A New Boundary Element Method for Accurate Modeling of Lossy

*
Substrates with Arbitrary Doping Profiles
Xiren Wang, Wenjian Yu, and Zeyi Wang
Eda Lab, Dept. Computer Science & Technology,
Tsinghua Univ., Beijing, 100084, China

Abstract - It is important to model substrate couplings for and converges slowly. In [4], a ‘numerically stable’ method
SoC/mixed-signal circuit designs. After introducing the was proposed to calculate the Green’s function with the
continuation equation of full current in lossy substrates, we acceleration of discrete cosine transform (DCT) [7].
present a new direct boundary element method (DBEM), which However, it is not actually stable, with further remedy
can handle the substrates with arbitrary doping profiles. Three
presented in [8]. In [2], there is an excellent idea of eigende-
techniques can speed up the DBEM remarkably, which include
reusing coefficient matrices for multiple-frequency calculation, composition, which results in a speedup of a dozen over the
condensing the linear system, and sparsifying coefficient matrix. above DCT-Green’s function method. The coefficient matrix
Numerical experiments illustrate that DBEM has high is dense for these methods, however. Storing and solving the
accuracy and high efficiency, and is versatile for arbitrary linear system are memory- and time-consuming.
doping profiles. On the other hand, the resistive model becomes invalid at
high frequencies, and a comprehensive frequency-dependent
impedance model is desired. The method in [4] is able to
I Introduction give such impedances. However, it also needs to find an
expensive Green’s function, whose derivation is based on
There are currently increasing demands for high-inte- the layered structure of substrates.
gration circuits [1]. High-speed digital components and Note that substrates are not always stratified. In fact, there
highly-sensitive analog components are often built on a are lots of realistic substrates with layout-dependent doping
common substrate. Although the high integration has some profiles, such as oxide wells, trenches, sinkers, buried
advantages, such as low power dissipation [2], there is a pro- diffusions and etc [9]. There are also special structures like
blem that the current noises injected by digital components Faraday shields and junction shields [10] for noise reduction.
travel the shared substrate and impact sensitive analog To simulate such substrates, the methods based on Green’s
components severely [3]. Besides, substrate losses also function meet much difficulty.
impact circuit performances considerably. For instance, the In this paper, we bring forward a direct boundary element
quality factor of such passive devices as inductors is impor- method (DBEM) [3, 11, 12] for lossy substrate modeling.
tant for circuit performances, especially in wireless commu- Conventionally it is applied in electrostatic DC capacitance
nication applications. But it is limited by substrate losses, extraction [12] or resistance extraction [11]. We will modify
especially at high frequencies. Thus acknowledge of lossy it to model lossy substrates at any frequency with the help of
substrates is very important even critical for designs [4]. the concept of complex permittivity [13].
A resistance network model of a substrate is often effi- In DBEM, only substrate boundary is discretized, then the
cient at frequency of up to several gigahertz. Some nume- variables are fewer than in FDM/FEM. Only the free-space
rical methods are presented for the resistive simulation, Green’s function is used, which brings two advantages. One
including finite element method (FEM), finite difference is that no derivation of the function is needed, and its
method (FDM) [5], and methods based on Green’s function computation is also straightforward. The other is that the
(sometimes also called boundary element method) [2, 4, 6, function is independent of structures, so DBEM can handle
7]. FEM and FDM are too slow, since they discretize entire substrates with arbitrary doping profiles. Besides, three
volume of a substrate. However, they have the advantage of kinds of accelerating techniques are presented, which are re-
versatility, i.e., they can handle substrates with arbitrary using coefficient matrices for multiple-frequency calculation,
doping profiles rather than only layered substrates. condensing the linear system before solving, and sparsifying
The Green’s function based methods [2, 6, 7] are gene- the coefficient matrix. They can enhance the efficiency of
rally faster than FDM and FEM, because they only discretize DBEM considerably, but preserve the accuracy exactly.
contact surfaces. A suitable Green’s function that satisfies The rest is organized as follows. In Section II, DBEM
boundary conditions needs to be found. For multilayered principle for impedance modeling is presented. In Section III,
substrates, the function consists of multiple infinite series, the accelerations for DBEM are described separately. Nume-
rical experiments follows, so as to demonstrate the accuracy
and efficiency of the method. A special case is a non-stra-
* This work is supported by the China National Science
Foundation under Grant 60401010. It is also partly supported tified substrate with lateral resistivity variation, and it can
by National High Technology Research and Development illustrate versatility of the method. The conclusions are
Program of China (No. 2004AA1Z1050). given at last.
 II. Frequency-Dependent Extraction of Substrate
Coupling with DBEM

Fig. 1 shows an example of lossy substrate, which is cons- M top ,σ top , ε top
tituted by layers of mediums Mi (with finite resistivity) and C2 Cj
contacts Cj. There is usually a grounded plane on the bottom.
In many cases, contacts are assumed to be on the top. Here, M i ,σ i , ε i
however, they can be placed anywhere if necessary.
For lossy substrates (such as Silicon), the coupling
between contacts can be modeled as frequency-dependent M 1 ,σ 1 , ε1
Back plane
impedances. To get the impedances needs to pre-set voltages
of contacts, and then to calculate the full current flowing Fig. 1 An example substrate with a back plane. The non-zero
through contacts. The reciprocal of the current is related to conductivity of medium Mi is σ i , and the permittivity is ε i .
the desired impedance.
Assume electric potential of contacts to be in u = ue jωt If En is known, we can get the current between contacts m
form, where u is the maximum voltage, ω is the angular and k. If the pre-set voltage of contacts m is 1V, and contact
frequency, and t is time. Electric field within the substrate is k is of 0V, the impedance between m and k will be
also in E = Ee jω t +θ form. Start from the Maxwell equation
1
Z mk = . (6)
∂D ∫ (σ + jωε ) En d Γ
∇× H = J + , (1) Γk
∂t
where Γk is the surface of contact k. σ and ε are the condu-
where H is the magnetic filed intensity, J is the conductance
ctivity (reciprocal of resistivity) and permittivity of the me-
(ohmic) current density, D is the electric displacement. For
dium surrounding contact k, respectively.
mediums with constant conductivity σ and permittivity ε ,
Note that only En is needed in (6). This can be made use
of later.
(σ + jωε )∇ ⋅ E = 0 . (2) Now we start solving En from (3). Utilizing the Green's
identities, and selecting free-space Green's function as
Because E = −∇u 1, the following Laplace equation holds: weighting function, we transform (3) into a boundary
integral equation (BIE), which is defined on the boundary of
∇ 2u = 0 , in medium Mi (3) medium Mi. Partition the boundary into Ni elements Γij. Then
we get a discretized BIE [11]:
with the mixed boundary conditions of
Ni Ni

u = u , on contact surfaces (Dirichlet boundary) (4a) csus + ∑ ∫ q*( s ) ud Γ = ∑ ∫ u*( s ) En d Γ, for medium Mi (7)
Γij Γi j
j =1 j =1
En = 0 , on natural boundary (Neumann boundary) (4b)
where collocation point s is on the boundary of Mi. cs is a
where u is pre-set voltage of contacts, usually 1V or 0V. En *
is normal electric field intensity. Besides, the potential and constant. u( s ) is the fundamental solution of Laplace equation
full current is continuous on the interface of adjacent *
which is related with point s, and q( s ) is its normal derivative.
mediums a and b:
Here, u* is equal to the free-space Green’s function.
List the equations (7) for each of the mediums in the
ua = ub , (5a) substrate. Combine these equations with the interface
(σ a + jωε a ) En ,a = (σ b + jωε b ) En,b . (5b) conditions (5). Substitute the conditions (4), and we get an
overall system
where σ a and σ b ( ε a and ε b ) are the conductivity (permit-
tivity) of regions a and b, respectively. Ax = b, (8)

where A is composed of complex entries. Vector x is

constituted by u and En unknowns of all mediums. b is
created after moving all the known values related to pre-set
1
Generally speaking, E = −∇u − ∂A , where A is the magnetic voltage u to the right side. Solving the system, we can
∂t directly get En, and in turn the impedance in (6). When the
vector potential. Here we adopt the Coulomb Gauge [15], which impedances between many pairs of contacts are desired, the
defines ∇ ⋅ A = 0 . Thus E = −∇u . system (8) will become
AX = B, (9) the diagonal as possible [3, 12]. These can shorten solution
time to some extent. But solving a large linear system takes
where B is also a matrix. still too much time.
However, if we discard inessential variables at first, and
the consequent solving the condensed system will be much
easier. Remind that only En variables are needed in (6), so
III. Effective Techniques for DBEM we can discard some u variables. Rewrite a system Ax = b in
terms of sub-matrices (for (9), we can do it similarly):
Although DBEM has much fewer variables in some sense,
it seems still too slow for substrates with many contacts.
⎡ A00 A0T ⎤ ⎡ x ⎤ ⎡ b0 ⎤
Accelerating techniques are preferred.
⎢A ⋅ =
ATT ⎥⎦ ⎢⎣ uT ⎥⎦ ⎢⎣bT ⎥⎦
, (10)
⎣ T0
A. Reuse Matrix where x and uT are two subsets of x. Remove the inessential
uT equivalently:
To evaluate a lossy substrate, we usually need to calculate
coupling impedances at many frequencies. Usually we need
to build the system (9) for many times. However, most A x = b, (11)
entries of matrix A and vector B can be reused so as to save
the computational time. where
For an interface element, there is a pair of variables En,a
and En,b related to it. In medium a, the coefficient of En,a, −1
A = A00 - A0T ATT AT 0 , (12)
denoted by Ia, can be determined by the integral in (7). In
medium b, Ib, the coefficient of En,b can be similarly and
obtained. En,a and En,b have the relationship of
σ + jωε a
En ,b = a En ,a . Actually, to overcome the singularity −1
b = b0 - A0T ATT bT . (13)
σ + jωε
of the matrix, we keep one of them in the system, say En,a.
The condensing procedure (12) (13) is generally expensive
Then the En,a coefficient in medium b becomes
in computation, because it involves the inversion of ATT and
σ + jωε a , which depends on ω .
Ib × a matrix-matrix multiplications.
σ b + jωε b We start to reduce the computational complexity of (12)
In a word, only those matrix entries related to the (13). For an example three-layer substrate, reordering the
interface variables like En,a are frequency-dependent. The non-interface variable v33 of medium 3:
other entries remain the same for any frequency. Besides, B
in (9) is also independent of frequency, because it is merely vO → EnC → uT , (14)
related to the pre-set voltage u and the integral similar to Ia.
When impedances at multiple frequencies are needed, we
where EnC represents the unknown En on contact surfaces, uT
need to build the system (9) for each frequency as usual.
denotes u unknowns on the top surface of medium 3, and vO
However, a better choice is to compute the aforementioned
denotes the other unknowns in v33.
frequency-independent entries only once, and reuse them for
With the ordering (14) and Fig.2 applied, the DBEM
any frequency. The other entries are re-calculated at each
matrix A for a substrate is shown in Fig.3a, where zero
frequency. In this way some CPU time can be saved. For the
first test case in Section IV, this reusing technique reduces entries are in white. The condensed matrix A is in Fig.3b.
the total running time for 20 frequencies from 144.91 to ATT is diagonal; A0T and bT are sparse. These properties can
124.70 seconds, or by 14%. be analytically proven. Refer to [3] for more details. Thus
calculating A and b through (12) (13) is easier.

B. Condense the Linear Equation System → v 11 → u 1 2 → u 13 → … → u 1 M

→ q 2 1 → v 22 → u 2 3 → … → u 2 M
Solving the complex system (9) consumes a great deal of
CPU time. Direct solutions have too high computational → q 3 1 → q 32 → v 33 → … → u 3 M
complexity. As we will see in subsection C, the coefficient
→ … … … … … … … … …
matrix is sparse for multi-medium problems, so we choose a
GMRES solver [14]. Moreover, reordering unknowns and → q M 1→ q M2→ qM 3→ … → vMM
corresponding collocation points in the same sequence can
make the matrix diagonally dominant [3]. Reorder unk- Fig. 2 Reorder unknowns. M is the number of mediums. vii
includes u on Dirichlet boundary and q on Neumann boundary of
nowns as in Fig.2 can arrange non-zero blocks as close to
medium i. uij and qji (i < j) are on the interface of mediums i and j.
 A00 A0T

AT0 ATT
(a) (b)
(a) (b) Fig. 4. (a) Original matrix, with 248,700 nonzero entries. (b) Matrix
Fig. 3. Distribution of nonzero entries in A: (a) original; (b) after after the QMM idea is used, with 142,716 nonzero entries.
using the condensing technique. Zero entries are in white.

In the first experiment in Section IV, it takes 4.1 seconds to

solve (8). If the condensing technique is applied, the solution IV. Numerical results
takes merely 1.5 seconds, or 60% is saved. Unfortunately,
the condensing procedure (12) (13) itself takes 5.0 seconds. The first experiment is a substrate involving a top contact
But when the sparsifying technique in the next subsection is and a back plane, as shown in Fig.5a. And the second case is
applied, the procedure needs only 0.58 seconds. shown in Fig.5b, involving a back plane and two top
On the other hand, there is only one right-hand side here. contacts separated by s µm. The test cases will be configured
If there are many sides as (9) indicates, the condensing with two types of substrate processes [4], as shown in Fig.6.
technique will bring higher efficiency [3, 11], since solution In order to be compatible with literature data, the contacts
time for each side is considerably reduced. are assumed to be of zero heights.
For the first test case, the impedances obtained with the
presented DBEM are depicted in Fig.7 (for the LR process)
C. Sparsify Matrix and Fig.8 (for the HR process), respectively. The data in [4]
are also depicted. The figures indicate that with any process,
From Fig.3, we learn that coefficient matrix A is blocked the magnitude of impedance decreases as the frequency
sparse. The sparsity can be analytically deduced from equa- increases. This is because the substrate current injection is
tion (7), where collocation points and variables related to capacitive in nature [4]. With the low-resistivity (LR)
each other are within the same medium. In other words, process, the impedance is smaller in magnitude than with the
collocation points and variables in different mediums have high-resistivity (HR) process. This is because the bulk is
no direct numerical relationship. Therefore, A is blocked with much smaller resistivity in the LR process.
sparse for multi-medium problems. Both the magnitude and the phase of the impedances
If a physical medium is quasi-cut into fictitious compo- obtained with DBEM are very close to literature data. The
nents, the matrix sparsity can be enlarged. This is the basic discrepancy is within 1.0%, which also includes the possible
idea of a quasi-multiple medium (QMM) technique in [11, error of experiment setups (we did not find the exact setups
12]. For an example substrate, if we cut the third medium in [4]). In a word, DBEM has a high accuracy.
into 6 smaller components, the original matrix Fig.4a turns
into a sparser one Fig.4b. The new matrix is a bit larger,
because the quasi-cutting will bring additional variables on
interface of adjacent fictitious components. What’s most
important is the total number of non-zero entries is
256 256
remarkably reduced. Since most time is spent in
matrix-vector multiplication in a GMRES solver, the
enhanced sparsity is certain to speed up the solution.
Similarly, the technique can also speed up the condensing 100
256 100 50
procedure (12) (13). s
For instance, in the first experiment in Section IV, 10
10 10
solution time of (9) is reduced from 4.1 seconds to 2.0
seconds, and the speed up is 2. (a) (b)
As we have seen, each of the three techniques above can Fig. 5. (a) A single contact at center. (b) Two contacts located
enhance the efficiency of DBEM without any accuracy loss. symmetrically along the center with distance of s µm. There is a
back plane in the two substrates.
What’s more, if they are combined together, the enhan-
cement will be more obvious, as shown in the next section.
 Contact Contact 1.E+05 50
eox=3.9 tox=1µm eox=3.9 tox=1µm *Zmag in [4]
55
eox=3.9 tox=1µm eox=3.9 tox=1µm Zmag
ρ epi =100k Ω-µm, tepi=5µm ρ epi =500 Ω-µm, tepi=1µm *Zph in [4]
60

1.E+04 65

Phase
Zph

Mag
70
ρ epi =100Ω-µm, tsub=300µm ρ epi =100kΩ-µm,tsub=300µm 75

1.E+03 80

85

90
(a) b)
1.E+02 95
Fig. 6. Side views of two doping profiles [4]. (a) Low-resistivity 0.1 2.1 4.1 6.1 8.1
(LR) process; b) high-resistivity (HR) process. Bottom bold lines Frequency (GHz)
denote back planes. Fig. 8. Magnitude and phase of the contact-ground impedance of
the single-contact substrate with the HR process.
Take a look at the efficiency. Since running parameters of
the method in [4] can not be obtained, we do not compare much difficulty in handling substrates with non-stratified
DBEM with it directly. But the superiority of DBEM over doping profiles. In reality, there are many such substrates.
Green’s function methods for substrate impedance modeling For example, there are many lateral variation components in
can be roughly deduced according to [3] and [11], where the substrates, such as oxide trenches and wells [9]. In order
DBEM is up to hundreds of times faster than the excellent to illustrate the versatility of DBEM, examine the simple
eigendecompostion method in [2] for resistance extraction. example shown in Fig. 10, where the top central block in
The total running time to get the impedances at 20 frequ- gray has a distinct resistivity from its lateral neighbor. We
encies is listed in Table 1. All DBEM programs run on a Sun call the block “LVB” (lateral variation block).
Fire 880 workstation, with CPU frequency of 750 MHz. The
TABLE I.
table tells us that each of the techniques in section III can
Running time for substrate Fig.5a with LR process
bring enhancement of efficiency. Furthermore, their combi-
nation performs much better, with a speed-up of 2.7 over Pure DBEM 144.91
pure DBEM. DBEM + Reusing (Sect III.A) 124.70
For the two-contact substrate Fig.5b, the contact-contact DBEM + Condensing (Sect III.B) 74.06
impedance depends on the spatial separation. The impedance DBEM + Sparsifying (Sect III.C) 70.54
of this substrate with the LR process is depicted in Fig. 9. DBEM+Reusing+Condensing+Sparsifying 52.90
We learn that the impedance is very sensitive to the distance
s. As s extends from 2µm to 10µm, the impedance increases Since the conductivity of epi layer (2000 s/m) is large,
by three times in magnitude. When s turns from 10µm to and at frequency of up to 10 GHz, the ohmic current is
50µm, the impedance rises by ten times. The same tendency dominant over the displacement current in the layer. Thus
is found in [4]. changing the dimension of the higher-resistivity LVB block
Note that the both doping profiles in Fig. 6 are stratified. will have limited impact on the contact-contact impedance,
However, these methods based on Green’s functions have as shown in Fig. 11.

1.E+05 50 1.E+08
*Zmag in [4] s=50
55
Zmag
1.E+07 s=10
*Zph in [4] 60 s=2
Zph
1.E+04 65
1.E+06
Phase

70
Mag
Mag

75 1.E+05
1.E+03 80
85 1.E+04
90
1.E+02 95 1.E+03
0.1 2.1 4.1 6.1 8.1
0.1 2.1 4.1 6.1 8.1 9.6
Frequency (GHz)
Frequency (GHz) Fig. 9. Magnitude of the contact-contact impedance of the
Fig. 7. Magnitude and phase of the contact-ground impedance of two-contact substrate Fig. 5b, with the LR process. s is the
the single-contact substrate with the LR process. distance between the contacts.
 Contact 1 Contact 2 1.E+10
L=70
1.E+09
L=20

L 1.E+08 L=0

Mag
L
1.E+07

1.E+06

Fig. 10. Substrate with lateral resistivity variation in the epi layer of 1.E+05
0.1 2.1 4.1 8.1
the HR process. The width and length of the substrate are 200 µm. Frequency (GHz)
The resistivity of the epi layer and the gray block are 5e-4 and 0.5 Fig. 11. Magnitude of the contact-contact impedance. Since the epi
Ωm, repectively. Note that each contact here is modeled as a block layer of HR process is very small, the central block is not possible
with 3-D shape, as well as the LVB block. to be large, and thus the curves are close to each other.

Fig.11 illustrates that when the LVB block is small (L=0, or coupling in mixed-signal IC’s,” IEEE Trans. Comput. Aided
L=20), the impedance is relatively small. As the block grows Design, Vol. 18, pp. 597-607, 1999.
large (L=70), the impedance increases, since it obstructs the [3] X. Wang, W. Yu, and Z. Wang, “Substrate Resistance Extra-
ohmic current flow considerably. Besides, for a specified ction with Direct Boundary Element Method,” Asia and South
frequency, the difference between the impedances under Pacific Design Automation Conference, pp.208-211, 2005.
different L is small. A reason is that the obstructing LVB [4] A.M. Niknejad, R. Gharpurey, and R.G. Meyer,
block can not be very large due to spatial limitation. “Numerically stable Green function for modeling and analysis
The results tell us that DBEM is able to simulate such of substrate coupling in integrated circuits,” IEEE Trans.
complex substrates without any difficulty. Comput. Aided Design, Vol. 17, pp. 305-315, 1998.
[5] N.K. Verghese and D.J. Allstot, “Fast simulation of substrate
Note that the accuracy of DBEM can be further improved
coupling effects in mixed-mode ICs,” IEEE Custom Integrated
by partitioning the substrate boundary more densely. On the
Circuits Conference, pp. 18.3.1–18.3.4, 1993.
other hand, the efficiency of it may also be enhanced by [6] T. Smedes, N.P. van der Meijs, and A.J. van Genderen,
partitioning more coarsely. This allows a tradeoff of them, “Extraction of circuit models for substrate cross-talks,”
which makes the method adaptive in various circumstances. IEEE/ACM International Conference on Computer-Aided
Design, pp. 199-206, 1995.
[7] R. Ghapurey and R.G. Meyer, “Analysis and simulation of
V. Conclusions substrate coupling in integrated circuits,” International Journal
of Circuit Theory and Application, Vol. 23, pp. 381–394, 1995.
In this paper, we present a new direct boundary element [8] C. Xu, T. Fiez and K. Mayaram, “On the numerical stability
method (DBEM) to extract the impedances of lossy of Green’s function for substrate coupling in integrated
substrates at any frequency. Based on Maxwell equations, circuits,” IEEE Trans. Computer-Aided Design, Vol. 24, pp.
we deduce the Laplace equation that holds in the substrate 653–658, 2005.
and its boundary conditions. Utilizing the complex current [9] S. Donnay and G. Gielen, Substrate Noise Coupling in
continuity conditions on the interfaces of lossy mediums, the Mixed-Signal ASICs. Kluwer Academic Publishers, 2003.
new DBEM becomes suitable for impedance modeling. [10] S. Ardalan and M. Sachdev, “An overview of substrate
Three accelerating techniques are proposed without any noise reduction technieques,” Proc. 5th International
Symposium on Quality Electronic Design, pp. 291-296, 2004.
sacrifice of accuracy. Reusing matrix entries can save lots of
[11] X. Wang, W. Yu, Z. Wang, and X. Hong, “An improved
time for building matrices for impedance extraction at
direct boundary element method for substrate coupling
multiple frequencies. Condensing the linear system and resistance extraction,” Proceedings of the ACM Great Lakes
sparsifying the coefficient matrix can speed up system Symposium on VLSI , pp. 84-87, 2005.
solution remarkably. Numerical experiments illustrate that [12] W. Yu, Z. Wang, and J. Gu, “Fast capacitance extraction of
DBEM can be of high accuracy and high efficiency. DBEM actual 3-D VLSI interconnects using quasi-multiple medium
is able to simulate substrates with arbitrarily processes, and accelerated BEM,” IEEE Trans. Microwave Theory Tech., Vol.
numerical experiment results are reasonable. 51, pp. 109-199, 2003.
[13] S. Ramo, J.R. Whinnery, and T. van Duzer, Fields and
References Waves in Communication Electronics, 2nd ed. New York: Wiley,
pp. 279-283, pp. 572-676, 1984.
[1] R. Gharpurey and E. Charbon, “Substrate coupling: [14] Y. Saad and M.H. Schultz, “GMRES: A generalized mini-
Modeling, simulation and design perspectives,” Proc. 5th mal residual algorithm for solving nonsymmetric linear sys-
International Symposium on Quality Electronic Design, pp. tems,” SIAM J. Sci. Statis. Comput., Vol. 7, pp. 856-869, 1986.
283-290, 2004. [15] S. Ramo, J.R. Whinnery, and T. van Duzer, Fields and
[2] J.P. Costa, M. Chou, and L.M. Silveira, “Efficient Waves in Communication Electronics, 3rd ed. New York: Wiley,
techniques for accurate modeling and simulation of substrate pp. 283-287, pp. 572-676, 1994.