1. A Comparative Study of Delay Analysis for
Carbon Nanotube and Copper based VLSI
Interconnect Models
By
HARPREET SINGH BHATIA
Under the supervision of
MR. MAYANK K. RAI
Assistant Professor, ECED
DEPARTMENT OF ELECTRONICS AND COMMUNICATION ENGINEERING
THAPAR UNIVERSITY, PATIALA
3. Introduction
• Interconnects – these are wires connecting transistors.
• As technology scales down :
Types
•Local/short
•Intermediate / semiglobal
•Global
Local Global
[1] Effect of SCALING
2
RC delay: t int = (RL)(CL) = RC.L
•Density of interconnect >> Density of gates; L
•Interconnects constitute the main source of Delay. Local delay: 2
t int = (R. S )(C) 2
= RC.L
2
S
2 2 2
Global: t int = (R. S )(C) L = RC.L .S
4. Interconnect Models
Parasitic Inductance
L Parasitic Resistance L ≈
M f( h.w
L
)
R = ρ.L
h h.w
L≈ nH/cm
Xox W
R≈ kΩ/cm
Parasitic
Capacitance
Cpp = ε.w.l
Xox
C= Cpp +Cfringe
RLC Model
C≈ pF/cm
RC model
5. Factors affecting Interconnect Performance
• Repeater design
→
W/L driver
n, W/L load
3n →
• Dielectric C
A
k 0
X ox
• Pitch Pitch = w + s
• Length R RS
l But, C pp
ε dielectric
hl
w s
• Aspect Ratio h .l
AR C
w AR
6. Limitations of Copper
• Surface Scattering
MFP of e- in Cu = 40 - 100 nm
Polycrystalline are composed of
• Grain Boundary many crystallites/ grains. e- s scatter at
Grain-boundaries.
effect
Increase in ρo- from 1.9μΩ-cm to 4.9μΩ-cm
@ 45nm [25]
• Barrier Width
7. Carbon Nanotube (CNT)
Future interconnect material?
• Graphene sheet rolled into a tube. Formation eliminates dangling bonds
• Single Wall CNT (SWCNT) and Multiple Wall CNT (MWCNT)
Advantages CNT Cu Problems
Mean free path (nm) @ room >1000 40 1. High resistance ~6.45kΩ
temp
2. Contact resistance~100Ω
Max current density (A/cm2) >1x1010 ~1x106 Therefore, the NEED FOR
CNT BUNDLE
Thermal conductivity (W/mK) 5800 385
3. Lack of control on chirality
8. CNT Model
For L< λCNT or MFP:
Fundamental Resistance Magnetic Inductance Kinetic Inductance
RF = h/4e2 = 6.45 KΩ LM = μ ln(y/d) LK = h/4e2vF
Since each nanotube has 2π •Only for L< λCNT
four conducting •Not observed upto 10GHz freq
channels in parallel
(N=4)
[16]
Ground-Plate Quantum Capacitance
Capacitance CQ = 2e2
CE = 2πε
hvF
ln(y/d) Quantum electrostatic energy
stored in the nanotube when
it carries current.
4 CQ for 4 channels
→
9. CNT v/s Cu
LOCAL INTERCONNECTS
ρ.l
R Cu small l small R
A
h L h
R CNT For L<Lo R CNT 2
=6.45 KΩ
4e
2
L0 4e
GLOBAL INTERCONNECTS
For Copper: large l large R Cu
R CNT grows more slowly than Copper because Rcontact is a constant
resistance
10. Analytical Delay Estimation
1. Driver Interconnect Model (DIL)
2. Modified Nodal Analysis (MNA) and Moment
Matching
3. Unified Time Delay Model
11. 1. Driver Interconnect Model (DIL)
RLC interconnect driven a CMOS driver • Region 1 : During this region,
the nMOS is cut off.
uses α-power law id iL 0
MOS model →
• Region 2 : During this region,
the nMOS operates in
saturation t
in id iL 0 V in V DD
τ
• Region 3 : During this region
V in V DD and nMOS is still in
saturation
• Region 4 : During this region,
[20]
and nMOS transistor operates
in linear region.
12. DIL Results
COPPER CNT
1.2 1.2
1 1
0.8
Voltage (V)
0.8
Voltage (V)
0.6 0.6
0.4 0.4
0.2
0.2
0
0
0 0.912 1.9 2.9
0 0.912 1.9
time (ns) spice
time (ns) spice
Analytical Analytical
13. 2. Modified Nodal Analysis (MNA)
Solving the above circuit through MNA, we obtain these two equations:
C 0 v c (t) Y E
M x (t) = - G x (t) + P u (t) where, M = x (t) = G = T
0 L i L (t) -E R
y (t) Q x (t)
Using Laplace transform we solve the equations in s-Domain
y(s) -1 -1 1
H (s) = = Q (1 + s A ) B where, A G M B G P
u(s)
Performing the inversion on matrix (1 + s A ) is complicated. Reduction is done by Matrix
Approximation (PVL and Arnoldi approximation)
14. Moment Matching
Maclaurin expansion of the transfer function (Taylor’s expansion around s=0)
The coefficients of this expansion are known as central moments
Focusing on the first moment
If y(t) is a step function, then this voltage-time area is equal to its 50% delay →
15. Moment Matching example
Rmos R1 L1 R2 L2
6.631k 1037.82k 884.5pH =R1 =L1
+ C2
Vs Cmos C1
1V 2.1fF =C1
0.0053pF
-
Using MNA and PVL approximation, we calculate the transfer function H(s) upto 3 terms:
-8 -16 2
H ( s ) = 1 - (1.656 × 10 ) s + (2.743 × 10 )s
Using Moment Matching of Maclaurin Series, we calculate the first moment m(1) =16.56 ns.
i.e. 50 % delay = 16.56 ns.
SPICE Simulation Results show this to be 13.59 ns. (error of 21.86 %.)
16. 3. Unified Time Delay Model
Meindl et al. (2001 [6-9]) presented simplified delay expressions in a series of four papers.
Unified Time Delay for distributed RLC lines
t d = max ( t d,rlc , t d,rc )
2
t d = [max( t f ,0 .377 rcL + 0 .693 R tr cL ) ] + [ 0 .693 .C L ( rL+ 0 .65 R tr + 0 .36 Z c ) ]
Unified Time Delay for repeater insertion
t d = max ( t d,rlc _ rep , t d,rc _ rep )
2
rcL R c cL
t d = [max( t f , 0 .377 + 0 .693 ) ] + [ 0 .693 .C C ( hrL+ 0 .65 kR c + 0 .36 hkZ c
)]
k h
22. Conclusion
• In the 45 nm node, CNT bundle interconnects show significant
improvement in delay performance as compared to copper
interconnects for the following cases-
– CNT offers a better reduction in delay when the pitch ratio
is 1:1.
– CNT gives a better delay reduction than copper when
repeater driver transistor W/L ratio above 80 (29.4 % to
36.43 %).
• When estimating the delay analytically, a tradeoff needs to
be made between the computational efficiency and accuracy.
– Unified Time Delay Model has simplified expressions but
higher inaccuracies.
– Driver Interconnect Model (DIL) is very accurate, but
requires higher computational efficiency.
23. References
1. H. B. Bakoglu, Circuits, Interconnections and Packaging for VLSI, Addison-Wesley, Reading, MA, 1990
2. Bakoglu, H.B. and Meindl, J.D. “Optimal interconnection circuits for VLSI”, IEEE Transactions on Electron
Devices, (1985), Vol. ED-32 No. 5, pp. 903-9.
3. P. Kapur, J.P. Vittie and K. C. Saraswat, “Technology and Reliability Constrainted Future Copper Interconnects-Part I”
IEEE Transactions on Electron Devices, (2002)
4. El-Moursy, M.A. and Friedman, E.G. “Optimum wire sizing of RLC interconnect with repeaters”, Integration, the VLSI
journal (2004)
5. Bakoglu, H.B. and Meindl, J.D. “Optimal interconnection circuits for VLSI”, IEEE Transactions on Electron Devices, (1985)
6. Davis, J.A., Meindl, J.D., Compact distributed RLC interconnect models—Part I: single line transient, time delay and
overshoot expressions, IEEE Trans. Electron Dev. 47 (2000) 2068–2077.
7. J.A. Davis, J.D. Meindl, Compact distributed RLC interconnect models—Part II: coupled line transient expressions and
peak crosstalk in multilevel interconnect networks, IEEE Trans. Electron Dev. 47 (2000) 2078–2087.
8. R. Venkatesan, J.A. Davis, J.D. Meindl, Compact distributed RLC interconnect models—Part III: transients in single and
coupled lines with capacitive load termination, IEEE Trans. Electron Dev. 50 (2003) 1081–1093.
9. R. Venkatesan, J.A. Davis, J.D. Meindl, Compact distributed RLC interconnect models—Part IV: unified models for time
delay, crosstalk, and repeater insertion, IEEE Trans. Electron Dev. 50 (2003) 1094–1102.
10. Chandel, R., Sarkar, S. and Agarwal, R.P. “Repeater insertion in global interconnects in VLSI circuits” (2005).
24. References
14. F. Kreupl, et al., “Carbon Nanotubes in Interconnect Applications,” Microelectronic Engineering, 64 (2002)
15. Sakurai, T. and Newton, A.R. “Alpha power law MOSFET model and its applications to CMOS inverter delay and other
formulas”, IEEE Journal of Solid State Circuits, (1990)
16. P J Burke, Luttinger Liquid Theory as a Model of the Gigahertz Electrical Properties of Carbon nanotubes ; IEEE
Transactions on Nanotechnology, Vol 1, no 3, 2002.
17. C. Ho, , A.E. Ruehli, P. A. Brennan, The Modified Nodal Approach to Network Analysis, IEEE Transactions on Circuits
and Systems, 1975
18. A.B. Kahng, S. Muddu, Efficient gate delay modeling for large interconnect loads, IEEE Multi-Chip Module Conf.
(1996)
19. Y. I. Ismail and E. G. Friedman, “Sensitivity of Interconnect Delay to On-Chip Inductance”, ISCAS (2000)
20. B.K. Kaushik et al., S. Sarkar, R.P. Agarwal, Waveform analysis and delay prediction for a CMOS gate driving RLC
interconnect load, INTEGRATION, the VLSI journal 40, 2007, pp. 394–405.
21. Shyh-Chyi Wong, Winbond TSM, “Estimation of Wire Parameters” IEEE, Proc. Feb 2000.
22. C. Thiruvenkatesan, J. Raja, “Studies on the Application of Carbon Nanotube as Interconnects for Nanometric VLSI
Circuits”, ICETET-09, (2009)
25. W. Steinhogl, et al., “Size-dependent Resistivity of Metallic Wires in the Mesoscopic Range,” , (2002).
26. APPENDIX: α-power law MOS model
ID = 0 ; V GS V T0 : cutoff region
( /2)
I D = k 1 (V GS - V T0 ) .V DS ; V DS < V (DS- sat) : linear region
I D = k S (V GS - V T0 ) ; V DS V (DS- sat) : saturation region
channel length is decreased
lateral electric field EY increases
drift velocity vd α electric field => for electric fields of E >105 V/cm, velocity saturates vd(sat) = 107 cm/s
saturation-mode current is no longer a quadratic function of VGS ←
27. APPENDIX: RC Models
R/3 R/3 R/3
Pi- Model
• Delay of Pi-Model C/2 C/2
= RC/3+RC/6 = RC/2
agrees with distributed model RC.
• 3-segment pi-model is accurate to 3% in simulation
• L-model needs 100 segments for same accuracy!
• Pi Model is often used in Spice instead of large number of segments as a
reasonable approximation of distributed RC. ←
28. APPENDIX: Repeater insertion
• Repeaters are buffers or inverters inserted at regular intervals.
• RC delay is proportional to l2
• Now, Delay linearly proportional to l (Bakoglu and Meindl Model →)
2
Td 2 . 3 R o C int R int C o Wire Length: l
Driver Receiver
Td 2 . 3 R o C int N Segments
Segment
l/N l/N l/N
← Driver Repeater Repeater Repeater Receiver
29. Repeater design (Bakoglu and Mendl)
Minimum size repeaters
• W/Ldriver = 1, W/Lload = 3
Optimal repeaters
•Increasing size of repeaters to improve propagation time
•Increase W/L by h, R becomes Ro/h, C becomes Co.h
•Area penalty + Power Penalty
Cascaded drivers
•Sequence of drivers that increase gradually in size
•Used to drive large capacitive loads
•Power penalty ←
30. Bakoglu and Meindl Model
RC Load C Load
Ro Rint, Cint
For a wire with k repeaters each
of size h times minimum size
Co
inverter is given by:
Ro C int R int C int
T 50 % k 0 .7 hC o 0 .4 0 . 7 hC o
h k k k
31. Bakoglu and Meindl…
• By setting dT/dk = 0 and dT/dh = 0, “optimal” values for k and
h are obtained
0 . 4 R int C int R o C int
k h
0 .7 R o C o R int C o
• Substituting these back, delay is given by
2
Td 2 . 3 R o C int R int C o
32. Bakoglu and Meindl…
• For long-distance interconnections, Cint is on the order of
picofarads and Co is on the order of femptofarads, and Rint and
R, have values around kilohms;
Therefore, RoCint>>RintCo, and the delay expression can be
further simplified to
T =2.3RoCint
• As a result, repeaters can effectively "transform" the RC
interconnection load into a capacitive load.
• Delay of a capacitive load is linearly proportional to l.
←
33. APPENDIX: Parameters affecting
interconnect performance
• Dielectric • Parallel plate equation: C = εA/ Xox
• Width • Dielectric constant
of Interconnect – ε= kε0
– ε0 = 8.85 x 10-14 F/cm
• Diffusion Barrier k = 3.9 for SiO2
–
• To reduce capacitance, we need to use low-
• Length k dielectrics
• Pitch – k 3 (or less)
• Aspect Ratio • But, low-k materials have lower thermal
conductivity
• leads to significant metal-temperature
34. Parameters affecting interconnect
performance
• Width But,
of Interconnect
• Dielectric Pitch = w + s
• Diffusion Barrier
• Length
• Pitch
• Aspect Ratio
←
35. Parameters affecting interconnect
performance
• Diffusion Barrier • Reduction of copper cross-sectional area
• Dielectric – Cu atoms diffuse into silicon and
damage FETs
• Width – Must be surrounded by a diffusion
barrier
of Interconnect – So, ρ increases
• Length • Barrier thickness doesn’t scale rapidly as
the interconnect dimensions because of
• Pitch reliability constraints
• Aspect Ratio
36. Parameters affecting interconnect
performance
• Dielectric • Aspect ratio: AR = h/w
• Aspect Ratio – Old processes had AR << 1
– Modern processes have
• Width AR 2
• h>>w
of Interconnect • C = ε.(wl)/h
• Diffusion Barrier • R = ρl/tw
• Pitch
• Length
Global interconnect lengths remain the same.
37. Challenges of VLSI interconnects in
deep sub-micron technologies
• Surface Scattering
• Grain Boundary effect
MFP of e- in Cu = 40 - 100 nm
• Barrier Width These effects lead to the increase in ρo- increases
from 1.9μΩ-cm to 4.9μΩ-cm @ 45nm [25]
→
38. [17]
• The 50% delay of Y(t) is essentially the median point of the impulse response.
• If H(t) is symmetric, the 50% delay is accurately –m(1) (the first moment of the
impulse response). This is Elmore delay.
←
39. APPENDIX: Kinetic Inductance
• The total energy associated with electric current is
In normal wires, the energy stored in the magnetic field is significantly larger than
the kinetic energy of electrons, and the second integral is negligible.
• Kinetic inductance per unit length is given by
• Reactive impedance is always going to be negligible compared to its resistance.
Resistance per unit length is
where τ is the average collision time for carriers.
• Kinetic inductance for a carbon nanotube becomes important as the mean free
path of electrons can be very large.
←
40. APPENDIX: Quantum Capacitance
• To add electric charge to a quantum wire, one must add electrons to available
states above the Fermi level (Pauli Exclusion Principle). The required energy to add
electric charge Q to a quantum wire is
where the first term is the energy stored in the electric field (CE is the electrostatic
capacitance) and e is electron charge.
• By equating this energy to an effective capacitance, the expression for the
quantum capacitance (per unit length) is obtained as shown
and has a value in the order of 100 aF/m, in the same order of the electrostatic
capacitance of a typical wire above a ground plane.
• As a CNT has four conducting channels as described in the previous sub-
section, the effective quantum capacitance resulting from four parallel
capacitances cQ is given by 4cQ.
←
41. APPENDIX
• As rise times decrease the bandwidth of the signal increase, as more no. of higher
frequency components (harmonics) need to be accompanied into the signal to
achieve this rise-time.
• approximation: BW = 0.35/RT
• So, the effective frequency of a 100 MHz wave is 1GHz if only 10 harmonics are
taken.
Notes de l'éditeur
Device delay improves, but Interconnect delay increases. Globallengths are not shrunk, so they getslower.
As interconnect length approaches MFP, or lowers down further, these effects become prominent.
Ballistic conduction is the unimpeded flow of charge or energy carrying particles over relatively long distances in a material.
CNT bundle local interconnects have larger delay than Cu due to their larger resistance.
Input=u(t). Output=y(t). MNA=KCL for R and C. KVL for L or voltage source.
if time of flight dominates, then the line behavior is inductive. Otherwise, it is a resistive line.
UDE can only be used to determine the trend for interconnect delay
delay of segments connected by the repeaters = delay of a repeater. In this case, the slow switching due to resistance is counterbalanced by the fast switching of the gates.
How many repeaters should we use? How large should each one be?. RC delay v/s LC delay.
A dielectric is an electrical insulator that may be polarized by an applied electric field. Dielectric Constant relates how much the capacitance is increased due to presence of the material.
Pitch remains constant : as w increases, s decreases.
Barriers have much higher resistivity. Materials used are Transition metal - silicon compounds: Adding Si to Ta, Ti, Mo, W yields amorphous refractory barriers with stability to 700°C. Diffusion - molecules spread from areas of high concnto areas of low concn
As interconnect length approaches MFP, or lowers down further, these effects become prominent.A crystal is a material whose atomsare arranged in an orderly repeating pattern. Polycrystalline are composed of many crystallites/ grains.Scatter- deviate from a straight trajectory
Higher order moments give more sophisticated measurement for the distribution of the impulse response.