trends in integrated circuits technology - stanford · pdf fileworld ic market by technology...
TRANSCRIPT
1
EE311/ Trends1 tanford Universityaraswat
Prof. Krishna Saraswat
Department of Electrical EngineeringStanford UniversityStanford, CA 94305
Trends in Integrated CircuitsTechnology
EE311/ Trends2 tanford Universityaraswat
Courtesy Prof. Tsu-Jae King
(Sources: VLSI Research Inc.; UnitedNation yearbook; World BankDatabase; IMF)
Miniaturization => Market growthMiniaturization => Market growth
Technology Scaling
Investment
Market Growth
Better Performance/Cost
Semiconductors havebecome increasingly moreimportant part of worldeconomy
2
EE311/ Trends3 tanford Universityaraswat
World IC Market by Technology
Silicon CMOS has become the pervasive technology
Ref: Chang and Sze, ULSI Technology, 1996
EE311/ Trends4 tanford Universityaraswat
Moore’s Law
108
107
106
105
104
103
102
101
100
109
1010
’60 ’65 ’70 ’75 ’80 ’85 ’90 ’95 ’00 ’05 ’10
Transistors Per DieTransistors Per Die
1011
1K1K4K4K 16K16K
64K64K256K256K256K
1M1M1M
16M16M4M4M
64M64M64M
40044004
80808080808080868086
8028680286i386™i386™
i486™i486i486™™PentiumPentium®®
MemoryMemoryMicroprocessorMicroprocessor
PentiumPentiumPentium®® IIIIPentiumPentium®® IIIIII
256M256M512M512M512M
PentiumPentium®® 4ItaniumItanium®®
1G1G2G2G 4G4G
128M128M
IC Performance to Gordon Moore’s Prediction
Source: Intel
3
EE311/ Trends5 tanford Universityaraswat
Feature Size Trend
0.001
0.01
0.1
1
1980 1990 2000 2010 2020
micron
Generation
LGATE
Gate length is not true measure of transistor size
EE311/ Trends6 tanford Universityaraswat
Generation:
Intel386™ DXProcessor
Intel486™ DXProcessor
Pentium® Processor
Pentium® II Processor
1.5µ 1.0µ 0.8µ 0.6µ 0.35µ 0.25µ
Example: Microprocessor Evolution
4
EE311/ Trends7 tanford Universityaraswat
MOS Device Scaling
Why do we scale MOS transistors?1. Increase device packing density ~ α2
2. Improve frequency response (speed) ~ a3. Power/ckt: ~1/α2, power density constant4. Improve current drive (transconductance gm)
!
gm ="ID
"VG VD = const
#W
Lµn
Ko x
to xVD for VD <VDSAT
, linear region
#W
Lµn
Ko x
to xVG $ VT( ) for V
D>VDSAT
, saturation region
Constant E Field ScalingAll device parameters are scaled bythe same factor α.
• Gate oxide thickness tox ↓• Channel length L ↓• Source/drain junction depth Xj ↓• Channel doping ↑• Supply voltage VD ↓
EE311/ Trends8 tanford Universityaraswat
Intel’s Transistor Research down to 10nm
65nm process65nm process2005 production2005 production
30nm30nm
45nm process45nm process2007 production2007 production 32nm process32nm process
2009 production2009 production
15nm15nm
22nm process22nm process2011 production2011 production
Source: Intel
10nm10nm
DNA is 15 nm wideDNA is 15 nm wide
20nm20nm
Electronics is NanotechnologyElectronics is Nanotechnology
5
EE311/ Trends9 tanford Universityaraswat
Speed increases as a result of scaling
Source: Mark Bohr, Intel
0.01
0.1
1
10
100
0.001 0.01 0.1 1LGATE (um)
Gate
Delay
(ps)
Limit
EE311/ Trends10 tanford Universityaraswat
Physical Limits in Scaling Si MOSFET
Substrate
Gate
Source Drain
Gate stack• Tunneling current ⇒ Increased Ioff• Gate depletion ⇒ Increased EOT
Source/Drain• Contact resistance • Doping level, abruptness
Channel• Surface scattering - the “universal mobility” tyranny• DIBL ⇒ drain to source leakage• Subthreshold slope limited to 60mV/decade (kT/q) ⇒ Increased Ioff• VG - VT decrease ⇒ reduced ION
Net result: Bulk-Si CMOS device performance increase commensurate withsize scaling is unlikely beyond the 70 nm node
High E-Field• Mobility degradation• Reliability
te
6
EE311/ Trends11 tanford Universityaraswat
MOSFET Scaling Limit: Leakage
Lo et al.,IEEE EDL, May 1997.
Gate Leakage S/D Leakage
Source: Marcyk, Intel
Total Leakage Trend
Total Power Trends
Ability to control Ioff will limit gate-length scaling– Thermionic emission over barrier– QM tunneling through barrier– Band-to-band tunneling from body to drain
To suppress D/S leakage, need to use:– Higher body doping to reduce DIBL
⇒ lower mobility, higher junction capacitance, increased junction leakage– Thinner gate dielectric to improve gate control ⇒ higher gate leakage– Ultra-shallow S/D junctions to reduce DIBL ⇒ higher Rseries
EE311/ Trends12 tanford Universityaraswat
45
MOSFET Scaling Problem: Saturation of Saturation of IIDsatDsat
200
400
600
800
0.1 0.2
0.25
0.3 0.4 0.6 0.8 1.0
NMOS
PMOS
IDsat (A/m) (drive current)
Channel Length (µm)
• Data from IBM, TI, Intel, AMD, Motorola andLucent
• Low OFF current desirable
0
0.4
0.8
1.2
1990 1995 2000 2005
1
10
Supply
Voltage (
V)
Drive C
urr
ent (m
A/µ
m)
Changhoon Choi, PhD Thesis, Stanford Univ., 2002
Constant OFF current Limit Relaxed OFF current Limit
Source: Intel
Low Vt
High Vt
IOFF,high Vt
IOFF,low Vt
Vg
log Id
0
7
EE311/ Trends13 tanford Universityaraswat
!
Eeff =q
"Si(Ndep +# $ NChannel ) Ndep= depletion charge density
NChannel = charge induced in the channel
Increases in substrate doping ⇒ Ndep ⇑ Gate oxide thickness decrease ⇒ NChannel ⇑ Eeff increases with scaling ⇒ µ ⇓ Reduced gate oxide thickness increases remote charge scattering ⇒ µ ⇓ High k dielectrics have higher coulombic scattering due to surface states
and phonon scattering ⇒ µ ⇓
Coul
ombi
c Phonon
SurfaceRoughness
µ
Eeff
Effects of Scaling Bulk MOSFET on Mobility
S. Takagi et al., IEEE TED, 41 (1994) 2357. srphCeff µµµµ
1111++=
EE311/ Trends14 tanford Universityaraswat
New Structures and Materials forNanoscale MOSFETs
1. Electrostatics - Double Gate - Retain gate control over channel - Minimize OFF-state drain-source leakage2. Transport - High Mobility Channel - High mobility/injection velocity - High drive current for low intrinsic delay3. Parasitics - Schottky S/D - Reduced extrinsic resistance4. Gate leakage - High-K dielectrics - Reduced power consumption5. Gate depletion - Metal gate
1
23
GG
Si
S D Si
SiO2
C
BULK SOI Double gate
Bottom Gate
Top Gate
Source Drain
High µchannel High-K
45
8
EE311/ Trends15 tanford Universityaraswat
Combining New Device Structures withCombining New Device Structures withNew MaterialsNew Materials
We will be here withthese innovations
We arehere today
• With better injection and transport we may be able toimprove MOSFET ION
• With better electrostatics we may be able to minimize Ioff
EE311/ Trends16 tanford Universityaraswat
0.001
0.01
0.1
2000 2010 2020
micron
1
10
100
nm
45 nm
65 nm
32 nm
22 nm
16 nm
11 nm
8 nm
Generation
LGATE
Evolutionary
CMOS Exotic Revolutionary
CMOS
Nanotechnology Eras
ReasonablyFamiliar
NanotubesNanowires
ReallyDifferent
Source: Mark Bohr, Intel
9
EE311/ Trends17 tanford Universityaraswat
Scaling of MOS Gate Dielectric
(Ref: S. Asai,Microelectronics Engg., Sept. 1996)
Gate SiO2 thickness is approaching < 10 Å to improve device performance• How far can we push MOS gate dielectric thickness?• How will we grow such a thin layer uniformly?• How long will such a thin dielectric live under electrical stress?• How can we improve the endurance of the dielectric?
ID ! gm !K
thickness
EE311/ Trends18 tanford Universityaraswat
•Below 20 Å problems with SiO2– Gate leakage => circuit instability, power dissipation– Degradation and breakdown– Dopant penetration through gate oxide– Defects
Problems in Scaling of Gate Oxide
Defects and
nonuniformity of film
Dielectric breakdown
Reliability due to
charge injection
Si substrate
Polysilicon gate electrode
Dopant
penetration
gate oxide
Leakage current
10
EE311/ Trends19 tanford Universityaraswat
Dielectric Degradation
What are the mechanisms for damage and breakdown?How can we engineer the gate dielectric to minimize the damage?
• Degradation during device operation due to high E field causing current injection• Degradation during fabrication due to charging in plasma processing
cathode
e
Anode
oxide
Eox
N(E)
E
EE311/ Trends20 tanford Universityaraswat
0.00001
0.001
0.1
10
1000
1801301007050
Technology Generation (nm)
Cur
rent
(µA
/µm
)
Igate
Ioff
Ion
Source G. Bersuker, et al. Sematech
Gate Oxide Scaling Issues: Leakage
• Ion is not increasing with scaling • Igate ⇑, power dissipation ⇑• Circuit instability
11
EE311/ Trends21 tanford Universityaraswat
High-k MOS Gate Dielectrics
K ≈ 20
Historically Cox has been increased by decreasing gate oxidethickness. It can also be increased by using a higher K dielectric
Si3N4 K ≈ 8
40 ÅToday Near future
Long term
20 Å SiO2 K ≈ 4!
ID"C
ox"
K
thickness
100 Å high K
Si
Higher thickness -> reduced gate leakage
Ichannel ∝ charge x source injection velocity∝ (gate oxide cap x gate overdrive) vinj∝ CCoxox (VGS - VT) Esource µµinjinj
EE311/ Trends22 tanford Universityaraswat
Perkins, Saraswat and McIntyre,Perkins, Saraswat and McIntyre,StanfordStanford Univ Univ. 2002. 2002
Capacitance and Leakage for High-k GateDielectric Films Grown Using ALCVD
Gat
e Le
akag
e (A
/cm
2 )
10-10
10-8
10-6
10-4
10-2
100
0 0.05 0.1 0.15 0.2
Leakage (A/cm
2) @ V
FB
± 1 V
1/C'
ox (µm
2/fF)
SiO2
4 nm
2.5 nm
ALCVD ZrO2
Gat
e C
urre
nt@
VFB
+1V
(A/c
m2 )
Equivalent SiO2 Thickness (nm)
Silicon Germanium
Chui, Kim, Saraswat and McIntyre,Chui, Kim, Saraswat and McIntyre,StanfordStanford Univ Univ. 2004. 2004
12
EE311/ Trends23 tanford Universityaraswat
!s = f VG( ) =1
"VG
!
ID"QI " eqVG /# kT
!
QI " eq#s /kT
!
" =1+1
2 # C ox
q$sNa
|%p |Where
Subthreshold Behavior
source
VG
• Diffusion of carriers over thebarriers to the channel.
• Fermi-dirac distribution ofcarriers: e-E/kT
• Gate reduces the barrier tocurrent flow.
Fermi Diracdistribution
draine
VT
VDD
ID
(log scale)
VG
ILEAK
or
IOFF
VDD- V
T
60 mV/decade60 mV/decade
ION
ON
OFF
(ln )
G
D
VS
I
!=!
EE311/ Trends24 tanford Universityaraswat
In devices with long channel lengths, the gate is completely responsible fordepleting the semiconductor (QB). In very short channel devices, part of thedepletion is accomplished by the drain and source bias
Since less gate voltage is required to deplete QB, VT↓ as L↓. Similarly, as VD
↑, more QB is depleted by the drain bias, and hence VT↓. These effects areparticularly pronounced in lightly doped substrates.
Effect of Reducing Channel Length
junctiondepletion
region
poly gate
n+ n+
p-substrate
poly gate
n+ n+
p-substrate
depleted by gate charge
VT
Drawn Channel Length, L
VT
Supply Voltage, VD
13
EE311/ Trends25 tanford Universityaraswat
p-well
poly gate
n+ n+
VDD
Drain-Induced Barrier Lowering (DIBL)• Exacerbates subthreshold leakage in short-channel devices• Soft punchthrough induced by drain-to-substrate depletion region
• |VT | ↓ as VD ↑ [drain-induced short channel effects (SCE)]• VD ↑ drain-to-substrate depletion region grows with more reverse bias• Lateral electric fields in drain-induced depletion region lowers source-to-
channel barrier, allowing more carriers to diffuse from source to channel
reduction of electron barrierheight in conduction band (CB)
at edge of source
CB
VDD
source drain
p-well
poly gate
n+n+
EE311/ Trends26 tanford Universityaraswat
Q depletedby source
Q depletedby drain
B B
N+ source N+ drain
Gate
P-Si
Depletion region
L!
L
rj
VT = VFB ! 2 "#F !QB
Cox" 1 ! 1+
2 "W
rj!1
$
% &
'
( ) "rj
L
*
+ ,
-
. /
•Roll-off in threshold voltage as the channel length is reduced•VT roll-off is reduced as junction depth(rj) is decreased•Sheet resistance increases as junction depth is reduced
L. Yau, Solid-State Electronics, vol. 17, pp. 1059, 1974
Why do we need to scale junction depth?
14
EE311/ Trends27 tanford Universityaraswat
Source: Jasonn Woo, UCLA
30 nm 50 nm 70 nm 100 nm0
20
40
60
80
100
120
140NMOSScaled by ITRS Roadmap
Rc
Rdp
Rext
Rov
Physical Gate Length
Serie
s R
esis
tanc
e (o
hms)
Rc RdpRext
Rov
x
y = 0
GateSiO2
Metal
Next(x)
Nov(y)
Source/Drain Resistance
Problem in junction scaling:• Sheet resistance of a junction is a strong function of doping density• Maximum doping density is limited by solid solubility and it does not scale• Silicidation can minimize the impact of junction sheet resistance (Rs,Rd)• Contact resistance Rc is one of the dominant components for future technology
EE311/ Trends28 tanford Universityaraswat
Contact Resistance
Specific contact resistivity
Doping density
(Ref: S. Swirhun, PhD Thesis, Stanford Univ. 1987)
!c = !co exp2"Bqh
#sm*
N
$
% & &
'
( ) ) ohm * cm2
• Contact resistance is a strong function ofdoping density at the metal/silicon interface
• Solid sulubility of dopants does not scale !
PROBLEMS
15
EE311/ Trends29 tanford Universityaraswat
Solutions to Shallow JunctionResistance Problem
Extension implants Elevated source/ drain
Silicidation Schottky Source/Drain
EE311/ Trends30 tanford Universityaraswat
Problems with Poly-Si Gate.This occurs because of high E - field due to a combination of highersupply voltage and thinner gate oxide.
• Effect of depletion is to increase effective tox and thus reduce Cox• A reduced Cox implies reduction in gm and thus ID(on)• Ionized impurities in the gate electrode cause “remote charge scattering”
⇒ Reduced mobility
tox(electrical)
tox(physical)
Gate depletion
Oxide
Substrate
Poly-Si gate
Need metal gate electrode with proper workfunction
16
EE311/ Trends31 tanford Universityaraswat
Ultra-Thin Body Single Gate SOI
Ultra-Thin Body DoubleGate SOI
Evolution of MOSFET Structures
GateGate
Silicon Substrate
Source Drain
TBOX
Si
SiO2
SOI
Source Drain
Gate 1Gate 1 Vg
Tox
SOI
Gate 2Gate 2
Si
Ref: Philip Wong, IEDM Short Course, 1999
BULK
Advantages of Ultra-Thin Body SOI• Depleted channel ⇒ no conduction path
is far from the gate• Short channel effects controlled by
geometry• Steeper subthreshold slope• Lower or no channel doping• Higher mobility• Reduced dopant fluctuation
EE311/ Trends32 tanford Universityaraswat
Non PlanarNon Planar MOSFETs
UC Berkeley
Gate
Source Drain
Intel
Tri Gate FET
SourceSource DrainDrain
GateGate
SiO2
Channel
Double Gate FinFETVertical FET
Stanford, AT&T
17
EE311/ Trends33 tanford Universityaraswat
Transport: Effects of Biaxial Tensile Strain on Si Energy BandsHoyt, 2002
Conduction BandAdditional splitting:Band repopulation
- reduced intervalley scattering- smaller in-plane effective transport mass
Δ2
Δ4
[001]
[010]
[100]
Δ Es ~ 67 meV/ 10% Ge
Δ4
Δ2
Δ6Ec
Bulk Si Strained Si
µ = q!
m*c
Strained Si grown on Relaxed Si1-xGex
biaxial tension
HH
LH
Spin-Orbit
E EValence Band
HH/LH degeneracy lifted at Γ
- reduced interband scattering- smaller in-plane transport mass due to band deformation
Γ
Bulk Si Strained Si
in-planeout-of-plane
kk
Δ Es ~ 40meV/10% Ge
ml
mtmt
mt < ml
Single ellipsoid
EE311/ Trends34 tanford Universityaraswat
Strained Si gate oxide
Si Substrate
Relaxed Si1-xGex
n+ poly LTOspacer
n+ n+
Si1-xGex Graded layer
0.80
1.0
1.2
1.4
1.6
1.8
2.0
0.0 0.10 0.20 0.30 0.40
Mobility enhancement ratio
Substrate Ge fraction, x
VDS
= 10 mV
300 K
Measured, J. Welser, et al.,
IEDM 1994.
Calculated for strained Si
MOS inversion layer
S. Takagi, et al., J. Appl. Phys. 80, 1996.Mob
ility
Enh
ance
men
t Fac
tor
NMOS
Mobility Enhancements in Strained-Si MOSFETs
Intel
PMOS
Gibbons Group, Stanford
18
EE311/ Trends35 tanford Universityaraswat
Nanowire and Nanowire and Nanotube FETsNanotube FETsALD HfO2 Coated of Ge NW FET
sourcegate
drain
dielectric
semiconductor
~20nmmetal
~10nm
Channel
Au
Nanoparticle
GeContaining
Vapor
Ge Nanowire
Ge NW Growth
Key Challenge: Controlled growthCatalyst Support
CnHmCnHm Fe
Gate
HfO2
10 nm SiO2
p++ Si
SD
Carbob Nanotube MOSFETCarbob Nanotube Growth
EE311/ Trends36 tanford Universityaraswat
B
+ =
Spintronics
Seemingly Useful Devices
Limited Current DriveCryogenic operation
Limited Fan-OutCritical dimension control
Need high spin injectionand long spin coherence time Limited thermal stability
New architectures needed
~ 2 nm
Challenging fabricationand process integration
CarbonNanotubes
Controlled growth
19
EE311/ Trends37 tanford Universityaraswat
• In general this device scaling methodologydoes not take into account many other chipperformance and reliability issues, e.g.,interconnects, contacts, isolation, etc.
• These factors are now becoming an obstaclein the evolution of integrated circuits.
EE311/ Trends38 tanford Universityaraswat
Device Isolation pitch as a function ofminimum dimension
With decreasing feature size the requirement onallowed isolation area becomes stringent.
0.0 0.2 0.4 0.6 0.8 1.0Minimum dimension [µm]
0.0
0.5
1.0
1.5
2.0
2.5
16M
64M
1G
P. Fazan, Micron, IEDM-93
Activ
e Ar
ea p
itch
(µm
)
20
EE311/ Trends39 tanford Universityaraswat
Scaling of Device Isolation
NitridePad oxide
Fully recessed LOCOS
Nitride Pad oxide
Semi-recessed LOCOSNitride
Field oxide
After field oxidation
After field oxidation
LOCOS based isolation technologies have serious problemsin loss of area due to bird’s beak.
Deep trench isolation
N-wellP-substrate
Shallow trench isolation
Trench isolation can minimize area loss
EE311/ Trends40 tanford Universityaraswat
Scaling of interconnections
OldNew (scaled)
• Bigger chip => longer interconnects• Scaling to smaller dimensions => reduced cross section• Larger R, L and C
21
EE311/ Trends41 tanford Universityaraswat
Interconnect Delay Is Increasing
• Chip size is continuallyincreasing due toincreasing complexity– Increase in R, L and C
• Device performance isimproving but interconnectdelay is increasing
• Need better materials– Metal with lower resistivity– Dielectrics with lower K– Other solutions, e.g., 3D,
optical interconnects
60 80 100 120 140 160 18010-2
10-1
100
101
Technology Node (nm)D
elay
Tim
e (n
s) Longest Interconnect Delay
Typical Gate Delay
scaling
EE311/ Trends42 tanford Universityaraswat
1970’s Poly-Si gateAluminum
1980’s Aluminum alloysSilicide contactsPolycide gatesLocal planarization
1990’s Layerd aluminum/titanium SalicidesCVD tungsten plugsShallow trench isolationGlobal planarization
Advances in Backend Technology
22
EE311/ Trends43 tanford Universityaraswat
Current Interconnect Technologies
Copper 6
Copper 1
TungstenLocal Interconnect
Copper 3
Copper 5
Copper 4
Copper 2
Current Al technology(Courtesy of Motorola)
Current Cu technology(Courtesy of IBM)
EE311/ Trends44 tanford Universityaraswat
1
2
3
4
5
6
7
8
9
10
11
12
13
14
Al & SiO2 (! = 4)
Cu & SiO2 (! = 4)
Al & low-! (! = 2)
Cu & lo w-! ( ! = 2)
0.092007
0.132004
0.182001
0.251998
0.351995
Tec hnology Generation
µmYear
Why Cu and Low-k Dielectrics?
Reduced resistivity and dielectric constant results in reduction in number ofmetal layers as more wires can by placed in lower levels of metal layers.
global
semiglobal
local
Source: Y.Nishi
23
EE311/ Trends45 tanford Universityaraswat
Cu Resistivity: Effect of Line Width Scaling• Effect of Cu diffusion Barrier
• Barriers have higher resistivity• Barriers can’t be scaled below a minimum thickness
• Effect of Electron Scattering• Reduced mobility as dimensions decrease
• Effect of Higher Frequencies• Carriers confined to outer skin increasing resistivity
Problem is worse than anticipated in the ITRS 1999 roadmap
EE311/ Trends46 tanford Universityaraswat
Problems in Scaling of Interconnections
Pure Metal Interconnect
Layered Interconnect
Surrounded Interconnect
Al
Al
Cu
ρav
Minimum Feature Size (λ)
• Resistivity increases asgrain size decreases
• Resistivity increases asmain conductor sizedecreases but not thesurroundingbarrier size
AS λ DECREASESBarrier
Barrier
Cu
24
EE311/ Trends47 tanford Universityaraswat
3-D Integration: Motivation
• Integration of heterogeneous technologiespossible, e.g., memory & logic, optical I/O
• Reduce Chip footprint• Replace long horizontal wires by short
vertical wires• Interconnect length ⇓ and therefore R, L, C ⇓
– Power reduction– Delay reduction
2-D System 3-D System
Wire-length
Num
ber o
f Int
erco
nnec
ts
(Log-Log Plot)2-D IC3-D IC
Area = A
Very Long Wire
Shorter Wire
2D
3D
A/2
A/2
EE311/ Trends48 tanford Universityaraswat
3-D Motivation: Integration Density
Time
No.
Tra
nsis
tors
/chi
p //
Perf/
Func
tiona
lity
No.
Tra
nsis
tors
per
cm
3 in
sys
tem
12-15 years
The Best Integrators of Electronic Devices Will Own theHeart of Every System – We have <15 Years to Figure it out
2-D Batch
3-D Packaging
3D batch
End-of-Moore’sLaw!
Source: D. Radack, DARPA
Logicn+/p+
n+/p+ n+/p+
Gate
Gate
T1
T2
M1
M2
M3
M4
n+/p+n+/p+
GateRepeaters oroptical I/O devices
n+/p+
M’1
M’2
MemoryorAnalog
25
EE311/ Trends49 tanford Universityaraswat
Can Optical Interconnects help?40Tb/s Optical I/O1024 x OC-768100Tb/s On-ChipBisection BWPMM64 Tiles64b Processor+ 4MB DRAMOn-Chip Optical Interconnects
Chip-to-chip Optical Interconnects
Can potentially address many problems ofCu/low-k wires
On-Chip LinksReduce delay
Clocking and SynchronizationReduce jitter and skew
High Bandwidth off-chip Links Reduce power
EE311/ Trends50 tanford Universityaraswat
Result: scaling of power components
Power increasingly becoming the performance bottleneck for high-endmicroprocessors
• Dynamic Power: CV2f• Leakage power: devices• Short circuit power during switching• Static power, e.g., analog components (sense amps etc.)
ITRS projections for totalpower dissipation on chip
Chandra, Kapur and Saraswat,IEEE IITC, June 2002
26
EE311/ Trends51 tanford Universityaraswat
• Thermal conductivity of low-k insulators is poor
• Thermal impedance increases
• Energy dissipated (CV2f) is increasing as performance improves• Average chip temperature is rising
Thermal Behavior in ICs
Source: ITRS
3550 70 130 180100
0
1
2
3
4
0
0.2
0.4
0.6
0.8
1
1.2
Technology Node [nm]
Die
lect
ric C
onst
ant Therm
al Conductivity [
W / m
K ]
35 50 70 100 130 180
0
5
10
15
20
25
30
0
1
2
3
4
5
Pow
er D
ensi
ty [
W /
cm2
]
Technology Node [nm]
Jm
ax [ MA
/ cm2 ]
EE311/ Trends52 tanford Universityaraswat
>100A will flow on these wires
The problems Caused by Increased Power
As T ⇑ R ⇑, RC delay ⇑10°C ⇑ , Speed ⇓ 5%
PERFORMANCE
RELIABILITY
10°C ⇑ , MTF ⇓ 50%
!
MTF =A
rmJnexp
Ea
kT
"
# $
%
& '
Mean time to failure
Electromigration induced hillocks and voids
Hillock VoidMetal
MetalDielectric
27
EE311/ Trends53 tanford Universityaraswat
Source Drain
Gate
Time
Conclusion: Technology ProgressionBulk CMOS
Si0.8Ge0.2
Si
Si
(tensile)
Si1-xGex
Strained Si
High k gate dielectric
Metal gate
3D ICs
2 nm
Cu interconnect
Low-k ILD
FD SOI CMOS
Optical interconnect
Molecular device
Detectors, lasers,modulators, waveguides
Spin device
B
+ =
Ge on Si hetroepitaxy
Ge on Insulator
Wafer bondingCrystallization
Nanowires
Ge/Si Heterostrcture
Double-Gate CMOS
Nanowire
Single etransistor
Nanotube
Feature Size
EE311/ Trends54 tanford Universityaraswat
MOS Transistor in 2010 A Circuit in 2010 A Factory in 2010
Gate oxide thickness < 1nm Channel Length < 2nmJunction depth < 1-2nmSize of an atom ~ 5 Å
1010 components Integrated digital, analog, sensors
Approaching $10 billion
Summary
Questions we are trying to answer• How can we continue the Moore’s law• What will be new materials, devices, circuits, sensors,
equipment, simulators, etc.• How will we design them?• How will we manufacture them?