interim report for the decadal plan for semiconductors€¦ · decada an interim report....
TRANSCRIPT
Semiconductor Research CorporationSemiconductor Industry Association
Decadal Plan for SemiconductorsI N T E R I M R E P O R T F O R T H E
Acronym DefinitionsAI ArtificialIntelligence
aJ attojoule(10-18joules)
bps bit-per-second
CPU CentralProcessingUnit
CMOS ComplementaryMetal-Oxide-Semiconductor
DNA DeoxyriboNucleicAcid
DRAM Dynamic Random-Access Memory
FPGA Field-ProgrammableGateArray
GPU GraphicsProcessingUnit
IoT InternetofThings
ICT InformationandCommunicationTechnologies
I/O Input/Output
IPblock semiconductorIntellectualPropertycore
HW Hardware
NANDflash highest-densitysilicon-basedelectronicnonvolatilememory
Mbps megabit-per-second
MIMO Multiple-InputandMultiple-Output
mm-Wave millimeterwave
nJ nanojoule(10-9joules)
NVM NonvolatileMemory
R&D ResearchandDevelopment
SIA Semiconductor Industry Association
SRC Semiconductor Research Corporation
SW Software
Tbps terabit-per-second
THz Terahertz
Zettabyte 1021 bytes
ZIPS 1021 compute instructions per second
Decadal Plan Executive CommitteeJamesAng,PNNL
DmytroApalkov,Samsung
FariAssaderaghi, Sunrise Memory
RalphCavin,SRC
RameshChauhan,Qualcomm
AnChen,IBM
RichardChow,Intel
RobertClark,TEL
MaryamCope,SIA
DebraDelise, AnalogDevices
CarlosDiaz,TSMC
BobDoering, TexasInstruments
SeanEilert,Micron
KenHansen,SRC
BaherHaroun, TexasInstruments
Yeon-CheolHeo,Samsung
GilbertHerrera,SNL
KevinKemp,NXP
TaffyKingscott,IBM
StephenKosonocky,AMD
MatthewKlusas,Amazon
SteveKramer,Micron
DonnyKwak,Samsung
LawrenceLoh,MediaTek
RaficMakki,Mubadala
MatthewMarinella,SNL
Seong-HoPark,SKhynix
DavidPellerin,Amazon
DanielRasic,SRC
BenRathsak,TEL
WallyRhines, Mentor Graphics
HeikeRiel,IBM
KirillRivkin,WesternDigital
GurtejSandhu,Micron
GhavamShahidi,IBM
SteveSon,SKhynix
MarkSomervell,TEL
GilroyVandentop,Intel
JeffreyVetter,ORNL
JeffreyWelser,IBM
JimWieser, TexasInstruments
IanYoung,Intel
DavidYeh,TI&SRC
VictorZhirnov,SRC
ZoranZvonar, AnalogDevices
The Decadal Plan Workshops that helped drive this report were supported by the U.S. Department of Energy, Advanced
Scientific Computing Research (ASCR) and Basic Energy Sciences (BES) Research Programs, and the National Nuclear
Security Agency, Advanced Simulation and Computing (ASC) Program. SIA and SRC are grateful for their support.
1
Decadal Plan Interim Report
Semiconductors, the tiny and highly advanced chips that power modern electronics, have helped give rise to the
greatest period of technological advancement in the history of humankind.
Chip-enabled technology now allows us to analyze DNA sequences to treat disease, model nerve synapses in
the brain to help people with mental disorders like Alzheimer’s, design and build safer and more reliable cars
and passenger jets, improve the energy efficiency of buildings, and perform countless other tasks that improve
people’s lives.
During the COVID-19 pandemic, the world has come to rely more heavily on semiconductor-enabled technology to
work, study, communicate, treat illness, and do innumerable other tasks remotely. And the future holds boundless
potential for semiconductor technology, with emerging applications such as artificial intelligence, quantum
computing, and advanced wireless technologies like 5G and 6G promising incalculable benefits to society.
Fulfilling that promise, however, will require taking action to address a range of seismic shifts shaping the
future of chip technology. These seismic shifts—identified in The Decadal Plan for Semiconductors by a broad
cross-section of leaders in academia, government, and industry—involve smart sensing, memory and storage,
communication, security, and energy efficiency. The federal government, in partnership with private industry,
must invest ambitiously in semiconductor research in these areas to sustain the future of chip innovation.
For decades, federal government and private sector investments in semiconductor research and development
(R&D) have propelled the rapid pace of innovation in the U.S. semiconductor industry, making it the global
leader and spurring tremendous growth throughout the U.S. economy. The U.S. semiconductor industry invests
about one-fifth of its revenues each year in R&D, one of the highest shares of any industry. With America
facing increasing competition from abroad and mounting costs and challenges associated with maintaining the
breakneck pace of innovation, now is the time to maintain and strengthen public-private research partnerships.
As Congress works to refocus America’s research ecosystem on maintaining semiconductor innovation and
competitiveness, The Decadal Plan for Semiconductors outlines semiconductor research priorities across
the seismic shifts noted above and recommends an additional federal investment of $3.4 billion annually
across these five areas. The interim report is included here, and the full report is scheduled to be released in
December 2020.
Working together, we can boost semiconductor technology and keep it strong, competitive, and at the tip of
the innovation spear.
JohnNeuffer
President&CEO
SemiconductorIndustryAssociation(SIA)
ToddYounkin
President&CEO
SemiconductorResearchCorporation(SRC)
Sincerely,
2
Decadal Plan Interim Report
Executive Summary
necessary generational improvements in the energy-efficiency with
which information is processed, communicated, stored, sensed
and actuated on. Long term sustainable ICT growth will rely on
breakthroughs in semiconductor technology capabilities that
enable holistic solutions to tackle information processing efficiency.
Disruptive breakthroughs are needed in the areas of software,
systems, architectures, circuits, device structure and the related
processes and materials that require timely and well-coordinated
multidisciplinary research efforts.
This Decadal Plan for Semiconductors outlines research priorities in
information processing, sensing, communication, storage, and security
seeking to ensure sustainable growth for semiconductor and ICT
industries by:
• informing and supporting the strategic visions of semiconductor
companies and government agencies
• guiding a (r)evolution of cooperative academic, industry and
government research programs
• placing ‘a stake in the ground’ to challenge the best and brightest
researchers, university faculty and students
The U.S. semiconductor industry leads the
world in innovation, based in large part on
aggressive research and development (R&D)
spending. The industry invests nearly one-
fifth of its annual revenue in R&D each year,
second only to the pharmaceuticals sector. In
addition, Federal funding of semiconductor
R&D serves as the catalyst for private R&D
spending. Together, private and Federal
semiconductor R&D investments have
sustained the pace of innovation in the U.S.,
enabling it to become the global leader in
the semiconductor industry. Those R&D
investments have nurtured the development
of innovative and commercially viable
products, and as a direct result, have led to a
significant contribution to the U.S. economy
and jobs.
The current hardware-software (HW-SW)
paradigm in information and communication
technologies (ICT) has made computing
ubiquitous through sustained innovation
in software and algorithms, systems
architecture, circuits, devices, materials,
and semiconductor process technologies
among others. However, ICT is facing
unprecedented technological challenges
for maintaining its growth rate levels
into the next decade. These challenges
arise largely from approaching various
fundamental limitations in semiconductor
technology that taper the otherwise
FigurecourtesyofAnalogDevices,Inc.
3
Decadal Plan Interim Report
The Grand ChallengeInformationandcommunicationtechnologiesmakeupover70%ofthesemiconductormarketshare.Theycontinuetogrow
withoutboundsdominatedbytheexponentialcreationofdatathatmustbemoved,stored,computed,communicated,secured
andconvertedtoenduserinformation.Therecentexplosionofartificialintelligence(AI)applicationsisaclearexample,andas
anindustrywehaveonlybeguntoscratchthesurface.
Havingcomputingsystemsmoveintodomainswithtruecognition,i.e.,acquiringunderstandingthroughexperience,reasoning
andperceptionisanewregime.Thisregimeisunachievablewiththestate-of-the-artsemiconductortechnologiesand
traditionalgainssincethereductioninfeaturesize(i.e.,dimensionalscaling)toimproveperformanceandreducecostsin
semiconductorsisreachingitsphysicallimits.Asaresult,thecurrentparadigmmustchangetoaddressaninformationand
intelligence-basedvaluepropositionwithsemiconductortechnologiesasthedriver.
Seismic shift #1
Fundamentalbreakthroughsinanaloghardwarearerequiredtogeneratesmarterworld-machineinterfacesthatcansense,
perceive and reason . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7
Seismic shift #2
Thegrowthofmemorydemandswilloutstripglobalsiliconsupplypresentingopportunitiesforradicallynewmemoryand
storagesolutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
Seismic shift #3
Alwaysavailablecommunicationrequiresnewresearchdirectionsthataddresstheimbalanceofcommunicationcapacityvs.
datagenerationrates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
Seismic shift #4
Breakthroughsinhardwareresearchareneededtoaddressemergingsecuritychallengesinhighlyinterconnectedsystemsand
ArtificialIntelligence. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
Seismic shift #5
Everrisingenergydemandsforcomputingvs.globalenergyproductioniscreatingnewrisk,andnewcomputingparadigms
offeropportunitieswithdramaticallyimprovedenergyefficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
Trends and driversCurrentlyinformationandcommunicationtechnologiesarefacingfivemajorseismicshifts:
The Semiconductor Industry Association (SIA) June 2020 report1 demonstrates that federal investment in semiconductor
R&D spurs U.S. economic growth and job creation and presents a case for a 3x increase in semiconductor-specific federal
funding. For every dollar spent on federal semiconductor research has resulted in a $16.50 increase in current GDP.
The Decadal Plan for Semiconductors complements this report and identifies specific goals with quantitative targets.
It is expected that the Decadal Plan will have a major impact on the semiconductor industry, similar to the impact of
the 1984 10-year SRC Research Goals document that was continued in 1994 as the National Technology Roadmap for
Semiconductors, and which later became the International Technology Roadmap for Semiconductors in 1999.
[1]SparkingInnovation:HowFederalInvestmentinSemiconductorR&DSpursU.S.EconomicGrowthandJobCreation,SIAReport,June2020
4
Decadal Plan Interim Report
Call to Action: Semiconductor Technology Leadership InitiativeMaintainingandstrengtheningtheleadershipoftheUnited
StatesinICTduringthisnewsemiconductorerarequires
asustainedadditional$3.4Bfederalinvestmentperyear
throughoutthisdecade(i.e.triplingFederalfundingfor
semiconductorresearch)toconductlarge-scaleindustry-
relevant,fundamentalsemiconductorresearch.(TheDecadal
PlanExecutiveCommitteeofferedrecommendationson
allocationoftheadditional$3.4Binvestmentperyear
amongthefiveseismicshiftsidentifiedintheDecadalPlan.
Thebasisofallocationisthemarketsharetrendandour
analysisoftheR&Drequirementsfordifferentsemiconductor
andICTtechnologies).
Theinvestmentsthroughnewpublic-privatepartnerships
mustcoverawidebreadthofinterdependenttechnicalareas
(compute,analog,memory/storage,communications,and
security)requiringmulti-disciplinaryteamstomaintainU.S.
semiconductortechnologyleadership.
Theseinvestmentsneedtobeorganizedandcoordinatedto
supportacommonsetofgoalsfocusedonmarketdemandto
providetechnologieswhichenablenewcommercialproducts
andservicesoverthecourseoftheprogram.TheDecadal
Planhasidentifiedfiveseismicparadigmshiftsrequiredto
accomplishthisoverarchingGrandChallenge.
A Note on Funding Strategic National Initiatives:
TheDecadalPlanservesasablueprintforpolicymakers
whorecognizethischallengeandseekguidanceonareasof
researchemphasisforscientificresearchagenciesandpublic-
privatepartnerships.
Semiconductor Technology Breakthroughs Rely OnHolistic Optimal Solutions
Driven by Hardware/Software Co-OptimizationInterlocked Multidisciplinary Research
5
Decadal Plan Interim Report
Seismic shift #1 Fundamentalbreakthroughsinanaloghardwarearerequiredtogeneratesmarterworld-machine
interfacesthatcansense,perceiveandreason.
Seismic shift #2 Thegrowthofmemorydemandswilloutstripglobalsiliconsupplypresentingopportunitiesforradically
newmemoryandstoragesolutions.
Seismic shift #3 Alwaysavailablecommunicationrequiresnewresearchdirectionsthataddresstheimbalanceof
communicationcapacityvs.datagenerationrates.
Seismic shift #4 Breakthroughsinhardwareresearchareneededtoaddressemergingsecuritychallengesinhighly
interconnectedsystemsandArtificialIntelligence.
Seismic shift #5 Everrisingenergydemandsforcomputingvs.globalenergyproductioniscreatingnewrisk,andnew
computingparadigmsofferopportunitieswithdramaticallyimprovedenergyefficiency.
Currentlyinformationandcommunicationtechnologiesarefacingfivemajorseismicshifts:
1. Introduction
TosupporttheDecadalPlandevelopment,aninternationalseriesoffiveface-to-faceworkshopshasbeenconductedtoassess
quantitativelyeachseismicshift,assigntargetsandsuggestinitialresearchdirections.Participantsandcontributorstothese
workshopsincludedacademic,governmentandindustrialdomainexperts.Theoutputoftheseworkshopshasguidedthe
recommendationsinthe2020DecadalPlanforSemiconductors.
Theseworkshopsprovidedhighlyinteractiveforumswherekeyresearchleadersevaluatedthestatusofnanoelectronicsresearch
andapplicationdrivers,discussedkeyscientificissues,anddefinedpromisingfutureresearchdirections.Thisisinstrumentalfor
the2020versionoftheDecadalPlanforSemiconductorstoreflectaninformedviewonkeyscientificandtechnicalchallenges
relatedtorevolutionaryinformationandcommunicationtechnologies,basedonnewquantitativeanalysesandprojections.
The primary objectives of the Decadal Plan include • Identify significant trends and applications that are driving Information and Communication Technologies and the
associated roadblocks/challenges.
• Assess quantitatively the potential and status of the five seismic shifts that will impact future ICT.
• Identify fundamental goals and targets to alter the current trajectory of semiconductor technology.
The Decadal Plan provides an executive overview of the global drivers and constraints for the future ICT industry, rather
than to offer/discuss specific solutions: The document identifies the what, not the how. In doing so, it focuses and
organizes the best of our energies and skills to the key challenges in a quantitative manner about which creative
solutions can be imagined and their impact measured.
6
Decadal Plan Interim Report
Seismic shift #1Fundamental breakthroughs in analog hardware are required to generate smarter world-machine interfaces that can sense, perceive and reason.
Analogelectronicsdealswithreal-worldcontinuouslyvariablesignalsofmultipleshapes(incontrasttodigitalelectronicswhere
signalsareusuallyofstandardshapetakingonlytwolevels,onesorzeros).Theanalogelectronicsdomainencompassesmultiple
dimensionsasshowninFigure1.Also,allinputshumancanperceiveareanalog,whichcallsforbio-inspiredsolutionsforworld-machine
interfacesthatcansense,perceiveandreasonbasedonultra-compressedsensingcapabilityandlowoperationpower(Figure2).
Thephysicalworldisinherentlyanalogandthe“digitalsociety”placesanincreasingdemandforadvancedanalogelectronicsto
enableinteractionbetweenthephysicalandcomputer“worlds.”
SensingtheenvironmentaroundusisfundamentaltothenextgenerationofAIwheredeviceswillbecapableofperceptionand
reasoning.Theworld-machineinterfaceliesattheheartofthecurrent information-centric economy.Asoneexample,thenext
waveoftheadvancedmanufacturingrevolutionisexpectedtocomefromnext-generationanalog-drivenindustrialelectronics,
thatincludessensing,robotics,industrial,automotive,medicaletc.Formissioncriticalapplications,thereliabilityofelectronic
componentsisapriority.Today,forexample,analogchipsconstitute80%offailuresinautomotiveelectronics,whichisten
timesworsethandigitalchips.
Figure1:TheDimensionsofAnalogElectronics
Conscious bits100 bps
8.75 Mbps
Figure2:Thebrain’sabilitytoperceiveandreasonisbasedonultra-compressedsensingcapabilitieswith100,000datareductionandalowoperationpower
7
Decadal Plan Interim Report
Theestimatedtotalanaloginformationgeneratedfromthe
physicalworldisequivalentto~1034bit/s.Asareference,
thetotalcollectivehumansensorythroughputpalesat
~1017bit/s(Figure3).Thus,ourabilitytoperceivethe
physicalworldissignificantlylimited.Therearetremendous
opportunitiesforfutureanalogelectronicsthataugmentthe
humansensorysystem,whichisexpectedtohavesignificant
economicandsocialconsequences.Examplesinclude,butare
notlimitedto,creatingmultimediathatspecificallytargets
humansensoryandcognitivesystems,includingnervous
systeminterfacesandcommunications.Thiscanresultin
newhuman-centrictechnologiessuchasmulti-sensing-based
medicaldiagnosticsandtherapy,completevirtualrealitywith
virtualaromasynthesizers,oractiveodorcancellationbased
onindoorairquality.
Todaytheabilitytogenerateanalogdataisgrowingfaster
thanourabilitytointelligentlyusethedata.Thissituationwill
becomeevenmoreseriousinthenearfuture,whendatafrom
ourlivesaswellasfromIoTsensorsmaycreateananalog
datadelugethatwillobscurevaluableinformationwhen
weneeditthemost.Sensortechnologiesareexperiencing
exponentialgrowthwithforecastsof~45trillionsensors
in2032thatwillgenerate>1millionzettabytes(1027bytes)
ofdataperyear.Thisisequivalentto~1020bit/s,thereby
surpassing the collective human sensing throughput.
Therefore,asignificantparadigmshifttowardsextractingkey
informationinthepredicteddatadelugeandapplyingitin
anappropriatewayiskeytoharnessingthedatarevolution.
Thus,theAnalog Grand Goalisforrevolutionarytechnologies
toincreaseuseful/actionableinformationwithlessenergy
anddatabitse.g.sensing-to-analog-to-informationreduction
withapracticalcompression/reductionratioof105:1.
Formanyreal-timeapplications,thevalueofsensory
dataisbrief,sometimesonlyafewmilliseconds.Thedata
mustbeutilizedwithinthattimeframeandinmanycases
locallyforlatencyandsecurityconsiderations.Therefore,
pursuingbreakthroughadvancesininformationprocessing
technologiessuchasdevelopinghierarchicalperception
algorithmsthatenableunderstandingoftheenvironment
fromrawsensordataisafundamentalrequirement.New
computingmodelssuchasanalog“approximatecomputing”
arerequired.ThisisalignedwiththeGrandGoal#5of
discoveringaradicallynew‘computingtrajectory’,outlined
later.Newanalogtechnologiescanalsooffergreat
advancementsincommunicationtechnologies.Evenin
computer-to-computercommunication,analoginterfaces
arerequiredatlongdistances.Theabilitytocollect,process
andcommunicatetheanalogdataattheinput/output(I/O)
boundariesiscriticaltothefutureworldofIoTandBigData.
AnalogtechnologyadvancementstotheTHzregimewillbe
requiredforboththesensingandcommunicationneedsof
thefuture.
Grand Goal #1:
Analog-to-informationcompression/reductionwitha
practicalcompression/reductionratioof105:1thatdrives
toapracticaluseofinformationversus“data”inaway
thatismoreanalogoustothehumanbrain.
Figure3:Trendinworld’sinstalledsensingcapacities
45 Trillion sensors in 2032
8
Decadal Plan Interim Report
Call for ActionTheanaloginterfacebridgesthephysicalanddigitalworlds.Ourcollectiveabilitytoaccesstheinformationofthephysical
worldthroughanalogsignalsis10,000trilliontimesbelowwhatisavailable,andradicaladvancesinanalogelectronicswillbe
requiredsoon.Newapproachestosensingsuchassensing to action,analog“artificialintelligence”(AI)platforms,braininspired/
neuromorphicandhierarchicalcomputation,orothersolutionswillbenecessary.Breakthroughadvancesininformation
processingtechnologies,suchasdevelopingperceptionalgorithmstoenableunderstandingoftheenvironmentfromraw
sensordata,areafundamentalrequirement.Newcomputingmodelssuchasanalog“approximatecomputing,”whichcantrade
energyandcomputingtimewithaccuracyofoutput(presumablyhowthebraindoes)arerequired.Newanalogtechnologieswill
offergreatadvancementsincommunicationtechnologies.Theabilitytocollect,processandcommunicatetheanalogdataat
theinput/outputboundariesiscriticaltothefutureworldofIoTandBigData.Additionally,analogdevelopmentmethodologies
requireastepincrease(10xorgreater)inproductivitytoaddresstheapplicationexplosioninatimelymanner.Altogether,
collaborativeresearchtoestablishrevolutionaryparadigmsforfutureenergy-efficientanalogintegratedcircuitsforthevast
rangeoffuturedatatypes,workloadsandapplicationsisneeded.
[2]TheDecadalPlanExecutiveCommitteeofferedrecommendationsonallocationoftheadditional$3.4BinvestmentamongthefiveseismicshiftsidentifiedintheDecadalPlan.ThebasisofallocationisthemarketsharetrendandouranalysisoftheR&DrequirementsfordifferentsemiconductorandICTtechnologies.
Invest $600M annually throughout this decade in new trajectories for analog electronics. Selected priority research themes
are outlined below.2
AnalogBioinspiredMachineLearning
THzRegimeAnalog
AnalogDevelopmentMethodology
Priority Research Themes
Trainableneuromorphicsignalconverters
Goal: 100,000:1 Data Reduction
Sensing to Action goalwillnotbepossiblewithoutintegratedsystem
solutionswithsignificantincreaseindesignmethodsandproductivity.
Analog-to-DigitalAnalog-to-Information
Sensing to Action
9
Decadal Plan Interim Report
Seismic shift #2The growth of memory demands will outstrip global silicon supply presenting opportunities for radically new memory and storage solutions.
RadicalnewsolutionsinMemory
andStoragetechnologieswillbe
needed for future ICT with major
innovationsindevices,circuitsand
architectures.Byendofthisdecade,
thecontinuingimprovementsinICT
energyefficiencyandperformancewill
stallastheunderlyingmemory,and
storagetechnologieswillmeetscaling
limitations.Atthesametime,training
dataforAIapplicationsisexploding
withnolimitinsight.Itisbecoming
increasinglyclearthatinfuture
informationprocessingapplications,
synergisticinnovationsfrommaterials
and devices to circuits and system-
levelfunctions,likelyusingunexplored
physicalprinciples,willbeakeyto
achievingnewlevelsofbitdensity,
energy-efficiencyandperformance.
Globaldemandfordatastoragegrows
exponentially,andtoday’sstorage
technologieswillnotbesustainablein
nearfutureduetoexcessivematerial
resources needed to support the
ongoingDataExplosion.Thus,new
radicalsolutionsfordata/information
storagetechnologiesandmethodsare
required.Figure4showstheprojections
ofglobaldatastoragedemand—both
a conservative estimate and an upper
bound.AsindicatedbyFigure4,future
informationandcommunicationtechnologiesareexpectedtogenerateenormous
amountsofdata,farsurpassingtoday’sdataflows.Currently,theproductionanduse
ofinformationhasbeengrowingexponentially,andby2040theestimatesforthe
worldwideamountofstoreddataarebetween1024and1028bitsasshowninFigure4.
Furthermore,asaresult,whiletheweight of a single bitinthecaseofultimatelyscaled
NANDflashmemoryis1picogram(10-12g),thetotal massofsiliconwafersrequiredto
store1026bitswouldbeapproximately1010kg,whichwouldexceedtheworld’stotal
availablesiliconsupply(Figure5).
ChallengeGlobaldemandforconventionalsilicon-basedmemory/storageisgrowing
exponentially(Figure4),whilesiliconproductionisgrowingonlylinearly(Figure5).
Thisdisparityguaranteesthatsilicon-basedmemorywillbecomeprohibitively
expensiveforZetta-scale“bigdata”deploymentswithintwodecades.
Figure4:Globaldemandformemoryandstorage(utilizingsiliconwafers),isprojectedtoexceedtheamountofglobalsiliconthatcanbeconvertedintowafers.
10
Decadal Plan Interim Report
Grand Goal #2a:
Developemergingmemoriesandmemoryfabricswith>10-100X
densityandenergyefficiencyimprovementforeachlevelofthe
memoryhierarchy.
Grand Goal #2b:
GrandGoal#3b:Discoverstoragetechnologieswith>100xstorage
densitycapabilityandnewstoragesystemsthatcanleveragethese
newtechnologies.
Inaddition,memory,suchasDRAM,isanessential
component of computers and further advances in
computingareimpossiblewithout‘reinventing’
thecomputememorysystemincludingdevice
physics,memoryhierarchyarchitectureand
physicalimplementation.Forexample,traditional
embeddednonvolatilememorycannolonger
bescaledbelow28nm,thusalternativesare
needed.Finally,newmemorysolutionsmustbe
abletosupportmultipleemergingapplications,
suchas,e.g.artificialintelligence,large-scale
heterogeneoushigh-performanceanddata-center
computing,andvariousmobileapplicationsthat
alsomeettheruggedenvironmentalrequirements
oftheautomotivemarket,etc.
Call for ActionRadicaladvancesinmemoryanddatastorage
arerequiredsoon.Collaborativeresearch‘from
materialstodevicestocircuitstoarchitectureto
processingandsolutions’forfuturehigh-capacity
energy-efficientmemoryanddata/information
storagesolutionsforthevastrangeoffuture
applicationsisneeded.
[3]TheDecadalPlanExecutiveCommitteeofferedrecommendationsonallocationoftheadditional$3.4BinvestmentamongthefiveseismicshiftsidentifiedintheDecadalPlan.ThebasisofallocationisthemarketsharetrendandouranalysisoftheR&DrequirementsfordifferentsemiconductorandICTtechnologies.
Figure5:GlobalSiwafersupply:1990-2020dataandfuturetrend.
Invest $750M annually throughout this decade in new trajectories for memory and storage. Selected priority research themes
are outlined below.3
Fast,dense,energyefficient,cost-effective,embeddednon-volatilememories
MemoryandStoragefornewinformationrepresentationparadigms
Memoryforquantumprocessors
Fundamentalnewstoragetechnologies(i.e.DNA)
Priority Research Themes
11
Decadal Plan Interim Report
growingtrendonedgecomputingforAIsystemstocaterforprivacy
andfasterresponsetime,theexplosionofinformationgenerated
andstoredwillrequiretremendousgrowthoncloudstorageand
communicationinfra-structure.
Thecurrentstateofthedevelopedworldis
characterizedby(almost)always-available
communicationandconnectivity,whichhas
atremendousimpactonallaspectsoflife.A
manifestationofthisisCloudStorageandComputing.
Theabilitytogetdatafromanywhereandsendit
to everywhere has transformed both the way we do
businessandourpersonalhabitsandlifestyle.Social
networksareanexample.However,themainconcept
ofthecloudisbasedontheassumptionofconstant
connectivity,whichisnotguaranteed.Furthermore,
thedemandgrowsdailyforcommunicationtobecome
moreubiquitousaswebecomemoreconnected.An
alarmingtrendisagrowinggapbetweentheworld’s
technologicalinformationstorageneed,asjust
discussed,andcommunicationcapacitiesshownin
Figure6.Forexample,whilecurrentlyitispossibleto
transmitallworld’sstoreddatainlessthanoneyear,
in2040itispredictedtorequireatleast20yearsfor
thetransmission.Aglobalstorage-communication
cross-overisexpectedtohappenaround2022which
mayhaveatremendousimpactonICT.Evenwith
Seismic shift #3Always available communication requires new research directions that address the imbalance of communication capacity vs. data generation rates.
Figure6:TheGlobalCommunicationDataGenerationCrossoveroccurswhenthedatageneratedexceedstheworld’stechnologicalinformationstorageandcommunicationcapacities,creatinglimitationstotransmissionofdata.
12
Decadal Plan Interim Report
Grand Goal #3a:
Advancecommunicationtechnologiestoenablemoving
aroundallstoreddataof100-1000zettabyte/yearatthe
peakrateof1Tbps@<0.1nJ/bit.
Grand Goal #3b:
Developintelligentandagilenetworksthateffectively
utilizebandwidthtomaximizenetworkcapacity.
Call for ActionRadicaladvancesincommunicationwillberequired
toaddressgrowingdemand.Forexample,thecloud
technologiesmayundergosubstantialchangeswithemphasis
shiftingtowardedgecomputingandlocaldatastorage.
Broadbandcommunicationswillexpandbeyondsmart
phonestoimmersiveaugmentedreality,virtualmeetings
andsmartofficesettings.Newcapabilitieswillenrichuser
experiencesthroughnewusecasesandnewverticalmarkets.
Thisrequirescollaborativeresearchspanningabroadagenda
aimingatestablishingrevolutionaryparadigmstosupport
futurehigh-capacity,energy-efficientcommunicationforthe
vastrangeoffutureapplications.TheDOEOfficeofScience
publishedareportinMarch2020toidentifythepotential
opportunitiesandexplorethescientificchallengesof
advancedwirelesstechnologies.4
Challengeswouldincludewirelesscommunicationtechniques
expandingtoTHzregion,wirelessandwirelinetechnologies
interplay,newapproachestonetworkdensification,
increasingimportanceofsecurity,newarchitecturesfor
mm-wave,devicetechnologytosustainbandwidthandpower
requirements,packagingandthermalcontrol.
[4]https://www.osti.gov/servlets/purl/1606539
[5]TheDecadalPlanExecutiveCommitteeofferedrecommendationsonallocationoftheadditional$3.4BinvestmentamongthefiveseismicshiftsidentifiedintheDecadalPlan.ThebasisofallocationisthemarketsharetrendandouranalysisoftheR&DrequirementsfordifferentsemiconductorandICTtechnologies.
Invest $700M annually throughout this decade in new trajectories for communication. Selected priority research themes are
outlined below.5
MIMOwith1000sofantennas
mm-Wavefiltersandisolators
“Wired”densityandefficiencycopperandoptical
Priority Research Themes
mm-Wave CMOS
New physics of communication
13
Decadal Plan Interim Report
Seismic shift #4Breakthroughs in hardware research are needed to address emerging security challenges in highly interconnected systems and Artificial Intelligence.
encryptionstandardsresistanttoquantumattackmust
bedeveloped,withconsiderationgiventotheimpactof
thesestandardsonsystemperformance.Also,privacyhas
emergedasamajorpolicyissuedrawingincreasedattention
byconsumersandpolicymakersacrosstheglobe.Technical
approachestoenhancingprivacyincludeobfuscatingor
encryptingdataatthetimeofcollectionorrelease.
Inanotherdirection,deviceshavepermeatedthephysical
world,andthustrustinthesedevicesbecomesamatterof
safety.Securityhasthusneverbeenmoreimportant.Safety
Today’shighlyinterconnectedsystemsandapplications
requiresecurityandprivacyforproperoperation(Figure7).
Corporatenetworks,social-networkingandautonomous
systemsareallbuiltontheassumptionofreliableandsecure
communicationbutareexposedtovariousthreatsand
attacksrangingfromexposureofsensitivedatatodenial
ofservice.Thefieldofsecurityandprivacyisundergoing
rapidfluxthesedaysasnewusecases,newthreats,and
newplatformsemerge.Forinstance,newthreatvectors
throughtheemergenceofquantumcomputingwillcreate
vulnerabilitiesincurrentcryptographicmethods.Thus,new
Figure7:Systemsviewofsecurity(courtesyofYiorgosMakris/UniversityofTexasatDallas)
14
Decadal Plan Interim Report
andreliabilityofsystemsneedtoconsidermaliciousattacks
inadditiontothetraditionalconcernsofrandomfailures
anddegradationofphysical-worldsystems.Securityof
cyber-physicalsystemsneedstoconsiderhowtocontinue
tofunctionorfailgracefullyevenafterattacks.Weneed
intelligentalgorithmsthatsiftthroughcontextualdatato
evaluatetrust,todosecuresensorfusionovertime.Thisisa
difficultproblemascontextualdatahastremendousvariety
andquantity—thesystemsofthefutureareactuallysystems
ofsystemswithlimitlesspossibilitiesforcommunicationand
signaling.Forinstance,carscancommunicatewitheachother
andalsowithroadsideinfrastructure.Likehumans,weneed
toaugmentsystemswiththeintelligencetotrustornottrust
alltheyperceive.
Ourhardwareisalsochanging.Complexityistheenemyof
security,andtoday’shardwareplatformsarehighlycomplex
duetothedriversofperformanceandenergy-efficiency.
Modernsystem-on-chipdesignsincorporateanarrayof
special-purposeacceleratorsandintellectualproperty(IP)
blocks.Thesecurityarchitectureofthesesystemsiscomplex,
as these systems are now tiny distributed systems where
wemustbuilddistributedsecuritymodelswithdifferent
trustassumptionsforeachcomponent.Furthermore,
thesecomponentsareoftensourcedfromthird-parties,
implyingtheneedfortrustinthehardwaresupplychain.
Thepursuitofperformancehasalsoledtosubtleissuesin
microarchitecture.Forinstance,manyexistinghardware
platformsarevulnerabletospeculativeexecutionside-
channelissues,famouslyexposedbySpectreandMeltdown.
Drivenbytheseproblemsandothers,thefuturerequires
fundamentallynewhardwaredesigns.
ThemajorworkloadoftodayisAI.Manysecuritysystems,
forinstance,useanomalydetectiontoidentifyattacks
oremployfeatureanalysisforcontextualauthentication.
AI’scapabilitiescontinuetoincrease,andapplicationsfor
thesetrustedsystemscontinuetogrow.However,the
trustworthinessoftheAIforthesesystemsisunclear.Thisis
aproblemnotjustforsecuritysystemsbutevenforgeneral
systemswithimplicittrustassumptions,forinstance,visual
objectdetectioninautonomousvehicles.Researchershave
shownthatsmallperturbationstoanimagecanswayneural
networkmodelsintothewrongconclusion.Awell-placed
smallstickeronastopsigncanmakeamodelclassifyitas
aSpeedLimit45sign.6Otherapplicationsofdeeplearning
systemshavesimilartrustissues:theoutputofspeech
recognitionmightbemanipulatedwithimperceptibleaudio
changes,ormalwaremightgoundetectedwithsmallchanges
tothebinary.Thebrittlenessofdeeplearningmodelsis
relatedtotheirfamousinscrutability.Neuralnetworksare
blackboxeswithnoexplanationfortheirdecisions.Other
importantproblemswithneuralnetworksarealgorithmbias
andfairness.Approachesareneededtomakedeeplearning
systemsmoretrusted,explainable,andfair.
Finally,inthelastdecade,thesystemsthatwemustsecure
havebecomeimmeasurablymorecomplex.Thecloudhas
becomethestandardforoutsourcingcomputationand
storage,whilemaintainingcontrol.Wearestillgrapplingwith
securitychallengesarisingfromthecloud—multi-tenancy,
providerassurance,andprivacy—whilecloudofferings
continuetoincreaseincomplexity.Cloudsnowoffertrusted
executionenvironmentsandspecialized,sharedhardware
andsoftware.Atthesametime,interestinedgecomputing
isgrowingaswerealizecloudslacktheperformanceand
privacyguaranteesofnearbycomputeinfrastructure.The
heterogeneousnatureoftheedgeimpliestrustinedge
serviceprovidersisamajorissueand,ofcourse,thesecurity
ofIoTdeviceshasplaguedusforyears.Developingsecurity
mustbemadeeasierforresource-constrained,oftenlow-cost
devices.Evenifcareistakeninthesecuritydesign,difficulties
arisefromextremeenvironments,suchasmedicalimplants.
Tocompoundtheproblem,systemshavebecomemore
complicatedateverylevel—modernsystem-on-chipdesigns
incorporateanarrayofspecial-purposeacceleratorsandIP
blocks,basicallytinydistributedsystemswherewemustbuild
distributedsecuritymodelswithdifferenttrustassumptions
foreachcomponent.
[6]“Whatisadversarialmachinelearning?”byBenDickson–July15,2020https://bdtechtalks.com/2020/07/15/machine-learning-adversarial-examples/
15
Decadal Plan Interim Report
Grand Goal #4:
Developsecurityandprivacyadvancesthatkeeppacewith
technology,newthreats,andnewusecases,forexample,
trustworthyandsafeautonomousandintelligentsystems,
securefuturehardwareplatforms,andemergingpost-
quantumanddistributedcryptographicalgorithms.
Call for ActionThepaceatwhichtoday’ssystemsareincreasinginintelligence
andubiquityisastounding.Atthesametime,theincreased
scaleandcomplexityofthesesystemshaveforcedhardware
specializationandoptimizationtoaddressperformance
challenges.Alltheseadvancesincapabilitymustgohand-
in-handwithadvancesinthesecurityandprivacy.Examples
includesecuringweaknessesinthemachine-learningor
conventionalcryptography,protectingprivacyofpersonaldata,
andaddressingvulnerabilitiesinthesupply-chainorhardware.
[7]TheDecadalPlanExecutiveCommitteeofferedrecommendationsonallocationoftheadditional$3.4BinvestmentamongthefiveseismicshiftsidentifiedintheDecadalPlan.ThebasisofallocationisthemarketsharetrendandouranalysisoftheR&DrequirementsfordifferentsemiconductorandICTtechnologies.
Invest $600M annually throughout this decade in new trajectories for ICT security. Selected priority research themes are
outlined below.7
Securityandprivacyoffuturehardwareplatforms,whicharecomposedof
heterogeneousandspecializedcomponentsandinvolvenewparadigmssuchas
quantumandneuromorphiccomputing
Emergingcryptography,suchashomomorphicencryptiontosupportnewuse
casesandpost-quantumalgorithmstothwartnewattacks
SecurityfornewsystemarchitecturesthatcomputeatallpointsfromIoTtoEdge
toCloud,aswellasmassivelydistributedprocessingsuchasinblockchains
Priority Research Themes
AIsystemsthatincorporatemechanismstoenabletrustandyetareascapableas
today’ssystems
16
Decadal Plan Interim Report
Seismic shift #5Ever rising energy demands for computing vs. global energy production is creating new risk, and new computing paradigms offer opportunities with dramatically improved energy efficiency.
Rapidadvancesincomputinghaveprovidedincreased
performanceandenhancedfeaturesineachnewgeneration
ofproductsinnearlyeverymarketsegmentwhetheritbe
servers,PCs,communications,mobile,automotive,and
entertainmentamongothers.Theseadvanceshavebeen
enabledbydecadesofR&Dinvestmentsbyboththeprivate
sectorandthegovernmentyieldingexponentialgrowth
incomputespeed,energyefficiency,circuitdensity,and
cost-effectiveproductioncapacity.Sustainedinnovation
insoftwareandalgorithms,systemsarchitecture,circuits,
devices,materials,andsemiconductorprocesstechnologies,
havebeenfoundationaltothatgrowthpace.Althoughthis
trendhaspersistedfordecadesbysuccessfullyovercoming
manytechnologicalchallenges,itisnowrecognizedthat
conventionalcomputingisapproachingfundamentallimits
intheenergyefficiencyandthereforepresentingchallenges
thataremuchhardertosurmount.Consequently,disruptive
innovationsininformationrepresentation,information
processing,communication,andinformationstorageareall
pressingandcriticaltosustainableeconomicgrowthandthe
UnitedStatestechnologicalleadership.
Asthecomputationsperyearincreases,thenumberofbits
usedtosupportthesecomputationsalsoincreases.Itis
projectedthatin2050wewillbedealingwithnearly1044
bits.AsshowninFigure8a,thetotalenergyconsumptionby
general-purposecomputingcontinuestogrowexponentially
and is doubling approximately every 3 yearswhiletheworld’s
energyproductionisgrowingonlylinearly,byapproximately
2%ayear.Therisingglobalcomputeenergyisdrivenby
evergrowingdemandsforcomputation(Figure8b)andthis
isinspiteofthefactthatthechip-levelenergyperone-bit
transitionincomputeprocessorunits(e.g.CPU,GPU,FPGA)
hasbeendecreasingoverlast40years(asmanifestedbythe
Moore’slaw),andis~10aJor10-17Jincurrentprocessors.
However,thedemandforcomputationgrowthisoutpacing
theprogressrealizedbyMoore’slaw.Inaddition,Moore’s
lawiscurrentlyslowingdownasdevicescalingisapproaching
fundamentalphysicallimits.Iftheexponentialgrowthin
computeenergyisleftunchecked,marketdynamicswilllimit
thegrowthofthecomputationalcapacitywhichwouldcause
aflatteningouttheenergycurve(the‘marketdynamics
limiting’scenarioinFigure8a).Thus,theradicalimprovement
inenergyefficiencyofcomputingisrequiredtoavoidthe
‘limiting’scenario.
Theunderlyingdifficultproblemisbit utilization efficiency in
computation,i.e.thenumberofsinglebittransitionsneeded
toimplementacomputeinstruction.ThecurrentCPUcompute
trajectoryisdescribedbyapowerformula(shownasinletin
Figure9)withanexponentboundedbyp~⅔.Thetheoretical
basisfortheobservedtrajectoryandforthevalueofthe
exponentisnotclearlyunderstoodandthusthetheoretical
basisforcomputationneedstobefurtherdeveloped.Asan
observation,ifitcouldbepossibletoincreasetheexponent
intheformulabyonly~30%,thecomputeefficiencyandthus
energyconsumptionwouldhavea1000,000ximprovement.
Thisisillustratedas“newtrajectories”inFigure9.
Grand Goal #5:
Discovercomputingparadigms/architectureswitha
radicallynew‘computingtrajectory’demonstrating
>1,000,000ximprovementinenergyefficiency.
Changingthetrajectorynotonlyprovidesimmediate
improvementsbutalsoprovidesmanydecadesofbuffer
(asshowninFigure8).Thiswouldbemuchmorecost
effectivethanattemptingtoincreasetheworld’senergy
supplydramatically.
17
Decadal Plan Interim Report
[8]M.HilbertandP.Lopez,“Theworld'stechnologicalcapacitytostore,communicate,andcomputeinformation,”Science332(2011)60-65.
Figure8(a):Totalenergyofcomputing:Thesolidyellowlineindicatescontinuingthecurrentcomputingtrajectorywhileimprovingthedeviceenergyperformance.Thedashedyellowlineindicatesa‘marketdynamicslimited’scenariostoppingfurtherincreaseintheworld’scomputingcapacityandresultinginaflatteningouttheenergycurve.Theblueboxindicatesascenariowherearadicallynewcomputingtrajectoryisdiscovered.TheDecadalPlanmodel(yellowline)iscomparedtoindependentdatabydifferentgroups(circleddots):
[1]W.VanHeddeghemetal(UGhent),“TrendsinworldwideICTelectricityconsumptionfrom2007to2012”,ComputerCommunications50(2014)64
[2]IEA(2017),Digitalizationandenergy,IEAPublications,Cedex,Paris
[3]EuropeanCommission(2015),Eco-designpreparatory study on enterprise servers and dataequipment.Luxembourg:PublicationsOfficeoftheEuropeanUnion
[4]E.Masanet,etal(NorthwesternU,LBNL,KoomeyAnalytics),“Recalibratingglobaldatacenterenergy-useestimates”,Science367(2020)984
[5]H.Fuchsetal(LBNL),“Comparingdatasetsofvolumeserverstoilluminatetheirenergyusein datacenters”,EnergyEfficiency13(2020)379
[6]Malmodinetal.(Ericsson,TeliaCompany),“ThefuturecarbonfootprintoftheICTandE&MsectorsProceedingsofthe1stInternationalConference on Information and Communication TechnologiesforSustainabilityETHZurich,February14-16,2013pp.12-20
[7]A.S.G.AndraeandT.Edler(Huawei–Sweden),“Onglobalelectricityusageofcommunicationtechnology:Trendsto2030”,Challenges6(2015)117-157
[8]L.BelkhirandA.Elmeligi(McMasterU,Canada),“AssessingICTglobalemissionsfootprint:Trendsto2040&recommendations”,J.CleanerProduction177(2018)448
Figure8(b):World’stechnologicalinstalledcapacitytocomputeinformation,inZIPS,for2010-2050.Thesolidyellowlineindicatesthecurrenttrend(basedonresearchbyHilbertandLopez4).Thedashedyellowlineindicatesa‘marketdynamicslimited’scenariostoppingfurtherincreaseintheworld’scomputingcapacityduetolimitedenergyenvelope.Theblueboxindicatesascenariowherearadicallynewcomputingtrajectoryisdiscovered.
18
Decadal Plan Interim Report
Call for ActionRevolutionarychangestocomputingwillberequiredsoon.
Computationalloadscontinuetogrowexponentiallyas
evidencedbythegrowthin“artificialintelligence”(AI)
applicationsandtrainingdemands.Newapproachesto
computingsuchasin-memorycompute,specialpurpose
computeengines,differentAIplatforms,braininspired/
neuromorphiccomputation,quantumcomputing,orother
solutionswillbenecessaryandwillneedtobecombined
inaheterogeneousmanner.Thescopeofpotential
heterogeneouscomputingarchitecturesaredescribedina
recentNationalScienceandTechnologyCouncil(NSTC)report
ontheFutureofComputing.9Thisresearchwillrequire
across-disciplinary,cross-functionalapproachtorealize
commerciallyviableandmanufacturablesolutionswithmulti-
decadelongevitytoreplacethemainstreamdigitalapproach.
Thisdocumentisintendedtostimulatecollaborativeresearch
'frommaterialstoarchitectureandalgorithms'toestablish
revolutionaryparadigmsthatsupportfutureenergy-efficient
computingforthevastrangeoffuturedatatypes,workloads
andapplications.Foradditionalbackground,seetheDOE
OfficeofScience,BasicResearchNeedsforMicroelectronics
workshopreport.10
[9]https://www.nitrd.gov/pubs/National-Strategic-Computing-Initiative-Update-2019.pdf
[10]https://science.osti.gov/-/media/bes/pdf/reports/2019/BRN_Microelectronics_rpt.pdf
[11]TheDecadalPlanExecutiveCommitteeofferedrecommendationsonallocationoftheadditional$3.4BinvestmentamongthefiveseismicshiftsidentifiedintheDecadalPlan.ThebasisofallocationisthemarketsharetrendandouranalysisoftheR&DrequirementsfordifferentsemiconductorandICTtechnologies.
Figure9:ThecurrentCPUcomputetrajectory
Invest $750M annually throughout this decade to alter the compute trajectory. Selected priority research themes are
outlined below.11
Shannonframeworkforcomputation
High-dimensionalrepresentation
AIprocessors:“Cambrian Explosion” needs to happen!
DecouplethecomputepowerfromenergyconsumptioninQuantumComputers
Priority Research Themes
Turing Shannon Approximate computing
19
Decadal Plan Interim Report
4819 Emperor Blvd Suite 300 Durham, NC 27703
919.941.9400
www.src.org