingsmighthavebeendifferentifIntelsmajorattempttoproduceagoodRISCchiphadbeenmoresuccessful.Iamreferringtothei860(notthei960,whichwassomethingdifferent).Inmanywaysthei860wasanexcellentchip,butitssoftwareinterfacedidnotfitittobeusedinaworkstation.Thereisaninterestingstinginthetailofthisapparentlyeasytriumphofthex86instructionset.ItprovedimpossibletomatchthesteadilyincreasingspeedofRISCprocessorsbydirectimplementationofthex86instructionsetashadbeendoneinthepast.Instead,designerstookaleafoutoftheRISCbook;althoughitisnotobvious,onthesurface,amodernx86processorchipcontainshiddenwithinitaRISC-styleprocessorwithitsowninternalRISCcoding.Theincomingx86codeis,aftersuitablemassaging,convertedintothisinternalcodeandhandedovertotheRISCprocessorwherethecriticalexecutionisperformed.第5页InthissummingupoftheRISCmovement,IrelyheavilyonthelatesteditionofHennessyandPattersonsbooksoncomputerdesignasmysupportingauthority;seeinparticularComputerArchitecture,thirdedition,2003,pp146,151-4,157-8.TheIA-64instructionset.Sometimeago,IntelandHewlett-PackardintroducedtheIA-64instructionset.Thiswasprimarilyintendedtomeetagenerallyrecognisedneedfora64bitaddressspace.Inthis,itfollowedtheleadofthedesignersoftheMIPSR4000andAlpha.HoweveronewouldhavethoughtthatIntelwouldhavestressedcompatibilitywiththex86;thepuzzleisthattheydidtheexactopposite.Moreover,builtintothedesignofIA-64isafeatureknownaspredicationwhichmakesitincompatibleinamajorwaywithallotherinstructionsets.Inparticular,itneeds6extrabitswitheachinstruction.Thisupsetsthetraditionalbalancebetweeninstructionwordlengthandinformationcontent,anditchangessignificantlythebriefofthecompilerwriter.Inspiteofhavinganentirelynewinstructionset,IntelmadethepuzzlingclaimthatchipsbasedonIA-64wouldbecompatiblewithearlierx86chips.Itwashardtoseeexactlywhatwasmeant.ChipsforthelatestIA-64processor,namely,theItanium,appeartohavespecialhardwareforcompatibility.Evenso,x86coderunsveryslowly.Becauseoftheabovecomplications,implementationofIA-64requiresalargerchipthanisrequiredformoreconventionalinstructionsets.Thisinturnimpliesahighercost.Suchatanyrate,isthereceivedwisdom,and,asageneralprinciple,itwasrepeatedassuchbyGordonMoorewhenhevisitedCambridgerecentlytoopentheBettyandGordonMooreLibrary.Ihave,however,hearditsaidthatthematterappearsdifferentlyfromwithinIntel.ThisIdonotunderstand.ButIamveryreadytoadmitthatIamcompletelyoutofmydepthasregardstheeconomicsofthesemiconductorindustry.AMDhavedefineda64bitinstructionsetthatismorecompatiblewithx86andtheyappeartobemakingheadwaywithit.Thechipisnotaparticularlylargeone.SomepeoplethinkthatthisiswhatIntelshouldhave第6页done.Sincethelecturewasdelivered,IntelhaveannouncedthattheywillmarketarangeofchipsessentiallycompatiblewiththoseofferedbyAMD.TheRelentlessDrivetowardsSmallerTransistorsThescaleofintegrationcontinuedtoincrease.Thiswasachievedbyshrinkingtheoriginaltransistorssothatmorecouldbeputonachip.Moreover,thelawsofphysicswereonthesideofthemanufacturers.Thetransistorsalsogotfaster,simplybygettingsmaller.Itwasthereforepossibletohave,atthesametime,bothhighdensityandhighspeed.Therewasafurtheradvantage.Chipsaremadeondiscsofsilicon,knownaswafers.Eachwaferhasonitalargenumberofindividualchips,whichareprocessedtogetherandlaterseparated.Sinceshrinkagemakesitpossibletogetmorechipsonawafer,thecostperchipgoesdown.Fallingunitcostwasimportanttotheindustrybecause,ifthelatestchipsarecheapertomakeaswellasfaster,thereisnoreasontogoonofferingtheoldones,atleastnotindefinitely.Therecanthusbeoneproductfortheentiremarket.However,detailedcostcalculationsshowedthat,inordertomaintainthisadvantageasshrinkageproceededbeyondacertainpoint,itwouldbenecessarytomovetolargerwafers.Theincreaseinthesizeofwaferswasnosmallmatter.Originally,waferswereoneortwoinchesindiameter,andby2000theywereasmuchastwelveinches.Atfirst,itpuzzledmethat,whenshrinkagepresentedsomanyotherproblems,theindustryshouldmakethingsharderforitselfbygoingtolargerwafers.Inowseethatreducingunitcostwasjustasimportanttotheindustryasincreasingthenumberoftransistorsonachip,andthatthisjustifiedtheadditionalinvestmentinfoundriesandtheincreasedrisk.Thedegreeofintegrationismeasuredbythefeaturesize,which,foragiventechnology,isbestdefinedasthehalfthedistancebetweenwiresinthedensestchipsmadeinthattechnology.Atthepresenttime,productionof90nmchipsisstillbuildingupSuspensionofLawInMarch1997,GordonMoorewasaguestspeakeratthecelebrationsofthecentenaryofthediscoveryoftheelectronheldattheCavendish第7页Laboratory.ItwasduringthecourseofhislecturethatIfirstheardthefactthatyoucanhavesiliconchipsthatarebothfastandlowincostdescribedasaviolationofMurphyslaw.orSodslawasitisusuallycalledintheUK.Mooresaidthatexperienceinotherfieldswouldleadyoutoexpecttohavetochoosebetweenspeedandcost,ortocompromisebetweenthem.Infact,inthecaseofsiliconchips,itispossibletohaveboth.Inareferencebookavailableontheweb,MurphyisidentifiedasanengineerworkingonhumanaccelerationtestsfortheUSAirForcein1949.However,wewereperfectlyfamiliarwiththelawinmystudentdays,whenwecalleditbyamuchmoreprosaicnamethaneitherofthosementionedabove,namely,theLawofGeneralCussedness.Weevenhadamockexaminationquestioninwhichthelawfeatured.Itwasthetypeofquestioninwhichthefirstpartasksforadefinitionofsomelaworprincipleandthesecondpartcontainsaproblemtobesolvedwiththeaidofit.InourcasethefirstpartwastodefinetheLawofGeneralCussednessandthesecondwastheproblem;Acyclistsetsoutonacircularcyclingtour.Deriveanequationgivingthedirectionofthewindatanytime.Thesingle-chipcomputerAteachshrinkagethenumberofchipswasreducedandtherewerefewerwiresgoingfromonechiptoanother.Thisledtoanadditionalincrementinoverallspeed,sincethetransmissionofsignalsfromonechiptoanothertakesalongtime.Eventually,shrinkageproceededtothepointatwhichthewholeprocessorexceptforthecachescouldbeputononechip.Thisenabledaworkstationtobebuiltthatout-performedthefastestminicomputeroftheday,andtheresultwastokilltheminicomputerstonedead.Asweallknow,thishadsevereconsequencesforthecomputerindustryandforthepeopleworkinginit.FromtheabovetimethehighdensityCMOSsiliconchipwasCockoftheRoost.Shrinkagewentonuntilmillionsoftransistorscouldbeputonasinglechipandthespeedwentupinproportion.Processordesignersbegantoexperimentwithnewarchitecturalfeaturesdesignedtogiveextraspeed.Oneverysuccessfulexperimentconcerned第8页methodsforpredictingthewayprogrambrancheswouldgo.Itwasasurprisetomehowsuccessfulthiswas.ItledtoasignificantspeedingupofprogramexecutionandotherformsofpredictionfollowedEquallysurprisingiswhatithasbeenfoundpossibletoputonasinglechipcomputerbywayofadvancedfeatures.Forexample,featuresthathadbeendevelopedfortheIBMModel91.thegiantcomputeratthetopoftheSystem360range.arenowtobefoundonmicrocomputersMurphysLawremainedinastateofsuspension.Nolongerdiditmakesensetobuildexperimentalcomputersoutofchipswithasmallscaleofintegration,suchasthatprovidedbythe7400series.Peoplewhowantedtodohardwareresearchatthecircuitlevelhadnooptionbuttodesignchipsandseekforwaystogetthemmade.Foratime,thiswaspossible,ifnoteasyUnfortunately,therehassincebeenadramaticincreaseinthecostofmakingchips,mainlybecauseoftheincreasedcostofmakingmasksforlithography,aphotographicprocessusedinthemanufactureofchips.Ithas,inconsequence,againbecomeverydifficulttofinancethemakingofresearchchips,andthisisacurrentlycauseforsomeconcern.TheSemiconductorRoadMapTheextensiveresearchanddevelopmentworkunderlyingtheaboveadvanceshasbeenmadepossiblebyaremarkablecooperativeeffortonthepartoftheinternationalsemiconductorindustry.AtonetimeUSmonopolylawswouldprobablyhavemadeitillegalforUScompaniestoparticipateinsuchaneffort.Howeverabout1980significantandfarreachingchangestookplaceinthelaws.Theconceptofpre-competitiveresearchwasintroduced.Companiescannowcollaborateatthepre-competitivestageandlatergoontodevelopproductsoftheirownintheregularcompetitivemanner.Theagentbywhichthepre-competitiveresearchinthesemi-conductorindustryismanagedisknownastheSemiconductorIndustryAssociation(SIA).ThishasbeenactiveasaUSorganisationsince1992anditbecameinternationalin1998.Membershipisopentoanyorganisationthatcancontributetotheresearcheffort.第9页EverytwoyearsSIAproducesanewversionofadocumentknownastheInternationalTechnologicalRoadmapforSemiconductors(ITRS),withanupdateintheintermediateyears.ThefirstvolumebearingthetitleRoadmapwasissuedin1994buttworeports,writtenin1992anddistributedin1993,areregardedasthetruebeginningoftheseries.Successiveroadmapsaimatprovidingthebestavailableindustrialconsensusonthewaythattheindustryshouldmoveforward.Theysetoutingreatdetail.overa15yearhorizon.thetargetsthatmustbeachievedifthenumberofcomponentsonachipistobedoubledeveryeighteenmonths.thatis,ifMooreslawistobemaintained.-andifthecostperchipistofall.Inthecaseofsomeitems,thewayaheadisclear.Inothers,manufacturingproblemsareforeseenandsolutionstothemareknown,althoughnotyetfullyworkedout;theseareasarecolouredyellowinthetables.Areasforwhichproblemsareforeseen,butforwhichnomanufacturablesolutionsareknown,arecolouredred.RedareasarereferredtoasRedBrickWalls.ThetargetssetoutintheRoadmapshaveprovedrealisticaswellaschallenging,andtheprogressoftheindustryasawholehasfollowedtheRoadmapsclosely.Thisisaremarkableachievementanditmaybesaidthatthemeritsofcooperationandcompetitionhavebeencombinedinanadmirablemanner.Itistobenotedthatthemajorstrategicdecisionsaffectingtheprogressoftheindustryhavebeentakenatthepre-competitivelevelinrelativeopenness,ratherthanbehindcloseddoors.Theseincludetheprogressiontolargerwafers.By1995,Ihadbeguntowonderexactlywhatwouldhappenwhentheinevitablepointwasreachedatwhichitbecameimpossibletomaketransistorsanysmaller.MyenquiriesledmetovisitARPAheadquartersinWashingtonDC,whereIwasgivenacopyoftherecentlyproducedRoadmapfor1994.Thismadeitplainthatseriousproblemswouldarisewhenafeaturesizeof100nmwasreached,aneventprojectedtohappenin2007,with70nmfollowingin2010.Theyearforwhichthecomingof100nm(orrather90nm)wasprojectedwasinlaterRoadmapsmovedforwardto2004andintheevent第10页theindustrygottherealittlesooner.Ipresentedtheaboveinformationfromthe1994Roadmap,alongwithsuchotherinformationthatIcouldobtain,inalecturetotheIEEinLondon,entitledTheCMOSend-pointandrelatedtopicsinComputinganddeliveredon8February1996.TheideathatIthenhadwasthattheendwouldbeadirectconsequenceofthenumberofelectronsavailabletorepresentaonebeingreducedfromthousandstoafewhundred.Atthispointstatisticalfluctuationswouldbecometroublesome,andthereafterthecircuitswouldeitherfailtowork,oriftheydidworkwouldnotbeanyfaster.Infactthephysicallimitationsthatarenowbeginningtomakethemselvesfeltdonotarisethroughshortageofelectrons,butbecausetheinsulatinglayersonthechiphavebecomesothinthatleakageduetoquantummechanicaltunnellinghasbecometroublesome.Therearemanyproblemsfacingthechipmanufacturerotherthanthosethatarisefromfundamentalphysics,especiallyproblemswithlithography.Inanupdatetothe2001Roadmappublishedin2002,itwasstatedthatthecontinuationofprogressatpresentratewillbeatriskasweapproach2005whentheroadmapprojectsthatprogresswillstallwithoutresearchbreak-throughsinmosttechnicalareas“.ThiswasthemostspecificstatementabouttheRedBrickWall,thathadsofarcomefromtheSIAanditwasastrongone.The2003Roadmapreinforcesthisstatementbyshowingmanyareasmarkedred,indicatingtheexistenceofproblemsforwhichnomanufacturablesolutionsareknown.Itissatisfactorytoreportthat,sofar,timelysolutionshavebeenfoundtoalltheproblemsencountered.TheRoadmapisaremarkabledocumentand,forallitsfranknessabouttheproblemsloomingabove,itradiatesimmenseconfidence.Prevailingopinionreflectsthatconfidenceandthereisageneralexpectationthat,byonemeansoranother,shrinkagewillcontinue,perha