Beyond-CMOS Research and Benchmarking: An SRC Perspective · 2017-12-12 · 4 The Pursuit of...
Transcript of Beyond-CMOS Research and Benchmarking: An SRC Perspective · 2017-12-12 · 4 The Pursuit of...
Beyond-CMOSResearchandBenchmarking:AnSRCPerspectiveNovember29,2017
AnChenExecutiveDirector,NRI&nCORE
Outline
• Introduction• Beyond-CMOSresearchinNanoelectronics ResearchInitiative(NRI)
q NRIresearchscopeq HighlightsofNRIachievements
• LessonslearnedfromtheNRIprogram• What’snext?• Summary
2
3
SemiconductorResearchCorporation(SRC)
Private foundation that manages and promotes SRC student training,
tracking, and hiring into areas of interest for our
sponsors
FAMELEAST
C-SPINTerraSwarmCFAR
SONIC
• Center based research program with 6 focus areas
• 5 year duration using [3+2] model
• Longer term research• MCs support all areas• US researchØGov’t sponsor = DARPA
$40M
INDEXCNFDSWAN
SupplementalNSF Grants
• Center based research program with 3 focus areas
• 5 year duration using [3+2] model
• Longer term research• MCs support all areas• US researchØGov’t sponsor = NIST, NSF
$7M
• Individual project management• 12 themes• Shorter duration projects
(2-3 yrs.)• Nearer term research• Flexible MC engagements• International research OKØ Gov’t sponsors = NIST, NSF
T3S
SLD
SSB
PKG NMPLMD
I3T ESHEP3C
CSRCADT
AMS-CSD
$30M
4
ThePursuitofLow-PowerSwitches
Traditional performance improvements (clock freq.andILP)havebeen flattening due topowerconstraint.
Needlower-powerswitchesbeyondCMOS
1.0E+02
1.0E+03
1.0E+04
1990 1995 2000 2005 2010
Clo
ck S
peed
(MH
z)
103
102
104
2004 Frequency Extrapolation
Microprocessor Clock Frequency
Peak cut-off freq. vs. 1/Lg(Lee et al., IEDM 2007)
DeviceFrequency CircuitFrequency
Exascale computingchallenges:• Reducingpowerrequirements• Copingwithrun-timeerrors• Exploitingmassiveparallelism
“TheOpportunitiesandChallenges ofExascale Computing,”SummaryReportoftheAdvanced ScientificComputingAdvisoryCommittee(ASCAC) Subcommittee,DOEOffice ofScience(2010)
5
NRITimelineandResearchVectors20
03
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
NR
I pl
anni
ng
star
ted
Cert
ifica
te o
f in
corp
orat
ion
1st
NR
I-N
SF
proj
ects
, 2yr
WIN
star
ted
IND
EXst
arte
dSW
ANst
arte
d1s
tan
nual
rev
iew
1st
NIS
T aw
ard
MIN
Dst
arte
d
Benc
hmar
king
ph
ase
1 st
arte
d
phas
e 1.
5 R
FP
Benc
hmar
king
ph
ase
2 st
arte
d
2nd
NIS
T aw
ard
phas
e 2
RFP
phas
e 2
star
ted:
IN
DEX
, SW
AN,
CNFD
NR
I-ST
ARne
tBe
nchm
arki
ng
Cent
erst
arte
dE2
CDA
NEB
202
0 st
arte
d
Phase 1 Phase 1.5 Phase 2
STAR
Tnet
star
ted
NISTpartnershipNSF partnershipNRIcenters
nCORE
NRIresearchvectors• Newlow-powerdeviceswithnewstatevariables
• Newwaystoconnectdevices• Newmethodsforcomputation• Newmethodstomanageheat• Newmethodsoffabrication
Outline
• Introduction• Beyond-CMOSresearchinNanoelectronics ResearchInitiative(NRI)
q NRIresearchscopeq HighlightsofNRIachievements
• LessonslearnedfromtheNRIprogram• What’snext?• Summary
6
7
Nanoelectronics ResearchInitiative(NRI)Takingcomputingbeyondthelimitationsofcurrenttechnology
UC Los AngelesUC BerkeleyUC IrvineUC Santa Barbara
Notre DamePurdue Penn StateUT-Dallas
UT-Austin RiceUT-Dallas NCSUU. Maryland Texas A&M
SUNY-AlbanyPurdue ColumbiaHarvard U. VirginiaGeorgia TechMIT
U. Nebraska-Lincoln
U. Delaware U. Wisconsin-
MadisonU. OaklandSUNY Buffalo UC-Irvine
U. AlabamaU. ArkansasBrownCaltechCMUU. ChicagoColumbiaCornellDrexelU. FloridaGeorgia TechHarvardUIUC
U. MarylandU. MassachusettsU. MichiganU. MinnesotaMITNebraska-LincolnNorthwesternNotre DameU. OklahomaPenn StateU. PennsylvaniaU. PittsburghPrincetonPurdueRochesterSUNY-BuffaloStanfordWisconsin-MadisonUC BerkeleyUC-IrvineUC-San DiegoUC-Santa BarbaraUC-RiversideU. Virginia Virginia CommonwealthYale
BenchmarkingGeorgia Tech
2008-20122006-2017
2013-2017
2006-2012
2006-2017
2015-2017
• >1500journalpapers
• >40patents• >300Ph.D./post-doc
Ø MRSECsupplement
Ø NEB2020Ø E2CDA
NRI: a program searching for the “next switch”
MajorNRIdevicecategoriesü Steepslopetransistorsü Spintronicsü vanderWaalsdevices
8
STARnet:SRC-DARPACollaborationSemiconductorTechnologyAdvancedResearchNetwork
Applied MaterialsNovellus
FAME Jane Pei-Chen Chang, Director The mission of FAME is to create and investigate new nonconventional atomic scale engineered materials and structures of multi-function oxides, metals and semiconductors to accelerate innovations in analog, logic and memory devices for revolutionary impact on the semiconductor and defense industries.
TerraSwarm Edward A. Lee, DirectorThe TerraSwarm Research Center aims to enable the simple, reliable, and secure deployment of a multiplicity of advanced distributed sense control-actuate applications on shared, massively distributed, heterogeneous, and mostly uncoordinated swarm platforms through an open and universal systems architecture.
C-SPIN Jian-Ping Wang, DirectorThe Center for Spintronic Materials, Interfaces and Novel Architectures (C-SPIN) seeks to overcome barriers to realizing practical spin-based memory and logic technology by assembling experts in magnetic materials, spin transport, novel spin-transport materials, spintronic devices, circuits, and novel architectures.
SONIC Naresh Shanbhag, DirectorSONIC will be guided by the following mission: To enable equivalent scaling in beyond-CMOS nanoscalefabrics by embracing their statistical attributes within statistical-inference-based applications, architectures, and circuits, to achieve unprecedented levels of robustness and energy efficiency.
LEAST Alan Seabaugh, DirectorThe Center for Low Energy Systems Technology (LEAST) explores the physics of new materials and devices to enable more energy-efficient integrated circuits and systems.
C-FAR Todd Austin, DirectorThe center's research agenda is guided by three initial technical vectors, whose intersections will help realize non-conventional architectures that address these pressing challenges: data-centric architectures, novel architectures based on emerging technologies, and beyond homogenous parallelism.
16 Universities 10 Universities
15 Universities
8 Universities14 Universities
10 Universities
Spintronics
vdW Mater.
Oxidesandinterfaces
Topologicalinsulators
Steep-slopedevices
2Dmaterials
NRI/WIN
NRI/MIND
SelectedHighlightsofNRIResearchAchievements
• InventedCu-basedlarge-scalegrapheneCVDgrowthtechnique§ Akeybreakthroughforgraphenegrowthachievedbyacademia-industrycollaboration;adoptedworldwideandcommercialized
• ProofofelectronopticsingraphenePNjunction§ Nametop10breakthroughsin2016byPhysicsWorld
• DemonstrationofelectricswitchingofboundarymagnetizationinCr2O3• Potentialforlow-powernonvolatilememoryandcomputing
• Exploreroom-temperatureexciton devicesin2Dheterostructures• Potentialforlow-powerswitchandelectronics-opticsintegration
• Inventedthe“negative-capacitanceFET”concept§ A low-powersteep-slopetransistordrivinginterestinferroelectrics
9S.Salahuddin andS.Datta,Nano Lett.8(2),405-410 (2008)
X.Li,etal,Science324,1312-1314 (2009)
S.Chen,etal,Science353,1522-1525 (2016)
X.He,etal,Nat.Mater.9,579-585 (2010);W.Echtenkamp, etal,Phys.Rev.Appl.7,034015-1-7 (2017)
M.M.Fogler,etal,Nat.Comm.5,1-5 (2014);E.V.Calman,etal,Appl.Phys. Lett.108,101901-1-4 (2016)
Outline
• Introduction• Beyond-CMOSresearchinNanoelectronics ResearchInitiative(NRI)
q NRIresearchscopeq HighlightsofNRIachievements
• LessonslearnedfromtheNRIprogram• What’snext?• Summary
10
LessonsLearnedfromtheNRIProgram
• Nobeyond-CMOSdevicehasbeenproventobecapableofreplacingCMOSforBooleanlogicandvonNeumannarchitectures.□ IfthereisaCMOSsolution,thedifficultyandrequirementofadoptingabeyond-CMOStechnologywillbesignificantlyhigher.
11
12
Beyond-CMOSDeviceBenchmarking
Benchmarkingconsiderations:• Theoretical projection vs.experimental demonstration• Accuracyand rigorousness ofassumptions andparameters
ThedevelopmentofNRIbenchmarking
C.PanandA.Naeemi,DATE(2017)
32-bitadderenergyvsdelay
Phase1: ledbyKerryBernstein(IBM)• Buildcircuitgates(inverter,NAND, adder)withNRIdevicestobenchmark performance(delay,power, area)againstCMOS
• Consider circuitparameters,e.g.,spanofcontrol,noise immunity, logicaleffort
Phase2: ledbyDmitriNikonov andIanYoung(Intel)• Develop uniformandtransparent criteriaandmethodologies forbenchmarking
• Improvequalityofassumptions anddevicedataforbenchmarking
Phase3: ledbyAzadNaeemi (GeorgiaTech)• Expand benchmarkingtoSTARnet devices• Expand benchmarkingtonon-Boolean logicandmemoryapplications
PracticalChallengesofBeyond-CMOSDevices
• Difficulttoachievesustainableimprovement• Generationsofdimensionalscalingisnolongeravailable
• Adoptionbarrier:expensiveandinflexibleinfrastructure• Manufacturing,materials,designstoolsandlibrary,…
• Challengingtimelineforadoptionofanewtechnology
13
Incubationtime:12-15yearsStrainedSi:1992– 2003
High-KMetalGate:1996– 2007Raisedsource/drain:1993–2009
Multi-gate:1997–2011
•Drop-inreplacementofCMOStransistors• CMOScompatibilityandenhancement• Significantlyimprovedperformance,reducedcost,ornovelfunctionalitybeyondthecapabilityofCMOS
Desirablefeaturesofanewtechnology
LessonsLearnedfromtheNRIProgram
• Nobeyond-CMOSdevicehasbeenproventobecapableofreplacingCMOSforBooleanlogicandvonNeumannarchitectures.□ IfthereisaCMOSsolution,thedifficultyandrequirementofadoptingabeyond-CMOStechnologywillbesignificantlyhigher.
□ Therearestillinterestinglow-powerdevicemechanismsworthexploring
14
15
LowPowerDeviceMechanisms
§ EnergyFiltering– Tunnel-FET–Graphene p-nJunctionDevice
§ InternalPotentialStep-up– Ferroelectric-GateFET
§ InternalTransduction– Spin-FET– ElectromechanicalRelay– PiezoElectronic Transistor(PET)
§ GatedPhaseTransition– Bi-layerPseudo-spinFET(BiSFET)
Deviceconcepts explored intheNRIprogram
Basicphysicalmechanisms definesthepotentialofadevice;however, non-ideal andparasiticcomponents determine deviceperformance.
16
Low-PowerDevicesinE2CDA(NSF-SRCCollaboration)
2DElectrostrictiveFET(EFET) 3HJTFET
E-field→Strain→bandgapmodulation Tripleheterojunction↓
Greaterbuilt-in potential↓
Thinner barrier
S.Das,Sci.Rep.6,34811 (2016)
LessonsLearnedfromtheNRIProgram
• Nobeyond-CMOSdevicehasbeenproventobecapableofreplacingCMOSforBooleanlogicandvonNeumannarchitectures.□ IfthereisaCMOSsolution,thedifficultyandrequirementofadoptingabeyond-CMOStechnologywillbesignificantlyhigher.
□ Therearestillinterestinglow-powerdevicemechanismsworthexploring□ Beyond-CMOSdevicesmayfindpromisingopportunitiesfornovelcomputingparadigmsbeyondBooleanlogicandvonNeumannarchitectures.
17
ExamplesofBeyond-CMOSDevices“Idiosyncrasy”
• Device-levelreconfigurability
• Built-inmemoryinlogicswitches
• Tunableanalogbehaviors
• Programmablerandomness
• Computingelementswithcomplexfunctions,e.g.,nativemajoritylogic
• ……18
Hardwaresecurity,e.g.,logiccamouflaging,…
In-memorycomputing,memory-in-logic,nonvolatilelogic,…
Neuromorphiccomputing,analogcomputing,…
Probabilisticlogic,randomnumbergenerator,…
Moreefficientlogicdesignandcomputingsystems
19
In/Near-MemoryComputingandNVLogic
• Breakthe“memorywall”:reducedatamovement, improveefficiencyandthroughput• Massiveparallelismandsuitabilityforbigdataapplications• Bit/gate-levelnonvolatiledatastorage
MTJ, ReRAM, FRAM, FeFET, PCM, Flash …
SRAM, FF, adder, CAM, LUT, FPGA, …
NV gates and logic
Nonvolatile switches
NVM
CMOS logic
FF: flip-flop; LUT: look-up tableCAM: content-addressable memory
Memory-enabledcomputing
Micron’s DRAM-based Automata Processor Nonvolatile (NV) logicbasedonNV-Memory elements andNVswitches
Challenges:• Speed• Endurance• Application specific
P.Dlugosch,etal,IEEETrans.Paral.Dist.Syst.25,3088 (2014)
LessonsLearnedfromtheNRIProgram
• Nobeyond-CMOSdevicehasbeenproventobecapableofreplacingCMOSforBooleanlogicandvonNeumannarchitectures.□ IfthereisaCMOSsolution,thedifficultyandrequirementofadoptingabeyond-CMOStechnologywillbesignificantlyhigher.
□ Therearestillinterestinglow-powerdevicemechanismsworthexploring□ Beyond-CMOSdevicesmayfindpromisingopportunitiesfornovelcomputingparadigmsbeyondBooleanlogicandvonNeumannarchitectures.
• Novelmaterialandprocessingareessentialforbeyond-CMOSresearch.□ Thereissignificantmaterialandprocessingbarrierformostbeyond-CMOSdevices.□ Needtofocusandaddressthehardproblem.
20
21
Beyond-CMOSDevicesExploredinNRI
Phonon
Photon
Phase
Exciton
Spin
Charge
Si, III-V Ferromagnetic materials
Ferroelectric materials
Graphene / CNT
TMD
nanophtonics
Thermal logic
BiSFET
Excitonic logic
Nanoribbon FET
PN junction e-lensing
spinFET
spin wave logic
spin torque logic
NML
spin Hall device
resonant injection FET
laterial TFET
TFET
ASL/CSL
NC-FET
ME-MTJ
Fe-FET
FTJ
Molecule / QD
QCA
spinFET
BisFET
ITFET
ITFET Excitoiclogic
TFET
TFET
spin valve/filter
spin tunnel junction
atomic switchVertical hetero TFET
1D TFET
superlattice TFET
spin wave
bi-layer device
graphene spin logic (spinFET, spin Hall, spin torque, etc.)
spin oscillator
Spin logic
spinFET
CDW
charge density wave (CDW)
MaterialsStat
e va
riabl
es
22
GraphenePNJunction andBisFET
Exp.
[C. Pan, M. Elahi, A. Naeemi]S.Chen,etal,Science,353,1522 (2016)
AlainDiebold, NRIINDEXReview(2017)
• Edgeroughness• Junction roughness• Gatecontrol• Contact resistance• Collimation
BilayerPseudo-Spin FieldEffectTransistor(BiSFET)
Challenges:• Room-temperature exciton condensate• Chargeimbalance, interlayerrotation,freecarrieranddielectricscreening,…
S.K.Banerjee,etal,EDL30,158 (2009)
Graphene PNjunction device
Veselago lensing Exciton condensate
23
Negative-CapacitanceFET(NC-FET)
NC-FETconcept(Proposedin2008)
Oneofthemostactivelypursuedsteepslopetransistors
FerroelectricHfOx alloys
FeFETmemory
NCFinFET
NC-FETwith1.5nmHfZrO
NC-pFET Ge/GeSn
K.S.Li,etal,IEDM2015
M.H.Lee,etal,IEDM2016
J.Zhou,etal,IEDM2016
24
STARnet andNRIMaterialBenchmarking
• AcollaborativeeffortconductedbySTARnet andNRIPIstosummarizeandcomparenovelmaterialsproperties,performance,challenges,anddeviceapplicationsacrossresearchprograms.
• GuidedbyaMaterialBenchmarkingCommitteeformedbyindustryandacademia.• CurrentaccesslimitedtoSTARnet/NRIparticipants,withlong-termgoalofsharingwithbroadcommunity.
25
Example:2DMaterialsandTopologicalInsulators
Southwest Academy of Nanoelectronics
Bilayer pseudo-spin FET (BiSFET) materialsInterlayer tunneling FET (ITFET) materials Topological Insulator Magneto-Electronics (TIME) materials
Monolayer graphene Bilayer graphene Hexagonal boron nitride MoS2 WSe2 MoTe2
Topological Insulators Ferromagnets
Bi2Se3 Bi2Te3 CoFeB Py Co
Depositionor Processing
Conditions
Deposition or Growth Method
CVD, exfoliation from bulk graphite
CVD, exfoliation from bulk graphite
Exfoliation from bulk,CVD
CVD, MBE, exfoliation from bulk
MBE, exfoliation from bulk
Exfoliation from bulk crystals
MBE for thin film, Bridgman
technique for bulk
MBE for thin film, Bridgman technique for
bulk
sputter deposition
Substrate(s)(typical) Copper foil, SiC Copper foil, SiC Copper Foil or SiC
CVD: SiO2, Si, Al2O3, HOPG
MBE: HOPG, Al2O3, epitaxial graphene
CVD: SiO2, Si, Al2O3, HOPG
MBE: HOPG, Al2O3, epitaxial graphene
N/A
Si, sapphire Si, sapphire Si
Deposition or Growth T (ºC)
~1,000 °C for Cu, ~1,400 °C for SiC
~1,000 °C for Cu, ~1,400 °C for SiC > 1,000 °C 700-800 °C 900-1000 °C N/A 150-200 °C 350 °C RT
Wafer scale process? YES YES NO
YES, but thickness uniformity unclear
YES, but thickness uniformity unclear
No
YES YES YES
Key Properties
Mobilityor
Curie Temp(ferromagnets)
~10,000 cm2/Vs at room temperature
~10,000 cm2/Vs at room temperature N/A
50 cm2/Vs at RT1,000-10,000 below
10K
100 cm2/Vs at RT1-3,000 cm2/Vs below 10 K
20 cm2/Vs at RT
3,000 cm2/Vs 1,700 cm2/Vs 728 – 1008 K 869K 1388
K
Bandgapor
Spin polarization (ferromagnets)
0 eV 0 – 0.2 eV 5-6 eV
Bulk: 1.2 eVSL: 2.1 eV (quasi-
particle gap); 1.9 eV (optical)
Bulk: 1.3 eV SL: 1.65 eV (optical
gap)
Bulk: 0.9 eVSL: 1.1 eV
3. eV 0.15 eV 55 % 48 –55%
35 –42 %
ApplicationDevice Driver BiSFET BiSFET, ITFET BiSFET, ITFET BiSFET, ITFET BiSFET, ITFET BiSFET, ITFET TIME TIME
Function in the Device
Channel (individual layer) material
Channel (individual layer) material
Substrate, inter-layer dielectric
Individual layer,tunnel barrier
Individual layer,tunnel barrier
Individual layer,tunnel barrier Spin channel Spin injector/collector
Material Propertiesor Attributes Critical to the
Device Application
Ability to dope extrinsically Yes (adatoms) Yes (adatoms) N/A
Yes (adatoms, oxygen
vacancies)
Yes(adatoms)
YesNot needed because of small
bandgap N/A
Ability to make ohmic contacts Yes Yes N/A
Yes(using graphene for
nFETs)
Yes(using Pt for pFETs) Yes N/A
Effective mass1 massless
~0.05me, varies with density and
transverse electric field
0.4me (DFT)not measured experimentally
~0.4me
Massless,Fermi velocity vF=4.5×105 m/s
N/ADirac point near mid-gap
Dirac point in valence band
Unique Properties Rotationally aligned double layers with accuracy < 3° Large effective mass Large effective mass
LessonsLearnedfromtheNRIProgram
• Nobeyond-CMOSdevicehasbeenproventobecapableofreplacingCMOSforBooleanlogicandvonNeumannarchitectures.□ IfthereisaCMOSsolution,thedifficultyandrequirementofadoptingabeyond-CMOStechnologywillbesignificantlyhigher.
□ Therearestillinterestinglow-powerdevicemechanismsworthexploring□ Beyond-CMOSdevicesmayfindpromisingopportunitiesfornovelcomputingparadigmsbeyondBooleanlogicandvonNeumannarchitectures.
• Novelmaterialandprocessingareessentialforbeyond-CMOSresearch.□ Thereissignificantmaterialandprocessingbarrierformostbeyond-CMOSdevices.□ Needtofocusandaddressthehardproblem.
• Aholisticapproachisneededtoaddressnewbeyond-CMOSresearchdirections.□ Benchmarkingisessentialandneedstostartfromthebeginning.
26
27
NewNSF-SRCInitiative:E2CDA
Electronic-Photonic Integration Using the Transistor Laser for Energy-Efficient ComputingU. Illinois/Urbana-Champaign, U. Chicago
EXtremely Energy Efficient Collective ELectronics (EXCEL)Notre DamePenn State, U. ChicagoGeorgia Tech, UC-San Diego
2D Electrostrictive FETs for Ultra-Low Power Circuits and ArchitecturesPenn State
Memory, Logic, and Logic in Memory Using Three Terminal Magnetic Tunnel JunctionsMIT
Center forExcitonic DevicesUC-San DiegoMIT UC-Santa BarbaraPrinceton
Energy Efficient Learning Machines (ENIGMA)UC-Berkeley Stanford
A Fast 70mV Transistor Technology for Ultra-Low-Energy ComputingUC-Santa BarbaraU. VirginiaPurdue
Energy Efficient Computing with Chip-Based PhotonicsColumbiaMITStanfordUC-San Diego
Self-Adaptive Reservoir Computing with Spiking Neurons: Learning Algorithms and Processor ArchitecturesTexas A&M
Target a 100X reduction or more in energy per delivered operation as compared to projected performance of conventional CMOS
architectures and deeply scaled technology at the end of the roadmap
Devices
Systemarchitectures
Interconnect
Circuits
Materials
Energy-Efficient Computing fromDevices toArchitectures
VerticalE2CDACenterExamples:EXCEL&ENIGMA
28
computation complexity of dynamical systems, pursue physical hardware demonstration and quantify their efficacy in solving computationally hard problems that are finding ever expanding applications in high-performance data centers, real-time cyber-physical systems and computational medicine.
Over the next decade (until 2025), energy efficiency gains are expected from improved device technologies like fully depleted FinFETs, steep slope FETs, embedded non-volatile memory, circuit design techniques like fine-grained DVFS, heterogeneous integration solutions like 3D monolithic chips, special-purpose digital accelerators and architectural innovations like in-memory computing. Circa 2025, standby power will constitute a smaller fraction of the energy consumed, and active power will once again hold hostage energy efficiency due to slowdown in Moore’s Law (3nm node). Now is the time to rethink and reinvent the fundamental paradigm of computing to transcend the energy efficiency wall imposed by Boolean computing. This pursuit should be complemented by the realization that, in a fast evolving socially interconnected world, we are observing a seismic shift in the amount of unstructured data that need to be processed in real-time. Such processing does not necessarily conform well to the traditional von Neumann representation of the Turing Machine. Thus, we propose a hardware fabric in EXCEL based on the time evolution of coupled dynamical systems that will lead to the next-generation processor or coprocessor executing tasks that are computationally and, hence, energetically hard to solve today.
We draw inspiration from a multitude of examples of real time complex information processing occurring in natural dynamical systems, such as activation patterns of neural circuits, intra and extra-cellular signaling mechanisms, flow of information in social networks etc. Analytical solutions of such systems are rare and numerical simulations run for weeks on our fastest digital supercomputers due to the large number of data points in multiple dimensions. Yet the physical system itself computes its own dynamics in seconds! In this proposal, we show how to harness this extraordinary computing power in an experimentally accessible and scalable solid-state system at a chip scale. We show empirically how a coupled network of oscillators exhibiting electronic phase transitions inherently holds multi-dimensional spatio-temporal information, and how their synchronization dynamics computes the evolution of this large amount of information with unprecedented parallelism, higher speed to task completion and extreme energy efficiency. We harness the same physical hardware platform in conjunction with low voltage stochastic synaptic memory to implement dynamics of several classes of deep learning neural networks relevant in machine learning applications including spike-based reservoir computing.
To achieve our goals, we have assembled a multi-disciplinary group of computer scientists, mathematicians, architects, integrated circuit designers, device physicists and material scientists into a collaborative research center, EXCEL. (Fig. 2.)
We utilize insulator-to-metal electronic phase transitions in complex oxides arising from electron-electron and electron-lattice interactions to construct low power digital FETs, relaxation oscillators and spiking neurons. Doped oxides and/or layered oxy-chalcogenides are utilized for realization of analog memories with extremely low programmable voltages. The dynamics of phase synchronization with interacting oscillators enables implementation of various classes of convex and combinatorial optimization solvers. Intrinsic stochasticity of nanoscale phase transition memories enables implementation of probabilistic spiking neural networks (similar to stochastic counterpart of Hopfield networks) that offer significant
Figure 2: Programmatic overview of EXCEL research S.Datta,EXCELcenter reviewpresentation,2017
EXCEL:EXtremely EnergyEfficient CollectiveELectronics ENIGMA:Energy Efficient Learning Machines
37XEDPimprovement3Xareareduction
29
Energy-Efficiencyvs.ComputingPrecision
• Oscillatory behavior• Analog modulation• Nonvolatile • Stochastic behavior• Low voltage/power• Small size• ……
Energy-efficient, data-intensive, cognitive applications
(low-precision)
High-performance scientific computing (high-precision)
Application space
q Isthereafundamentaltradeoffbetweenenergyefficiencyandprecision?
q Cantheseapplication-specificapproacheshelpHPC?
• Massively parallel • Highly scalable • Co-located data
storage & computing • Fault tolerant • Energy efficient• ……
Emerging devicecharacteristics
Novelarchitecturefeatures
Manyapproaches achieveenergyefficiency• For specificapplications• For finiteprecision requirement
Applications
Energy Delay
Area Precision
30
ExtendBenchmarkingtoNon-BooleanComputing
Benchmarkingchallengesfornon-Booleancomputing:• Largevarietyofapplications, models, algorithms, implementations …• Numerous technology options withvarying levelofmaturity• Newfigures-of-merit andcriteria
Spindiffusion
Domainwall
SpinHalleffect
SpindiffusionDomainwall
SpinHall effect
AdvantagesofSpintronics forCNN
LessonsLearnedfromtheNRIProgram
• Nobeyond-CMOSdevicehasbeenproventobecapableofreplacingCMOSforBooleanlogicandvonNeumannarchitectures.□ IfthereisaCMOSsolution,thedifficultyandrequirementofadoptingabeyond-CMOStechnologywillbesignificantlyhigher.
□ Therearestillinterestinglow-powerdevicemechanismsworthexploring□ Beyond-CMOSdevicesmayfindpromisingopportunitiesfornovelcomputingparadigmsbeyondBooleanlogicandvonNeumannarchitectures.
• Novelmaterialandprocessingareessentialforbeyond-CMOSresearch.□ Thereissignificantmaterialandprocessingbarrierformostbeyond-CMOSdevices.□ Needtofocusandaddressthehardproblem.
• Aholisticapproachisneededtoaddressnewbeyond-CMOSresearchdirections.□ Benchmarkingisessentialandneedstostartfromthebeginning.
• Public-privatecollaborationiscriticalforfundingthebasicresearchneededtoadvancethesemiconductorandcomputingtechnologies.
31
TheValueofResearchInvestment
• TheimportanceofscientificresearchshouldnotbeevaluatedsolelybasedonthemonetaryROI(returnoninvestment),butshouldalsoconsidertheRONI(riskofnoinvestment).
• Industryinvestmentisinevitablydrivenby“return”.□ Governmentfundingandpolicyguidanceareneededtodrivebasicresearch.
• Consortiummodelreducesrisk,leveragesresources,andpromotescollaboration.□ Achallenge:“sharedmind”inthepost-scalingandbeyond-CMOSera
32
Public-PrivateCollaborationThroughSRC
IP
Research
NewTalent
ü TechnologyTransferü BestandBrightestü Create Solutions
SRCMemberCompanies
GovernmentAgencies
ü 1700 Studentsü 600Facultyü 120Universities
SRCin2016
Commercialization
Amechanism foreffectivepublic-private partnership&pre-competitive research
SRCConnection
ü IndustryRelevant Challengesü CompetitorswithCommonGoalsü Leveraged Researchü Gov’t&IndustryConnections
BestUniversities
33
UniversityOutput
Outline
• Introduction• Beyond-CMOSresearchinNanoelectronics ResearchInitiative(NRI)
q NRIresearchscopeq HighlightsofNRIachievements
• LessonslearnedfromtheNRIprogram• What’snext?• Summary
34
NewScienceTeam(NST)https://www.src.org/about/nst/
Ext. CMOS
BeyondCMOS
RF, Analog
$$
$ $
JUMP nCORE
$
NST
35
Joint University Microelectronic Program(DARPA)1. RF to THz Sensor and Communication Systems2. Distributed Computing and Networking3. Cognitive Computing4. Intelligent Memory and Storage5. Advanced Architectures and Algorithms6. Advanced Devices, Packaging, and Materials
nanoelectronic COmputing REsearch program(NSF, NIST, & Other Gov’t Agencies such as DOE*)I. Novel Computing ParadigmsII. Device, Interconnect, and Materials ResearchIII. Advanced Manufacturing and NanofabricationIV. Innovative Metrology and CharacterizationV. Computational models
*Ongoing talks,noconcreteplansyet!
NRIandSTARnet bothendin2017NRI nCORESTARnet JUMP
36
JUMPProgramSixJUMPCentersRepresentSixJUMPThemes
Center1
Center3
Center2
Center4
RFtoTerahertzSensors and
CommunicationSystems (V)
DistributedComputingandNetworking(V)
CognitiveComputing(V)
IntelligentMemory andStorage (V)
AdvancedArchitectureandAlgorithms(H)
AdvancedDevices,Packaging,andMaterials (H)
Center6
Center5
Systems-Level
Circuit-Level
Ext.CMOS
BeyondCMOS
RF,Analog
JUMPreceived18fullproposalsrepresenting 304professors at47USuniversities
37
nCORE ResearchScope5researchvectorsdefine vertically integrated research
I. Novelcomputingandstorageparadigms,andtheoryofoperation,beyond conventional CMOSdevices,beyond vonNeumann architectures, andbeyond classicalinformation processing andsensing
II. Fundamentalmaterial, device,andinterconnectresearch toenable novel computing andstorageparadigms III. Advancedmanufacturingandnanofabrication toenable thefabrication ofemerging devicesandsystemsIV. Innovativemetrologyandcharacterization tosupport basicdeviceandmaterial research,andtestplatforms
andstandards tobenchmark performance from devicesuptosystems V. Computationalmodels tosupport basic researchfrom emerging devicesandmaterials tonovel systems
Novel&computing¶digms
Device,&interconnect,&&&material&research
Enable
Characterization&/metrology=Test&platforms/standards=
Advanced&manufacturing
Support
Modeling
Support
Support
I
II IVIII
V
AnIntegratedResearchStructure
Summary:LessonsLearnedfromtheNRIProgram
• Nobeyond-CMOSdevicehasbeenproventobecapableofreplacingCMOSforBooleanlogicandvonNeumannarchitectures.□ IfthereisaCMOSsolution,thedifficultyandrequirementofadoptingabeyond-CMOStechnologywillbesignificantlyhigher.
□ Therearestillinterestinglow-powerdevicemechanismsworthexploring□ Beyond-CMOSdevicesmayfindpromisingopportunitiesfornovelcomputingparadigmsbeyondBooleanlogicandvonNeumannarchitectures.
• Novelmaterialandprocessingareessentialforbeyond-CMOSresearch.□ Thereissignificantmaterialandprocessingbarrierformostbeyond-CMOSdevices.□ Needtofocusandaddressthehardproblem.
• Aholisticapproachisneededtoaddressnewbeyond-CMOSresearchdirections.□ Benchmarkingisessentialandneedstostartfromthebeginning.
• Public-privatecollaborationiscriticalforfundingthebasicresearchneededtoadvancethesemiconductorandcomputingtechnologies.
38