[IEEE 2012 IEEE/MTT-S International Microwave Symposium - MTT 2012 - Montreal, QC, Canada...
Transcript of [IEEE 2012 IEEE/MTT-S International Microwave Symposium - MTT 2012 - Montreal, QC, Canada...
Fast Extraction of High-Frequency Parallel Admittance of Through-
Silicon-Vias and their Capacitive Coupling-Noise to Active Regions
Chuan Xu1, Roberto Suaya
2, and Kaustav Banerjee
1
1 Dept. of Electrical and Computer Engineering, University of California, Santa Barbara, CA 93106, USA
2
D2S Division, Mentor Graphics Corporation, 38334 St. Ismier, France
Abstract — We introduce an accurate and efficient method to
extract high-frequency parallel admittance (capacitance and
conductance) among multiple Through-Silicon-Vias (TSVs). Our
method utilize the analytical expression of TSV’s MOS
capacitance and the 3D quasi-electrostatic (QES) scalar potential Green’s function in layered media for extracting the inverse of complex capacitance matrix of TSV segments, with consideration
of low conductivity silicon bulk region and high conductivity well regions with VDD/VSS contact rails. The elements of the complex
capacitance matrix of TSV segments can be summed up to the
final results of parallel admittance matrix of TSVs. Our method is
verified against a full-wave Finite-Element-Method (FEM) electromagnetic solver HFSS, and shows more than 103X speed-up
with good accuracy in the frequency range from digital circuit clock frequency to 100 GHz (<7% error for the self-admittance, the dominant quantity, and <13% error for the mutual-admittance, the smaller quantity). The complex capacitance matrix of TSV
segments and the scalar potential Green’s function is also used to
extract the noise coupling coefficient from TSV to active regions, which is also verified by HFSS simulation.
Index Terms — 3-D ICs, extraction, interconnect, parallel admittance, quasi-electrostatic, scalar potential Green’s function, through-silicon-via.
I. INTRODUCTION
High aspect ratio through-silicon-vias (TSVs) constitute a
key component for 3-D ICs [1],[2]. Both fabrication
technologies of TSVs [3]-[6] and electrical/electromagnetic
modeling and simulation [7]-[13] have been investigated. Full-
wave Finite-Element-Method (FEM) simulations [7] are
accurate but are very slow and memory intensive. Curve fitting
based empirical TSV admittance/impedance models [8] are in
closed form, but lack generality when geometrical and/or
material configurations change. The TSV’s MOS capacitance is
derived in [9]-[11], which is shown to be essential to the TSV
admittance. Analytical expressions for self-admittance of a
single TSV and coupling-noise from a single TSV to active
regions are derived in [12]. Numerical method in computational
electromagnetics to analyze multiple TSVs has been discussed
in [13], and the computation time is ~30 sec per data point for
the case of 4 TSVs. A fully analytical model for the high-
frequency series impedance (resistance and inductance) of
TSVs with consideration of substrate eddy current effect and
coupling with horizontal interconnects is discussed in [14]. The
TSV parallel admittance (capacitance and conductance, CG),
which is shown to be more important [10], needs investigation.
In this paper, we develop an accurate and efficient method to
extract the high-frequency (up to 100 GHz) parallel admittance
of multiple TSVs, which is based on analytical expression of
TSV’s MOS capacitance and scalar potential Green’s function
in layered media. Our extraction method is verified against a
full-wave FEM software (HFSS [15]). A method to extract the
coupling-noise from TSV to active regions when multiple TSVs
are present is also discussed.
II. QUASI-ELECTRO-STATIC (QES) SCALAR POTENTIAL
GREEN’S FUNCTION IN 3-D IC CONFIGURATION
Twin-well configuration is the most typical structure in bulk
CMOS and thin-BOX SOI technologies. In this paper, we
mainly discuss face-to-back 3-D IC with twin-well low
conductivity substrate (Fig. 1). Our methodology can be
extended to other type of configurations, once scalar potential
Green’s function in layered media is properly derived.
Quasi-electro-static (QES) assumption is valid since the
electromagnetic quarter wave length in SiO2 (0.37 mm at 100
GHz) is much greater than the TSV dimension. In such
assumption, the scalar potential in layered media at any (x,y,z)
due to a point source charge at (x',y',z'), can be expressed as:
( , , ) ( , , ; ', ', '); ( ', ', ')V x y z G x y z x y z qN x y zΦ= , (1)
where “<;>” is an operator, which captures the space integration
of sources at various (x',y',z'); qN(x',y',z') is the charge density at
Hbulk
(x=x2,y=y2,z=z2)+(x'=x1,y'=y1,z'=z1)bulk Si
Treat high conductive wells as ground plane
upper dielectric
lower dielectric
x
z
2 ( ') ( ') ( ')( , , ; ', ', ')
Si Si
x x y y z zG x y z x y z
j
δ δ δ
ε σ ωΦ − ⋅ − ⋅ −
∇ = −−
2 ( , , ; ', ', ') ( ') ( ') ( ') dielG x y z x y z x x y y z zδ δ δ εΦ∇ = − − ⋅ − ⋅ −
Fig. 2 The high conductivity wells (connected to VDD/VSS interconnects) can be treated as ground plane. The scalar potential Green’s function GΦ(x,y,z;x',y',z') satisfies Poisson’s equation
Hbulk
Si depletion region
Si bulk region
(P-type, 10 Ω-cm)
Si d
eple
tio
n r
egio
n
TS
V m
eta
l
STI STISTI STI STIN-Well, 0.1 Ω-cm
Si d
eple
tio
n r
egio
n
STI STIP-Well, 0.1 Ω-cm P-Well
TSV isolation SiO2
VDDVSS NMOS PMOS NMOSPMOS
Both p-well and n-well
are present in twin-well
configuration. Substrate
conductivity (~10 S/m)
is typically much lower
than the wells in most
advanced (< 32 nm)
CMOS technologies,
due to less concern of
latch-up effect in lower
supply voltage and more
concern of the substrate
eddy current for RF
components.
Fig. 1 Cross sectional view of TSV in twin-well face-to-back 3-D IC configuration with low conductivity substrate
978-1-4673-1088-8/12/$31.00 ©2012 IEEE
(x',y',z'); GΦ
(x,y,z;x',y',z') is the space domain scalar potential
Green’s function in layered media. The high conductivity wells,
which are connected to VDD/VSS interconnects, can be treated as
a ground plane (Fig. 2). GΦ
can be transformed from the
spectral domain vector potential Green’s function (GΦ ):
00
( , , ') (1 2 ) ( , , ') J ( ) dG z z G z zρ π λ λ ρ λ λ∞
Φ Φ∆ = ⋅∆∫ (2)
where 2 2( ' ) ( ' )x x y yρ∆ = − + − ; λ is the horizontal component
of wave vector; J0 is the 0th
order Bessel function of the 1st
type.
For 0<z<Hbulk and 0<z'<Hbulk (Hbulk is the thickness of Si bulk
region in thinned substrate),
( ) | '| ( ')
, 1
( ') ( ') ( ')
2 3 4
( , , ') 1 2 e e
e e e
z z z z
Si eff
z z z z z z
G z z c
c c c
λ λ
λ λ λ
λ λεΦ − − − +
+ − − −
= +
+ + +
(3)
where εSi,eff = εSi − jσSi/ω is the effective complex permittivity in
the Si bulk region; εSi and σSi are the permittivity and
conductivity in Si bulk region, respectively; ω is the radial
frequency; coefficients c1, c2, c3 and c4 are independent of z and
z', and can be extracted from boundary conditions. Eq. (3) can
be approximated from making a Taylor series of 2
e bulkHλ− :
( )
, 2| '| ( ')
, ,
, 2 4( ') ( ') ( ')
,
1( , , ') e 1 e e
2
e e e e (e )
bulk
bulk bulk
Si eff diel Hz z z z
Si eff Si eff diel
Si eff diel H Hz z z z z z
Si eff diel
G z z
O
λλ λ
λ λλ λ λ
ε ελ
λε ε ε
ε ε
ε ε
−Φ − − − +
− −+ − − −
−= − − +
−+ − − +
+
(4)
where εdiel is the permittivity in lower dielectric. From (2)&(4),
2 2 2 2,
, 2 2
, ,
2 2 2 2
2 2
1 1 1( , , ')
4 ( ') ( ')
11 ( ' 2 )
4
1 ( ' 2 ) 1 ( ' 2 )
1 ( ' 2 )
Si eff
Si eff diel
bulk
Si eff Si eff diel
bulk bulk
bulk
G z zz z z z
z z H
z z H z z H
z z H
ρπε ρ ρ
ε ερ
πε ε ε
ρ ρ
ρ
Φ
∆ ≈ − ∆ + − ∆ + +
− + ∆ + + −+
− ∆ + − + − ∆ + − −
+ ∆ + + +
(5)
Eq. (5) can be interpreted as the contribution from the source
charge itself and its 5 image charges from reflections at the
ground plane and/or the interface between Si bulk region and
lower dielectric. GΦ
with z<0 or z>Hbulk or z'<0 or z'>Hbulk can
be obtained using similar way, but is not presented in this paper.
III. TSV PARALLEL ADMITTANCE EXTRACTION AND
CAPACITIVE COUPLING-NOISE FROM TSVS TO ACTIVE REGIONS
TSVs are segmented as shown in Fig. 3. The segmentation
rule: the TSVs are firstly cut at all interfaces (e.g., the interface
between Si bulk region and lower dielectric); for high aspect
ratio parts (e.g., parts within Si bulk region), they are cut into
more than one segment. For the test cases in Section IV, the
TSV parts within Si bulk region are cut into 5 segments, which
make 8 segments per TSV. Fig. 3 also shows the correlation
among scalar potential Green’s function, P and C matrices of
TSV segments, Y matrix of TSVs and charge on each TSV
segment Qi'. Note that the correlated equation from Green’s
function to Pi,i' is for the case that segments i and i' are within
the two boundaries of Si bulk region, but similar expression can
be found for either segments i or i' outside the two boundaries.
The noise coupling coefficient from a TSV to the active
regions when multiple TSVs are present can be obtained from
current density injected into the well (J(x,y)) and the well
resistance network. The general idea has been presented in [12],
where only one TSV is present. In this paper, J(x,y) is obtained
in a more accurate way, which is proportional to the gradient of
scalar potential at the boundary of Si bulk region and p-well:
( )' ' ' ' 0'
( , ) ( ) ( , , ; , , )Si Si i i i i z
i
J x y j Q G x y z x y z zσ ωε= +
= + ∂ ∂∑ (6)
J(x,y) can then be used to estimate the coupling-noise voltage in
the active regions, as is done in [12].
IV. VERIFICATION AGAINST FULL-WAVE FEM SIMULATIONS
Our extraction method (coded in Maple [16], a symbolic
computation and mathematical programming tool) is verified
against full-wave FEM simulations (HFSS [15]) for two cases.
A. Test Case with 2 TSVs
The calculation results of the self- and mutual-admittance of
the two TSVs (Y11 and Y12) in the test case (top view shown in
Fig. 4) are compared with HFSS simulation in Fig. 5. The
TSV segmentationTSV segmentation
C=P-1
segment segmentTSV ' TSV
, , '
'
TSV TSV
TSV TSV
i i i i
i i i i
i i
Y j Cω′∈ ∈
′ = ∑ ∑
segment TSV
' ',
TSV
TSV
TSV
i i
i i i i
i i
Q V C∈
= ⋅
∑ ∑
When 0<zi<Hbulk and 0<zi' <Hbulk
, '
'
1 1ln
2
1ln
2
( , , )
( ', same TSV and same segment)
( , , )
( ', same TSV but different segm
via ox dep
via ox dep
i i
via ox
i ox via
via ox dep
Si via ox
i i r t w
i i r t w
P
r t
h r
r t w
r t
G z z
i i
G z z
i i
ρ
ρ
πε
πε
ρ
ρ
∆ = + +
∆ = + +
=
+
+ + +
+
+ ∆
=
∆
≠
' ' '
ents)
( , , ; , , )
( ', different TSVs)
i i i i i iG x y z x y z
i i
≠
Fig. 3 Schematic view of TSV segmentation (the TSVs are cut by the dashed
lines) and correlation among scalar potential Green’s function, P and C
matrices of TSV segments, Y matrix of TSVs, and charge on each TSV
segment Qi'. C is the complex capacitance matrix of TSV segments (due to the
conductive silicon substrate, εSi,eff is complex); P=C-1; Y is the parallel admittance matrix of TSVs; ViTSV is the voltage on TSV iTSV. tox and εox are the
thickness and permittivity of TSV oxide, respectively; rvia is the radius of TSV
metal; wdep is the depletion width surrounding the TSV oxide.
VDD
Pla
cin
g
NM
OS
Pla
cin
g
PM
OS
Pla
cin
g P
MO
STSV
Pla
cin
g N
MO
S
VSS
Pla
cin
g
NM
OS
Zone 1 Zone 2 Zone 3 Zone 4 Zone 5
VSS
VSS
1.5 µm 1.5 µm 1.5 µm
TSV
y
xd1 d2
OP1
d1 d2
OP2y1
y1
VDD
Pla
cin
g
NM
OS
Pla
cin
g
PM
OS
Pla
cin
g P
MO
STSV
Pla
cin
g N
MO
S
VSS
Pla
cin
g
NM
OS
Zone 1 Zone 2 Zone 3 Zone 4 Zone 5
VSS
VSS
1.5 µm 1.5 µm 1.5 µm
TSV
y
xd1 d2
OP1
d1 d2
OP2y1
y1
Fig. 4 Schematic top view of 2 TSVs in bulk CMOS technology. The 2 TSVs have center-to-center (c2c) distance of 2y1; observation points OP1 & OP2 and geometrical parameters d1 (=1.0µm) & d2 (=0.5µm) are defined in the plot.
978-1-4673-1088-8/12/$31.00 ©2012 IEEE
greater error of both Y11 and Y12 at higher frequencies is due to
the fact that the ground plane treatment of wells is more valid at
lower frequencies for the condition |σwell/(σSi+jωεSi)|>>1. Our
method gives <7% error for Y11, the dominant quantity, and
<9% error from 4 GHz (typical digital circuit clock frequency)
to 100 GHz for Y12, the smaller quantity. We compare the
results of noise coupling coefficient g31 (ratio of voltage
variation at observation point to that at one of the 2 TSVs) with
HFSS in Fig. 6. The total computation time per data point using
our method is 0.10 sec as compared to 130 sec using HFSS.
B. Test Case with 4 TSVs
The top view of the test case and the results are shown in Fig.
7. Due to the symmetry of the structure, our extraction method
gives Y11=Y22=Y33=Y44, Y12=Y23=Y34=Y14 and Y13=Y24. In HFSS
simulation, there is slight and negligible difference within the
three categories (Y11, Y12 and Y13), due to the slight asymmetry
from the VSS lines. Nevertheless, it is sufficient to just compare
the results of Y11, Y12 and Y13. Similar to the case of 2 TSVs, in
the case of 4 TSVs, our method is generally more accurate
(<6% error) for Y11, the dominant quantity, and less accurate
(<13% error from 4 GHz to 100 GHz) for Y12 and Y13, the
smaller quantities. The computation time per data point using
our method is 0.31 sec as compared to 680 sec using HFSS.
V. SUMMARY
An accurate and efficient method to extract the high-
frequency parallel admittance of multiple TSVs is developed
through utilizing the analytical expression of TSV’s MOS
capacitance and 3D QES scalar potential Green’s functions in
layered media, and is extended to the estimation of electrical
coupling-noise voltage from TSV to active regions. Our method
is verified against full-wave FEM electromagnetic simulations,
which shows good accuracy and 103X speed-up. Our method
gives <7% error for the TSV self-admittance, the dominant
quantity, and <13% error from digital circuit clock frequency to
100 GHz for the TSV mutual-admittance, the smaller quantity.
REFERENCES
[1] K. Banerjee et al, “3-D ICs: A novel chip design for improving deep-submicrometer interconnect performance and systems-on-chip integration,”
Proceedings of the IEEE, vol. 89, pp. 602-633, 2001. [2] G. L. Loi et al, “A thermally-aware performance analysis of vertically
integrated (3-D) processor-memory hierarchy,” 43th ACM/IEEE DAC, 2006,
pp. 991-996. [3] N. Sillon et al, “Enabling technologies for 3D integration: From packaging
miniaturization to advanced stacked ICs,” Tech. Dig. IEDM 2008, pp. 595-598. [4] F. Liu et al, “A 300-mm Wafer-level three-dimensional Integration Scheme
Using Tungsten Through-Silicon Via and Hybrid Cu-Adhesive Bonding,”
Tech. Dig. IEDM 2008, pp. 599-602. [5] G. Katti et al, “3D stacked ICs using Cu TSVs and die to wafer hybrid
collective bonding,” Tech. Dig. IEDM 2009, pp. 357-360. [6] L. Cadix et al, “Integration and frequency dependent electrical modeling of
Through Silicon Vias (TSV) for high density 3DICs,” 2010 IITC, pp. 1-3. [7] J. S. Pak et al, “Electrical characterization of through silicon via (TSV)
depending on structural and material parameters based on 3D full wave
simulation,” Int. Conf. Electronic Materials and Packaging, 2007, pp. 1-6. [8] I. Savidis et al, “Closed-form expressions of 3-D via resistance, inductance, and capacitance,” IEEE TED, 56(9), pp. 1873-1881, 2009. [9] G. Katti et al, “Electrical modeling and characterization of through silicon
via for three-dimensional ICs,” IEEE TED, 57(1), pp. 256-262, 2010. [10] C. Xu et al, “Compact AC Modeling and Performance Analysis of
Through-Silicon Vias in 3-D ICs,” IEEE TED, 57(12), pp. 3405-3417, 2010. [11] T. Bandyopadhyay et al, “Rigorous Electrical Modeling of Through
Silicon Vias (TSVs) With MOS Capacitance Effects,” IEEE Trans. CPMT, 1(6), pp. 893-903, 2011. [12] C. Xu et al, “Compact Modeling and Analysis of Through-Si-Via-Induced
Electrical Noise Coupling in Three-Dimensional ICs,” IEEE TED, 58(11), pp. 4024-4034, 2011. [13] K. J. Han et al, “Electromagnetic Modeling of Through-Silicon Via (TSV)
Interconnections Using Cylindrical Modal Basis Functions,” IEEE Trans. Advanced Packaging, 33(4), pp. 804-817, 2010. [14] C. Xu et al, “A Fully Analytical Model for the Series Impedance of
Through-Silicon Vias With Consideration of Substrate Effects and Coupling
With Horizontal Interconnects,” IEEE TED, 58(10), pp. 3529-3540, 2011. [15] User’s Guide: High Frequency Structure Simulator (HFSS v10.0), Ansoft Corporation, Pittsburgh, PA, 2005. [16] Maple 14 Student Edition. [Online]. Available: http://www.maplesoft.com
109
1010
1011
-0.5
0.0
0.5
1.0
1.5
2.0
2.5
3.0
VSS
TSV1
VSSVSSVSS
6 µm
TSV3
TSV2 TSV4
6 µm VSS
TSV1
VSSVSSVSS
6 µm
TSV3
TSV2 TSV4
6 µm
Lines: This workSymbols: HFSS simulation
Im(Y11
)
Re(Y11
)
Im(Y12
)
Re(Y12
)
Im(Y13
)
Re(Y13
)Ad
mitta
nce (
10
-3 S
)
Im(Y11
)
Re(Y11
)
Im(Y12
)
Re(Y12
)
Im(Y13
)
Re(Y13
)
Frequency (Hz) Fig. 7 The self and mutual admittance of 4 TSVs (Y11, Y12 and Y13) as a function of frequency from our methodology in this work and from HFSS simulation. The schematic top view is shown in the inset. The TSVs are located in between VSS lines and form a square shape with diagonal distance of 6 µm. All other parameters are the same as those in the test case of 2 TSVs.
Symbols: HFSS simulation Lines: our method
109
1010
1011
-0.50.00.51.01.52.02.5 2y
1 = 4
µm
Im(Y11
)
Re(Y11
)
Im(Y12
)
Re(Y12
)
Ad
mitta
nce (
10
-3 S
)
Im(Y11
)
Re(Y11
)
Im(Y12
)
Re(Y12
)
Frequency (Hz)
4 6 8 10 12-0.5
0.0
0.5
1.0
1.5
f = 50 GHzA
dm
itta
nce
(10
-3 S
)
2y1 (µm)
(a) (b)
Fig. 5 The self and mutual admittance of two TSVs (Y11 and Y12) as a function of frequency and c2c distance (2y1) from our methodology in this work and from HFSS simulation. The cross section view and top view of the configuration are shown in Fig. 1 and Fig. 4, respectively. The technology configuration, default material and geometrical parameters are the same as those in Section III.A of [12]. (a) Y11 and Y12 as a function of frequency (f) with 2y1 = 4 µm; (b) Y11 and Y12 as a function of 2y1 with f = 50 GHz.
Symbols: HFSS simulation Lines: our method
109
1010
1011-80
-70
-60
-50
-40
-30
-20
OP2
OP1
dB
(|g
31|)
Frequency (Hz)
2y1 = 4
µm
4 6 8 10 12
-50
-40
-30 OP1
50 GHz
10 GHz
dB
(|g
31|)
2y1 (µm)
(a) (b)
Fig. 6 The noise coupling coefficient g31 from one of the 2 TSVs to the OPs, with the other TSV keeping a constant voltage. (a) dB(|g31|) as a function of frequency (f) with 2y1 = 4 µm for both OP1 and OP2; (b) dB(|g31|) as a function of 2y1 for OP1 with f = 10 GHz and f = 50 GHz.
978-1-4673-1088-8/12/$31.00 ©2012 IEEE