Driver Models For Timing and Noise Analysis: Bogdan Tutuianu and Ross Baldick
Driver Models For Timing and Noise Analysis: Bogdan Tutuianu and Ross Baldick
Driver Models For Timing and Noise Analysis: Bogdan Tutuianu and Ross Baldick
Abstract:
In the recent years, the impact of noise on chip level signals has become a significant source of static timing
errors. This paper presents a new technique to generate accurate non-linear driver models which can be used for static
timing and noise analysis, with inductive interconnect and multi-source nets. The new technique is efficient because it
relies on existent gate characterization for timing, does not require additional non-linear circuit simulations and gen-
erates re-usable models.
Introduction:
One of the problems that has gathered much attention recently is the effect of switching
noise on chip level timing (delay noise or dynamic noise). Static timing analysis determines the
extremes of signal propagation, being the main tool used for predicting the speed performance of
the digital ICs. Since switching noise can overlap with and affect logic signals, it will directly
impact the chip level timing and the reliability of the final product. A good description of the dif-
ferent types of noise, their impact on circuit activity and ways to model and analyze it is given in
[23]. Other tools and methodologies for functional noise analysis are proposed in [19], [10] and
[1]. Special circuit modeling techniques to asses global noise impact have been proposed in [12],
[13] and [18].
The impact of switching noise on chip level timing is generally split into functional noise
and delay noise. Functional noise is noise induced in quiet nets (victims) by switching neighbors
(aggressors). For high levels of induced currents, it can cause unwanted logic activity and even
functional failures. The delay noise is caused by the same switching activity on the neighboring
nets, but it happens while the victim net is itself active. In this case the noise can modify the time
of flight and slew-rate of the useful signal and it can cause delay (timing) errors.
Switching noise analysis is performed in two steps: a first stage where all possible aggres-
sors are considered, some being filtered based on functional constraints, clock domains, timing
windows, etc., and a second stage where the actual effect of noise on delay is being determined
through circuit simulation. Most of the research in this area has been focused on the first step,
quiet victim noise pulse switching victim delay noise
victim victim
input near end input near end
aggressor far end aggressor far end
In all the examples shown in the paper we have used the same victim driver, a medium
sized 4-input NAND gate from a Motorola MPC755 PowerPC-Compatible microprocessor. The
gate has 48 transistors, 6 diodes, 725 parasitic capacitors and 322 resistors. In all examples, the
active input signal has been A1 while all others (A0, A2 and A3) were tied to Vdd. The gate is
driving a long wire routed on the upper layers of metal (M5). For the aggressor net, we have used
A0
A1 X near end victim wire far end
+ A2 gate
- A3 capacitive coupling load
(~40%)
+
-
aggressor wire
Figure 2: Test circuit used throughout this paper. A NAND4 gate driving a very
long wire coupled to an aggressor. Each nets load is within the range specified for
its driving gate.
a strong inverter as driver and it has been routed along to victim at minimum spacing (about 40%
of total wire capacitance is coupling to its neighbor). The two nets were coupled for 75% of the
victim nets length. The aggressor signal has been offset such that its effect is overlapping the far
end victim signal obtained in the absence of noise. In Figure 3, the victim input signal and the vic-
voltage
0 20 40 60 80 100 0 20 40 60 80 100
time time
a) far end signals b) delay and functional noise
Figure 3: a) The effect of switching noise on delay b) Comparison between func-
tional noise and delay noise.
tim far end signal without noise are shown. In addition, the functional noise has been determined
with the victim driver holding low. Then the far end signal in the presence of noise has been deter-
mined using accurate spice simulation of the full circuit (with delay noise waveform) and by
adding the functional noise to the far end signal without noise (with functional noise wave-
form). It is apparent from figure 3 that functional noise can be a very poor estimate of delay noise.
The last stage of the gate, the driving port, can be seen as comprised of two variable resis-
tors: one modeling the P FETs and one for the N FETs. These two resistors will have opposite
variation during the transition of the output and, as a consequence, any noise pulse injected in the
output pin will see this variable resistive path to ground. In Figure 4 we are showing the equiva-
1000 966.2
resistance (ohms)
800
600
400
200
108.8 170.4
0
0 20 40 60 80 100
time
Figure 4: The equivalent resistance seen at output pin during output transition.
lent driver resistance seen at the output port during the output transition (for the input-output sig-
nals pair shown in Figure 3). Note how the resistance varies between 100 and 1000 ohms during
the active interval. Since one of the existing methods to model the driver for delay noise [24]
relies on computing a transient holding resistance, it is clear from Figure 4 that such a single
value resistor model will not be able to accurately capture the complex driver behavior.
2 Background
In section 2.1 we are giving a succinct presentation of the different linear driver models
currently used in timing and noise analysis. In Section 2.2 we give a brief introduction to the
Galerkin method for Finite Elements, which is being used at the core of our modeling technique.
2.1 Logic gate models for static timing and noise analysis.
In the timing pre-characterization process of a logic block, detailed simulations of all the
possible signal paths are performed for different input signals and output loads. The delay mea-
surements are stored in table format or even post-processed as delay equations. This delay data
(equations) is usually generated for simple output capacitive loads. However, due to interconnect
resistance and inductance, the output load of the gate is modeled by complex RLC circuits which
vary from simple models to high order models. During timing analysis, the simple delay data
(equations) is used to generate driver equivalent circuit models using an iterative process (often
called C-effective algorithm). These models were first developed by [21], later their accuracy has
been greatly improved by [8].
In the C-effective algorithm, the driver is modeled by a Thevenin like circuit: an ideal volt-
age source - step or saturated ramp- and a driver equivalent resistance (Figure 5). The iterative
driver I
model Rdrv Rw
wire load
driver driver Vdrv +- w
model
inputs output Cw_1 Cw_2
u w sinks
driver ICeff
model Rdrv
w effective
Vdrv +- Ceff load model
Figure 5: Driver logic gate is modeled (using C-effective) by a Thevenin like linear
circuit while the interconnect input impedance is modeled (using AWE) by a
model and transfer functions between driver output and sink pins.
procedure tries to determine an effective output capacitance load such that for a specific time
interval the total charge stored on the simple capacitance is the same as the total charge stored on
the complex load and the delays and rise times derived from pre-characterized data for this simple
effective capacitance match the ones obtained through the simulation of the linear driver model
and the RC load:
t1 t1
Q = I ( t ) dt = Q Ceff = I Ceff ( t ) dt ,
t0 t0
T D = GateTD ( T X u, Ceff ) T X = GateTX ( T X u, Ceff )
where Q is charge, I is current, TD is delay time, TX is rise time, GateTD and GateTX are the gate
delay and rise time coming from pre-characterized data.
In practice, the C-effective technique is stable and converges rapidly and it has been in use
for almost a decade with different flavors in most static timers, commercial as well as corporate
EDA tools. Switching noise effects, on-chip interconnect inductance and multiple source nets are
relatively recent issues for static timing and there has been significant work done to extend this
simple algorithm to these cases. In [2] the authors are extending the algorithm to general output
load models in reduced order format [20][14]. Later [17] and [16] developed an extension of the
RC model to a stable RLC model with good accuracy for chip level timing.
For delay noise analysis there are a couple of models proposed in the literature:
1) An extension of the C-effective algorithm to model the injected current as additional capac-
itive load [9]. The C-effective algorithm is applied simultaneously to the victim and the aggressor
resulting in a system of non-linear equations solved efficiently using the successive-chord
method. It is worth noting that the same algorithm can be applied to the situation of a single net
with multiple sources.
2) The transient holding resistance model proposed in [24] which models the reaction of the
gate to the injected current with the help of a fitted resistance. Each iteration contains the follow-
ing steps: a) for each aggressor in isolation (with victim and other aggressors grounded) the cur-
rents injected in the victim is recorded; b) a non-linear simulation is performed to determine the
response of the gate with the induced current at its output. From the comparison of this output
with the one obtained in the absence of noise a delay noise pulse is obtained; c) a transient resis-
tance value for the victim driver is then computed to match the area of the delay noise with a func-
tional noise pulse.
In [24] the authors have compared the two methods and reported much better results for
the later technique. It is important to note that, in order to be accurate, the second modeling tech-
nique requires a non-linear circuit simulation at each iteration.
b j t
tM
( j ( t ) g ( t ) + j ( t ) ( c 1 g ( t ) + c 0 g ( t ) f ( t ) ) ) dt = 0 b j (6)
m
j=1
which is valid for any bj values and that gives us m independent equations:
tM
t m
( j ( t ) g ( t ) + j ( t ) ( c 1 g ( t ) + c 0 g ( t ) f ( t ) ) ) dt = 0 j = 1, , n (7)
n
If the approximation of the actual solution is g ( t ) = ai i ( t ) , equation (7) becomes:
i=1
n n n
tM
tm j ( t ) ai i ( t ) + j ( t ) c1 ai i ( t ) + c0 ai i ( t ) f ( t ) dt = 0 (8)
i=1 i=1 i=1
j = 1, , m
n
ai t
tM
( j ( t ) i ( t ) + j ( t ) ( c 1 i ( t ) + c 0 i ( t ) ) ) dt =
tM
m
t m
j ( t ) f ( t ) dt
(9)
i=1
j = 1, , m
Equation (4) has been reduced to a system of linear equations (9) which can be written as
Q a = r where Q = [ q ij ] is a m n matrix with its entries defined as:
in & jm
tM
q ij = t m
( j ( t ) i ( t ) + j ( t ) ( c 1 i ( t ) + c 0 i ( t ) ) ) dt , (10)
T T
a = a1 an is the vector of scalar coefficients for the solution and r = r 1 r m is the
vector of right hand side terms with each entry defined as:
tM
rj = t m
j ( t ) f ( t ) dt . (11)
At this point, the remaining step is to evaluate the integrals of equations (10) and (11),
usually done through Gaussian integration.
There is a trade-off between the number of nodes and the accuracy of the approximation.
In order to keep the computational cost low, the domain is usually split into elements as
= E p . This allows us to use low order basis functions and low order Gaussian integration.
p
In the case when there are two variables (as in our models), the most flexible domain par-
tition is obtained with triangular elements. However, this is cumbersome for the automated mod-
eling process and rectangular elements are used instead. The Lagrange interpolating polynomials
in two dimensions are also straightforward to obtain as products of one-dimensional polynomials.
In Figure 7 we are showing a rectangular element with 9 nodes and the expression of a two-
y
y3
2, 3 ( x, y ) = 2 ( x ) 3 ( y )
E
y2
( x x1 ) ( x x3 ) ( y y1 ) ( y y2 )
= ------------------------------------------- -------------------------------------------
y1
x1 x2 x3 x ( x2 x1 ) ( x2 x3 ) ( y3 y1 ) ( y3 y2 )
Figure 7: An example of a rectangular element (E) for the two-dimensional case.
The basis function ((x,y)) that takes value 1 in the point (x2,y3).
dimensional basis function derived using second order one-dimensional basis functions. In Figure
8 we are plotting two of these basis functions, 1, 1 ( x, y ) and 2, 2 ( x, y ) . The use of two-dimen-
z z
1 1
1, 1 ( x, y ) 2, 2 ( x, y )
0
0 0 0
0 1 y 0 1 y
1 1
x x
Figure 8: Two dimensional basis functions using second order one-dimensional
Lagrange interpolating polynomials.
sional Lagrange interpolating polynomials on rectangular elements (bi-linear, bi-quadratic basis
functions, etc.) guarantees the continuity of the approximation at the boundary between elements.
3. Non-linear driver models for timing and noise analysis.
The switching noise pulses inject/draw charge in/from the victim net, effectively changing
the size of the interconnect load seen by the victim net driver. As a consequence, the driver
response depends simultaneously on the input signal and the noise pulse and it is not possible to
separate these effects without incurring errors. Our solution is a simple non-linear model which
has either a Thevenin or a Norton form. In the following, the Thevenin type model is used to
present the modeling process and its properties.
This Section is divided in two sub-sections: the main steps of our modeling technique are
presented in section 3.1 followed by a discussion on the properties of our models in section 3.2.
PFET Rd
u w u w
Vd(u,w)
+
NFET Cload - Cd Cload
u1 w w2 w3 w
1
u2
u3
u D
Figure 10: A simple voltage source model defined on a grid as a function of the
input (u) and output (w) signal values as a PWL function of u and w.
under the difference operator) while we actually have the solution, but we do not know the func-
tional coefficients of the equation (i.e. Vd). In some sense, we are applying the FE method in
reverse. Our goal is to find a representation of V d ( u, w ) which satisfies the differential equation
at the measurement points (ui,wj). If the values V ij = V d ( u i, w j ) are known, then its approxima-
tion is:
V d ( u, w ) = V ij ij ( u, w ) i = 1, , N u and j = 1, , N w , (14)
i, j
where each ij is the two-dimensional Lagrange interpolation polynomial. The same interpolation
process is applied to all other time dependent functions: u, w and . For a particular input-output
signal pair, Figure 11, the time domain is partitioned by the measurement points. For the end
ti0 ti10 ti50 ti90 ti100 input signal
voltage representative
time points
V ij ij
k k
where k = 1, , I + J . Every equation can be concisely written as: = 0
i, j
which is part of the system of equations: V = where all the coefficients are ordered in
the matrix , all the unknown voltage points V ij are ordered in the vector V and all the free terms
are in .
tmax tmax tmax
= w j k j dt R d C
k k
ij = k ij ( u, w ) dt k j d t (20)
tmin j tmin tmin
By solving the system of equations (20) we can obtain the set of voltage points that define our
voltage source model.
The number of equations obtained in this process must be related to the number of ele-
ments needed for the Vd function. Since one input-output signal pair will provide a limited num-
ber of equations, we have to extend the analysis to more than one pair. It is easier to understand
that by visualizing every input-output signal pair as a path in the input-output domain D. In Figure
12.a we are showing a typical set of paths for an inverter with rising input and falling output.
These paths can be obtained by varying the input signal and/or the output pin capacitive loads. In
order to cover the lower left region of the domain we need to take other paths into account with
falling input and rising output (Figure 12.b). In order to better model the hold-up and hold-down
resistances of our model we need to better cover the lower right and upper left corners of the
domain which can be done with static noise signals on the input. Their distinctive paths are shown
in Figure 12.c. Note that the points known from measurements (marked with black squares) are
situated on one or both of the measurement thresholds. In the case of noise characterization, other
1 1 1
.9 .9 .9
output (w)
output (w)
output (w)
.5 .5 .5
.1 .1 .1
0 0 0
0 .1 .5 .9 1 0 .1 .5 .9 1 0 .1 .5 .9 1
input (u) input (u) input (u)
a) b) c)
Figure 12: Input-output signal pairs as paths through the domain of an invertor: a)
rising input falling output paths, b) falling input rising output paths and c) static
(positive and negative) noise paths for output holding high and low.
measurement rules can be applied. For example we are interested in the peak value of noise both
on input and output. Another point easy to describe is the point where input noise and output
noise pulses have the same height (points situated on the diagonal of the domain).
It is apparent from the distribution of points that we may need more accurate models of the
gate in some regions of the domain while others are sparsely populated and/or used. The D
domain can be split into elements in various ways. It is more efficient and more accurate to use the
measurement thresholds as boundaries between elements because in that case we have precise
information about the time-points at which the paths are traversing the element boundaries.
The variety of basis functions and the flexibility in the choice of a domain partition pro-
vides us with the adequate means of controlling the accuracy of our models. Depending on the
application, the user can choose to fit a model to a larger number of data points (equations) and
can use curve-fitting techniques such as Singular Value Decomposition to generate optimal (in the
least square sense) driver models.
0.5
1
Vout
Figure 13: Variation of output port current w.r.t. input and output pin voltages
(center). Contour plots of the output current for fixed input pin voltage levels (left)
and fixed output pin voltage levels (right).
1
Vinp
0.5 Iout=0
0 0.5 1
Vout
Figure 14: The convergence points of the original NAND4 gate which correspond
to the points where the dc output gate current is 0.
ure 15 we are showing the port current of our model which has been obtained for a one element
Iout Iout Iout
0.004 0.004 0.004
1
Vout
Figure 15: Variation of output pin current with the input and output pin voltages
(center) for the non-linear driver model of the NAND4 gate. Contour plots of the
output current for fixed input pin voltage levels (left) and fixed output pin voltage
levels (right).
partition of the domain and using second order Lagrange interpolating polynomials as base func-
tions. From the contour plots it can be seen that the port current is not monotonic inside the
domain and that results in multiple operating points for the same input-output voltage pair.
In Figure 16 the convergence curve of the driver port model and stability region(s) are
1.5
Vinp
0.5
Iout=0
-0.5
-0.5 0 0.5 1 1.5
Vout
Figure 16: The convergence points of the driver model, the stable model domain
(dark shaded area) and the absolute stability domain (light shaded area).
shown.
It is in general desirable to have a close match between the original convergence curve and
the model because that impacts the steady state accuracy which is important in cases when multi-
ple drivers are driving simultaneously the interconnect (see Example 5 from Section 4). From Fig-
ure 16 it is also apparent the impact of the holding resistance value. Our choice for Rd was the
hold down resistance value (108ohms) and the model tries to compensate with current in the hold
high case where the actual resistance is larger. However, in Example 2 of Section 4 showing the
hold high and hold low functional noise pulses, we can see (Figure 19) that the accuracy in both
cases is comparable.
Another important issue for our model is the domain of stability. For example, in our case
the basis functions are second order Lagrange polynomials and for any input voltage value there
are exactly two points where the port current is zero. One point is the convergence point and is
characterized by:
d
I ( u, w ) = 0 and I ( u, w ) < 0 , (22)
dw
and the other one is the limit of the stable region and is characterized by:
d
I ( u, w ) = 0 and I ( u, w ) > 0 . (23)
dw
The stable region boundary is marked in Figure 16 by the border of the dark shaded areas. The
light shaded areas are marking the boundary of the absolutely stable region. The points in this
region are characterized by:
d
I ( u, w ) 0 (24)
dw
in which the port current source offers a negative feedback with respect to the output voltage vari-
ation. In general, the situation in which the model has regions of instability inside its domain is
the result of sparse measurements data present in those regions.
4. Results
In this section we are presenting some results obtained with our proposed model. We are
showing for comparison the performance of our model in the case of basic timing signal propaga-
tion, functional noise and delay noise. We are also exemplifying the robustness of our model in
the case when input signals are outside the characterization range (over-shoot and under-shoot)
and with highly inductive interconnect. We present the performance of the model in a multi-
source net case and how the steady state is captured.
Using the test case described in Section 2, the new modeling technique has been used to
characterize the timing arc of our NAND4 gate from the input pin A1 to output pin X. We have
used a Thevenin type model on a domain with 1 two-dimensional element (similar to the one
shown in figure 9) characterized by 9 points (a 3x3 grid), two of them with known values, the hold
high V d ( 0, 1 ) = 1 and hold low V d ( 1, 0 ) = 0 conditions. So, 7 points were unknowns in the
characterization process. We have used 8 input-output signal pairs, 4 for rising output and 4 for
falling output, with 2 equations for each pair (one for 0% to 50% and one for 10% to 90%). The
Rd and Cd values have been determined using a simple small signal analysis on the output port
with the gate set-up to hold low.
Example 1: The first example is the test used in Figure 3. The near and far end waveforms
in the case without noise are shown in Figure 17.a. The near and far end waveforms in the case
1.0 1.0
0.8 0.8
voltage
0.6
voltage
-0.05
-0.10 our model
-0.15
0 20 40 60 80 100
time
Figure 18: Delay noise comparison between our model and actual driver.
Example 2: In Figure 19 we are showing the accuracy of the model for functional noise
estimation. Our model is compared with the actual gate and the hold-up/down resistors.
actual driver actual driver
0.0 with hold-down resistor 0.0 with hold-up resistor
-0.05 -0.05
voltage
voltage
our model our model
-0.10 -0.10
-0.15 hold-down noise -0.15 hold-up noise
0 20 40 60 80 100 0 20 40 60 80 100
time time
Figure 19: Functional noise at sink pin (far end) with the actual driver, holding
resistance model and our model for holding down (left) and holding up (right).
Example 3: One of the more difficult cases to model in static timing (for the C-effective
algorithm) is the gate response with inductive output loads. In order to show the robustness of the
driver model, a large amount of inductance has been added to the interconnect RC model. The
simulation results for the near and far end nodes of the interconnect wire are shown in Figure 20.
1.0 1.0
0.8 0.8
voltage
1.0 1.0
0.8 near end signals 0.8 near end signals
0.6 0.6
voltage
voltage
voltage
0.6
C 0.4 O1
C
0.2 O4
O3 O4 0.0
I3 I4
0 20 40 60 80 100
time
Figure 22: Signal propagation in a multi-source large RC grid (left). The driver
inputs, outputs and the signal in the center of the grid are shown (right).
40 wire sections driven from the four corners. Two of the drivers have a significant input offset to
magnify the voltage division across the interconnect. In Figure 22.b the waveforms at the first and
last driver outputs and in the center of the grid are shown, both with real drivers and with our
models. During the first part of the response, the two active drivers are driving each an amount of
capacitance outside their model characterization range but good accuracy is maintained.
Example 6: We have used the algorithm proposed in [24] to generate the transient hold-
ing resistance for comparison with our model. Through full net simulation, a transient holding
resistance has been determined (712.8ohms) such that the area of the noise pulse with resistance
model matches within 0.004% the area of the real delay noise pulse. The superposition of the
quiet response and the noise pulse with resistance model produces the approximation of the noise
impact on delay. The errors are tabulated in Table 1:
In Figure 23, all the near end point delay noise pulses (using actual driver model, our
non-linear model and transient holding resistance model) are shown. The plot spans 150 time
units which is the interval used for matching the noise pulse areas.
0.00 transient holding
resistance model
voltage
-0.05
-0.10 actual driver
-0.15 our model
In this paper we have proposed a new technique to model logic gates for timing and noise
analysis. The proposed models have quite a few distinct advantages over the existent driver mod-
els for timing and/or noise:
The modeling process is using the already existent measurements data generated for static
timing analysis for each logic block. No new data or special characterization work is needed.
No non-linear spice simulations are required in the modeling process.
The models are simple Thevenin/Norton-like circuits with voltage/current sources dependent
on the input and output pin voltages and are represented using elementary functions (polyno-
mials). This makes their simulation extremely efficient.
The models have variable accuracy both in terms of the range of input rise time and output
capacitance load that is being covered and in terms of the error with respect to the actual mea-
surement values used in the process.
The models are covering large ranges of input rise time and output capacitive load and they
are re-usable (do not depend on a particular input-output situation or noise pulse).
The models are very robust and maintain accuracy outside the characterization range.
The examples presented in Section 4 are showing the versatility of the proposed model.
We have demonstrated the accuracy of our model in different situations of practical interest:
normal signal propagation (static timing analysis) with very good behavior throughout the
characterization domain,
simulation of the driver response with complex output load models including inductance,
computation of the delay variation due to switching noise,
functional noise analysis,
simulation of special cases such as nets with multiple drivers with significant time offsets and
complex interconnect models.
As future work, our attention is focused on circuit simulation. One draw-back of our mod-
els is that in order to simulate them we need a non-linear circuit simulator. The driver models that
we have used are piece-wise polynomial models. These models can be simulated very efficiently
by available special purpose simulators (such as ACES [11]). One can also make a simple obser-
vation that the same FE method used to generate the models can be used to simulate them and, in
conjunction with reduced order interconnect models, one can develop a very efficient simulation
engine.
References:
[1][Ain00] - Aingaran K. et al. Coupling noise analysis for VLSI and ULSI circuits, IEEE First
International Symposium on Quality Electronic Design 2000, page(s) 485-489.
[2][Aru97] - Arunachalam R., Dartu F. and Pileggi L.T. CMOS gate delay models for general
RLC loading IEEE 1997 page(s) 224-229.
[3][Aru00] - Arunachalam R., Rajagopal K. and Pileggi L.T. TACO: timing analysis with cou-
pling Design Automation Conference 2000, page(s) 266-269.
[4][Aru01] - Arunachalam R., Blanton R.D. and Pileggi L.T. False coupling interactions in static
timing analysis Design Automation Conference 2001, page(s) 726-731.
[5][Bec81] - Becker E.B., Carey G.F. and Oden J.T. Finite elements - an introduction - Prentice-
Hall Inc. 1981.
[6][Che97] - Chen W., Gupta S.K. and Breuer M. Analytic models for crosstalk delay and pulse
analysis under non-ideal inputs International Test Conference 1997, page(s) 809-
818.
[7][Che99] - Chen P. and Keutzer K. Towards true crosstalk analysis International Conference
on CAD 1999, page(s) 132-137.
[8][Dar96] - Dartu F., Menezes N. and Pileggi L.T. Performance computation for pre-character-
ized CMOS gates with RC loads IEEE Transactions on CAD, vol. 15, issue 5,
May 1996, page(s) 544-553.
[9][Dar97] - Dartu F. and Pileggi L.T. Calculating worst-case gate delays due to dominant capac-
itance coupling Design Automation Conference 1997, page(s) 46-51.
[10][Del00] - Delaurenti M. et al. Switching noise analysis framework for high speed logic fam-
ilies 14th International Conference on VLSI Design 2000, page(s) 524-530.
[11][Dev94] - Devgan A. and Rohrer R.A. Adaptively controlled explicit simulation IEEE
Transactions on CAD, vol.13, no.6, Jun.1994, page(s) 746-762.
[12][Dev97] - Devgan A. Efficient coupled noise estimation for on-chip interconnects Digest of
Technical Papers, International Conference on CAD 1997, page(s) 147-151.
[13][Fel97] - Feldman P. and Freund L.W. Circuit noise evaluation by Pade approximation based
model-reduction techniques Digest of Technical Papers, International Conference
on CAD 1997, page(s) 132-138.
[14][Fre98] - Freund R.W. Reduced-order modeling techniques based on Krylov sub-spaces and
their use in circuit simulation Numerical analysis manuscript No. 98-3-02, Bell
Laboratories, Feb. 1998.
[15][Gro98] - Gross P.D. et.al. Determination of worst-case aggressor alignment for delay calcu-
lation International Conference on CAD 1998, page(s) 212-219.
[16][Kas00] - Kashyap, C.V. and Krauter, B.L. A realizable driving point model for on-chip
interconnect with inductance Design Automation Conference, 2000, page(s) 190-
195.
[17][Kra99] - Krauter, B.L., Mehrotra S. and Chandramouli V. Including inductive effects in
interconnect timing analysis IEEE Custom Integrated Circuits Conference 1999,
page(s) 445-452.
[18][Kuh01] - Kuhlmann M. and Sapatnekar S.S. Exact and efficient crosstalk estimation IEEE
Transactions on CAD, vol.20, no.7, July 2001, page(s) 858-866.
[19][Lev00] - Levy R. et al. ClariNet: a noise analysis tool for deep sub-micron design Design
Automation Conference 2000, page(s) 233-238.
[20][Oda98] - Odabasioglu A., Celik M. and Pileggi L.T. PRIMA: Passive reduced-order inter-
connect macro-modeling algorithm IEEE Transactions on CAD, vol. 17, issue 8,
Aug.1998, page(s) 645-654.
[21][Qia94] - Qian J., Pullela S. and Pillage L.T. Modeling the effective capacitance of RC
interconnect IEEE Transactions on CAD, vol. 13, issue 12, Dec. 1994, page(s)
1526-1535.
[22][Sas00] - Sasaki Y. and DeMicheli G. Crosstalk delay analysis using relative window
method ASIC/SOC Conference 1999, page(s) 9-13.
[23][She99] - Shepard K.L., Narayanan V. and Rose R. Harmony: static noise analysis of deep
submicron digital integrated circuits IEEE Transactions on CAD, vol.18, no.8,
Aug.1999, page(s) 1132-1150.
[24][Sir01] - Sirichotiyakul S. et al. Driver modeling and alignment for worst-case delay noise
Design Automation Conference 2001, page(s) 720-725.
[25][Xia00.1] - Xiao T., Chang C.-W. and Marek-Sadowska M. Efficient static timing analysis in
presence of crosstalk ASIC/SOC Conference 2000, page(s) 335-339.
[26][Xia00.2] - Xiao T. and Marek-Sadowska M. Worst delay estimation in crosstalk aware
static timing analysis International Conference on Computer Design 2000,
page(s) 115-120.