A Performance-Driven Router for RF and Microwave
Analog Circuit Design
Edoardo Charbon, Gary Holmlundy , Bruce Doneckery and Alberto Sangiovanni-Vincentelli
Department of Electrical Engineering and Computer Sciences,
University of California, Berkeley
y HP-EEsof R/D Labs, Santa Rosa
Abstract
Techniques are proposed for the routing of very high-frequency circuits.
In this approach, performance sensitivities are used to derive a set of
bounds on critical parasitics and to generate weights for a cost function
which drives an area router. In addition to these bounds, design often
requires that the length of interconnect lines be equal to predefined values.
The routing scheme enforces both types of constraints in two phases.
During the first phase all parasitic constraints are enforced on all nets.
Equality constraints are enforced during the second phase by expanding
each net simultaneously while ensuring that no additional violations to
parasitic constraints are introduced in the layout. During both phases
accurate and efficient parasitic estimations are guaranteed by compact
analytical models, based on 2-D and 3-D field analysis. Finally, a global
check on all distributed parasitics is performed. If the original constraints
are not satisfied, the weights are updated based on severity of the violation
and routing is applied iteratively. Several layouts synthesized using this
technique have been fabricated and successfully tested, confirming the
effectiveness of the approach.
1 Introduction
Great effort has been devoted over the years to create general purpose CAD
tools for schematic design and optimization of RF and microwave circuits, e.g.
[1, 2]. Layout synthesis has not received comparable attention in the literature
due to the inherent complexity of the problem and to the lack of designs
whose size could justify a CAD based approach. More recently however, the
increasing complexity of monolithic RF and microwave integrated circuits
(MMIC) suggests that an automated approach to the physical synthesis is
preferable for efficiency, reliability and yield considerations.
LINMIC [3] and later [1] proposed a knowledge-based interactive approach aimed at designers with minimum expertise by providing aids for low
frequencies design of MMICs. However, due to the lack of an explicit reference to performance, the synthesis process could result in a large number of
time-consuming iterations necessary to satisfy specifications.
Other semi-automated topology-driven approaches to the routing of
MMICs have also been proposed [4, 5]. These systems include templatebased routines for the abutment of pre-defined cells implementing devices
as well as interconnect. The knowledge of the relative position of all cells
provides the starting point of the layout realization, thus strongly limiting the
flexibility of the approach and its applicability to complex designs.
We propose a constraint-based approach to the layout synthesis of MMICs.
In this paper we focus on the routing problem. The flow diagram of the approach is shown in Figure 1. Firstly, high-frequency performance specifications are mapped onto a set of bounds on all classes of distributed parasitics.
Then, using sensitivity analysis, a set of weights is calculated for the area
router. The role of the weights is to control a cost function which penalizes
those realizations with highly critical parasitics. Layout synthesis of RF an
microwave circuits almost always requires that the dimensions of some interconnect lines be fixed. However, length constraints on interconnect cannot
be effectively enforced during this phase. Hence, the routing or constructive
phase is followed by by a refinement phase. The refinement consists on progressive expansion of all nets simultaneously thus allowing enforcement of all
net constraints while no new violations are created on the remaining parasitic
constraints. If infeasibility is detected a new set of weights is generated and the
cycle is repeated. At the completion of the layout, the entire circuit is checked
against constraint violations so as to verify that all performance specifications
are met. Parasitics are directly extracted from all physical geometries and
estimated by means of ad hoc analytical models based on 2-D and 3-D field
analysis. In case a specification violation occurs, the parasitics responsible
for the violation are identified and a sensitivity-based scheme is used to create
a new set of weights. The cycle is then restarted. The finite step loop ends
when all specifications are met.
There are several advantagesto this approach. A constraint-basedapproach
to the layout of RF and microwave circuits helps drastically reduce the number
design iterations by carefully modeling and controlling all relevant parasitic
effects in the circuit. For a given technology, compact and accurate models
for physical parasitics are derived only once. Hence the synthesis and analysis
of a circuit is efficient and can be effectively used within a larger semi- or
fully automated design cycle. Furthermore, using a different cost weighting
scheme, rapid circuit redesign for different realization and performance requirements can be efficiently accomplished. This is particularly useful in the
design of large scale MMIC libraries.
The paper is organized as follows. In Section 2 the models for interconnect used in our layout synthesis approach are presented and the process of
generating constraints on layout components is described. In Section 3 the
constructive and the refinement phases of the router are outlined and in Section
4 the practical suitability of the approach is demonstrated through an example.
2 Parasitic Modeling and Constraint
Generation
Accurate interconnect modeling is a fundamental requirement of a constraintdriven approach to the routing of RF and microwave circuits. For reasons
of efficiency closed formulae and analytical models for interconnect lines
and all relevant parasitics are desirable. Since the area of application of this
work is MMICs, all interconnect lines are modeled as microstrip transmission
lines. Parasitic effects such as inductive and capacitive crosstalk are modeled
im terms of the degradation induced on the characteristic impedance Zo and
loss . Alternatively at low frequencies discrete (R,C,L) parasitics can be
used. Analytical models of all considered parasitics are obtained by fitting
appropriate mathematical expressions to data obtained from 2-D or 3-D field
solvers as proposed by [6].
Interconnect discontinuities are modeled using discrete components, while
radiation and surface-wave propagation effects have been neglected due to
Specifications
Technology
Constraint Generation
Weight
Generation
no
Constructive
Phase
Refinement
Phase
constraints
met?
Parasitic
Extraction
Figure 1: Flow diagram of the tool.
yes
STOP
Z0
L1
L1
Z0
C0
1
L1
L1
L1
C0
L1
Z0
C0
Figure 3: Interconnect model for microstripline with multiple bends.
1
0
(a)
(b)
Figure 2: (a) Microwave specification. (b) Cost function for constrained
optimization.
the the relative small circuit size if compared with the signal wavelengths
[7]. If needed, surface-wave propagation could easily be modeled using the
closed forms reported in [8]. Substrate-dependent losses have been taken into
account in the full model. A summary of the formulae used in our approach
for estimating Zo and can be found in [9].
The constraint generation techniques outlined in [10] and [11] need be
modified to account for the distributed nature of parasitics in microwave
circuits. For a given a performance Ki , let us define performancespecification
the set of inequalities which determine lower- and upper-bounds for Ki and
the range of frequencies for which they are valid. For example, consider the
input reflection coefficient S11 illustrated in Figure 2 (a). The solid line is the
actual value of S11 obtained from the complete layout after extraction and the
dotted lines are the frequency dependent specifications to performance S11 ,
or
jL0 j jS11 j jU0 j; 6 L0 6 S11 6 U0 ; for f0 ? 4f0 f f0 + 4f0
jL1 j jS11 j jU1 j; 6 L1 6 S11 6 U1 ; for f1 ? 4f1 f f1 + 4f1
:::
jLn j jS11 j jUn j; 6 Ln 6 S11 6 Un ;
for fn ? 4fn f fn + 4fn :
f g
In general, for a set of performances Ki ; i = 1; :::; Nw and a set of parasitics pj ; j = 1; :::; Np , parasitic constraint generation is defined as the
process of creating an inequality constraint on a subset of all parasitics pj :
f g
f g
pj pj(bound) , to guarantee the fulfillment of all performance constraints in
the entire frequency range or 4Ki(f ) 4Ki (f ) ; 8 i = 1; : : : ; Nw ; 8 f 2
[fmin ; fmax ] . 4Ki (f ) represents the total and 4Ki (f ) the maximum allowed degradation of performance Ki with respect to all parasitics within the
entire range of operation. Parasitics such as cross-couplings, characteristic
impedance degradation, losses and microstrip line length tolerances are constrained in this fashion. Constraints on parasitics related to short and open
microtrip terminations are derived from the above constraints as described in
[9].
Assuming that performance is linear near its nominal value at all frequencies, it can be represented as a linear combination of its sensitivity with respect
to all distributed parasitics at all frequencies. Since parasitics can contribute
constructively as well as destructively to performance, positive and negative
sensitivities must be considered separately for constraint computation. Positive and negative components of the performance degradation Ki+ and
Ki? with respect to all parasitics belonging to the set pj are, in first
approximation
4
f g
4 Ki (f )+ =
4
X i ?
X i +
Sj (f ) pj
Sj (f ) pj ; 4Ki (f )? =
j
j
(1)
Technology-related process variations can be taken into account by replacing nominal values of sensitivities with worst-case values [10] or by finding
bounds on parasitic and device mismatches [11].
To insure that all relevant parasitics are considered during the optimization, each interconnect line to be implemented in the circuit is modeled as
a microstrip line with inductive and capacitive coupling with all the other
nets. A numerical sensitivity analysis is performed for each performance
function using simulator MNS [4]. All parasitics, whose cumulative contribution to performance degradation is negligible, are automatically discarded.
Bounds on critical distributed parasitics are calculated by using constrained
optimization, the objective being the maximization of the flexibility of the
layout generation process. The objective function of the optimizer, PMW ,
is a polynomial of second order monotonic in the range from zero to one, as
shown in Figure 2 (b). The goal is to obtain large bounds on critical parasitics, while at the same unnecessarily loose bounds on uncritical parasitics
are allowed to be tightened.
RF and microwave circuits rely for functionality on transmission lines with
a specific length. However tolerances L
j from nominal value Lj should
be permitted to provide more flexibility to the router:
4
`j = Lj ; 4`+j 4L+j ; and 4 `?j 4L?j ;
+=?
While Lj is determined by design, constraints 4Lj
must be computed
numerically. Using MNS the sensitivity of each performance with respect
to length variations can be quantified, thus, using constrained optimization,
constraints on the tolerances are calculated.
Parasitics involving the detailed realization of interconnect such as bends,
gaps and steps need be considered in a somewhat different way. A bended
interconnect realization is shown in Figure 3, where the bends have been
replaced with appropriate models [9]. Consider the problem of finding a
constraint on the number of bends in the microstrip line.
Suppose a microstrip line of nominal length L is partitioned into N segments of length L=N each. Let us model N 1 bends as in Figure 3. Due to
the repetitive character of the model, performance sensitivities with respect to
K of each bend are necessarily equal. Therefore
the discrete components SL=C
the cumulative degradation K of N bends is
?
4
N ?1
X
4K = (SLK1 L1 + SCK0 C0 ) = (N ? 1)(SLK1 L1 + SCK0 C0 )
i=1
Consequently only one (L1 ; C0 ) pair must considered by the optimizer. After
the optimization two scenarios are possible: (a) parasitics are not critical and
therefore they have been eliminated. In this case, given the maximum and
minimum values for C0 and L1 , N (bound) can be computed as following
N (bound) = 1 + minfbC0(max) =C0(min) c; bL(1 max) =L(1 min) cg
(min)
(min)
Note that if C0
= 0 and L1
= 0, the constraint on N is 1, thus it
can be neglected.
(b) parasitics are critical and bounded. In this case N (bound) can be obtained
by replacing the maximum value of C0 and L1 with its bound.
N (bound) = 1 + minfbC0(bound) =C0(min) c; bL(1 bound) =L(1 min) cg
3 Routing
For reasons of flexibility and algorithmic reliability, an area routing scheme
has been used in our approach. In this scheme all inequality constraints can be
easily implemented in the cost function driving the router. Equality constraints
however cannot be handled effectively via a cost function, due to the instability
they induce in the algorithm [12, Chp.3]. Thus, the routing scheme has been
partitioned into two phases. During the first phase, the constructive phase, all
inequality constraints, i.e. constraints related to interconnect parasitics, are
enforced. The second phase or refinement, enforces all equality constraints
while no new parasitic violations are introduced. Both phases have been
implemented in a tool called CORAL .
The first phase of the routing scheme consists of a maze router based on
the A* algorithm [13]. The A* algorithm is based on a heuristic estimation
of the cost of a path connecting the propagation node on the grid x and a
terminal target. An optimal path for a net j is defined as the path minimizing
a cost function f (x). f (x) is the estimate of the wiring length ` weighted by
a factor proportional to interconnect crowding Kc (x) and to the sum of all
violations to parasitic constraints.
f (x) = `(x)
Kc (x) + wjR V iol[RR0j (x)] + wjC V iol[CC0j (x)]
P C
Z V iol[Zj (x)] ;
1
V
iol
[
C
(
x
)]
+
w
w
jk
j
jk
C0 jk
Z0
1+
+
Source
Source
where
Rj (x) = integral of the estimated transmission line loss of net j at x
Cj (x) = integral of the estimated substrate capacitance of net j at x
Cjk (x) = integral of the estimated coupling between nets j and k at x
Zj (x) = local estimated characteristic impedance from nominal for net j at x
Every violation (V iol[:]) is calculated as the difference between the estimate
of the parasitic component (Section 2) and its pre-computed constraint, when
this value is positive, otherwise it is set to zero. Parameters R0 , C 0 and Z 0 are
C
and wjZ are specific weights, calculated
normalization factors. wjR , wjC , wjk
according to the equation
X Sji (f )?
:
wj = w 0
4Ki (f )?
i;f
Sji (f )+
+
4Ki (f )+
where Ki (f )+=? represents the specification of performance Ki at frequency f , Sji (f )+=? its sensitivity with respect to parasitic pj and w0 a
normalization factor.
The constructive phase yields interconnects with a minimized length `.
Since ` cannot be further reduced, the algorithm stops if loss and substrate
capacitance violations occur or if the maximum transmission line length L
is exceeded. This technique can also be used for implementing stubs simply
by creating a virtual target in an area where no parasitic violations can occur
[9]. The routing order of the nets is determined automatically giving priority
to the wiring of stubs and of nets on which parasitic constraints are tightest.
This is done to not compromise the ability of the router to meet all parasitic
constraints by routing first non-critical nets.
The refinement phase is responsible for the enforcement of all equality
constraints while insuring that no additional violation be introduced in the
layout. Refinement is only applied to the set of all nets for which at least a
violation exists, call C such set. Consider the constraint on microstripline
length of net j , Lj . Since previously obtained length is guaranteed to be
smaller than the constraint, the interconnect needs be expanded. However
the expansion must occur inside a space where no constraint violations are
possible. Call this space feasibility zone of net j , or j . j is defined as the
intersection of all spaces j (pk ), for which parasitic pk associated with net
j does not exceed its predefined constraint.
Assume that each parasitic can be expressed in form of a nth order polynomial n (x; s; V), where x is the location of a point in the interconnect
and s is the position of any objects responsible for pk . V is the vector of
all known design parameters (interconnect width, via size, etc.). Then, the
location of j (pk )’s boundaries is the locus of all x that solve
4
F F
B
P
B
)
P (x; s0 ; V) ? p(
=0;
(2)
for given object position s0 and parameter V. As illustration consider the
bound
n
k
effect of a via structure on the characteristic impedance of a microstrip line.
Given the physical dimensions of both objects, a model for the deviation
of microstrip impedance Zo is derived (Section 2). The resulting space
( Zo ) is a two-dimensional sphere centered in s0 with radius d0 , where
d0 is computed solving Equation (2) for a given constraint Zo(bound) .
Hence any implementations of the microstrip outside this space will satisfy
the constraint. Given Pj , the set of all constrained parasitics relevant to net j ,
j becomes
B4
F
4
4
F
j
=
\
pk
F
2
Pj
B (p ) :
j
Target
(a)
(b)
Figure 4: Distributed parasitics acting on the interconnect determine the
feasibility zone.
zone, the maximum possible number of rectangles r is selected for expansion,
where r = n4 . This ensures that the minimum possible horizontal expansion
is performed, thus maintaining the distance between interconnect and zone
boundary as large as possible. The interconnect can be expanded in two
opposite directions in two adjacent rectangles to maximize the length increase
`, or in two equal direction to minimize the number of bends.
Hence, a large number of combinations for the expansion rectangles exists
that obtains the desired expansion. The problem can be formulated as
b c
4
Find R, a subset of R, such that the total expansion
subject to the constraint on number of bends.
4` is maximized
This problem can be solved using exhaustive enumeration techniques or linear
programming. Due to the low number of vertices per zone, in C ORAL the
former solution has been adopted.
An expansion step
4 ` =K
e = jmin
2C
j
is selected,where K is a constant
proportional to the worst-case number of expansion steps of the algorithm.
For each net j the set of all expansion rectangles Rj is calculated and each net
is expanded by e. The expansion is continued until either a net j reaches its
nominal length Lj or a the location of a zone vertex changes. If the first event
occurs, net j is dropped from C , e is recomputed and the expansioncontinues.
Otherwise, Rj is recomputed for all nets in C and the expansion proceeds
until an infeasibility is detected, i.e. interconnect crosses its assigned zone.
If the latter case the partially expanded layout is analyzed for performance
violations, a new set of weights is generated and routing is repeated. All
parasitics responsible for the violation can be easily identified, thus the set of
weights associated with these parasitics can be modified so as to insure that
a new routing attempt with a modified cost function will reach a satisfactory
solution. In our approach, weights are increased by an amount wj: , inversely
proportional to the entity of the violation that the corresponding parasitics
induced. m0 is a normalization factor.
4
4w
:
j
=
1
?w X
m0
:
j
i;f
Sji (f )?
4Ki(f )? ? 4Ki (f )?
+
Sji (f )+
4Ki (f )+ ? 4Ki(f )+
The sequence of constructive/refinement steps terminates when no more
constraint violations are present in the layout or in case the maximum number
of iterations is reached.
4 Results
A number of analytical models for MMIC microstrip lines has been computed
for the GaAs technology in which the chips were designed. Table 1 shows a
k
The boundaries of feasibility zone j are often a complex function of relative
position and size of the surrounding layout objects and by the constraint
imposed on each parasitic pk as shown in Figure 4 (a) (dotted lines).
If the initial number of bends b exceeds the maximum allocated number
the refinement phase terminates. If, on the other hand, n = N (bound)
b 4, then the expansion algorithm is applied. Since only Manhattan
expansion is allowed, the feasibility zone for each net is horizontally sliced in
correspondence of each vertex as shown in Figure 4 (b). Between each pair
of horizontal cuts a rectangle is built of width equal to the minimum zone
diameter between the cuts. Call R the set of all the rectangles non-adjacent
to source and target obtained in this fashion. Since it is preferable to expand
interconnect as uniformly as possible in order to keep it at the center of the
feasible
polytope
Target
?
Vdd
INPUT
Vdd
Figure 5: Schematic of “pcn38”.
Vdd
OUT
Parasitic
parameter
substrate cap.
cross-over cap.
parallel line cap.
char. impedance
bend
via hole
C1
type of model
single models
1st ord. polynomial
2nd ord. polynomial
2nd ord. polynomial
logarithmic
lumped passive
1st ord. polynomial
414.8
6281.7
1888.3
2401.0 (HP9000/750)
correction factors
2nd ord. polynomial
2nd ord. polynomial
2nd ord. polynomial
3040.9
2543.7
36 hrs (HP9000/750)
C12
C12
w
Z
C; L1 ; L2
L1
4
4
4
adjacent line
adjacent pad/spiral
adjacent via
C1
w
Zw
Z
computed bounds have then been used by CORAL to route the pre-placed
circuit, the resulting layout is shown in Figure 7 and the extracted deviations
from nominal performance are listed in Table 3. The CPU time for the routing
phase was 348 sec on a DECstation 5000/240. As expected, all performance
specifications were met. Figure 6 shows the results of jS11 j, jS22 j and jS21 j
respectively.
generation time (sec)
5 Conclusions
Table 1: Analytical models used in synthesis for parasitic control.
Circuit
Frequency range (GHz)
# parasitic constr.
CPU (sec)
specs met
TWA
pcn38
Wilk
tq
1-10
20-40
1-20
2-7
39
72
145
131
749
2802
3450
2991
3/3
3/3
2/2
4/4
Table 2: Performance of a set of commercial RF benchmarks.
freq. range (GHz)
j j (dB)
j j (dB)
j j (dB)
S11
S22
S21
25-30
30-35
35-40
40-43
-0.15
+1.22
+0.54
-0.03
-1.43
+1.21
-0.54
+1.59
+1.57
-0.12
+1.24
+0.53
Acknowledgements
Table 3: Worst-case performance degradation form nominal of “pcn38”.
The authors would like to thank J. Perdomo of HP for providing some of
the examples presented in this work. The authors are also grateful to Enrico
Malavasi of University of Padua, Italy, Brad Brim, Brett Carver, Julie Chapman, Jerry Gladstone and Brian Hughes of HP for their support and useful
discussions.
summary of the models, along with the CPU time required for their generation
on a DECstation 5000/240 (See [9] for a detailed description of the models).
The via model required a large CPU time due to the high level of accuracy
selected for HFSS in the 3-D field analysis. Notice however that all models
have been generated for a given technology, hence no further computational
effort is needed for all future designs realized in the same technology.
The routing methodology and the CAD tools described in this paper has
been applied to a number of commercial microwave circuits. The circuits
were fabricated in HP’s GaAs technology and tested. (See Table 2)
The measured performance confirmed the predictions in full with a yield a
fraction of a percent lower than that of hand layout. The results of the layout
synthesis are shown hereafter in detail for one circuit. The discussion and
results of the synthesis of all tested benchmarks can be found in [9].
Consider the three-stage amplifier shown in Figure 5. The specifications
are illustrated in Figure 6. All relevant parasitic constraints were generated
by PMW in a total CPU time of 2454 sec on a DECstation 5000/240. The
−10
−5
−20
−10
References
30
−8
20
−15
10
−30
25
| S11 | / dB
43
25
| S22 | / dB
43
Figure 6: Performance of “pcn38”.
Figure 7: Final layout of “pcn38”.
25
| S21 | / dB
A novel technique for the routing of RF and microwave circuits has been
proposed. First, sensitivity analysis is performed on the circuit. Using sensitivities and constrained optimization a set of bounds on all critical parasitics
in the circuit is generated. These bounds are then directly used by the router,
which consists of two phases. During the first phase a sub-set of all constraints
is enforced. A cost function drives this routing step where violations to the
bounds are weighted differently according to their criticality. During the second phase a refinement is performed on all nets simultaneously so as to enforce
all remaining constraints. A final performance verification is performed on the
layout. In case one or more violations are detected, the weighting of violations
in the cost is modified based on their severity. Routing is iteratively performed
until all violations have been eliminated or infeasibility is detected. A number
of circuits has been fabricated and successfully tested, thus confirming the
suitability of the approach.
43
[1] R. H. Jansen, R. G. Arnonld and I. G. Eddison, “A Comprehensive CAD
Approach to Design of MMICs up to MM-Wave Frequencies”, MTT-T,
vol. 36, n. 2, pp. 208–219, February 1988.
[2] W. L. Williams R. C. Compton and B. Rutledge, “Puff, an interactive
Microwave Computer Aided Design Program for Personal Computers”,
in MTT-S, pp. 707–708, 1987.
[3] R. H. Jansen, “LINMIC: A CAD Package for the Layout-Oriented
Design of Single- and Multi-Layer MICs/MMICs up to mm-Wave Frequencies”, "Microwave Journal", pp. 151–161, February 1986.
[4] Hewlett-Packard Microwave and RF Design System, Manuals, Santa
Rosa Systems Division, Santa Rosa, CA, December 1992.
[5] J. F. Zürcher, “MICROS - A CAD/CAM Program for Fast Realization
of Microstrip Masks”, in MTT-S, pp. 481–484, 1985.
[6] U. Choudhury and A. Sangiovanni-Vincentelli, “An Analytical-Model
Generator for Interconnect Capacitances”, in Proc. IEEE CICC, pp.
861–864, May 1991.
[7] M. E. Goldfarb and A. Platzker, “Losses in GaAs Microstrip”, in MTT-S,
pp. 563–565, 1990.
[8] J. R. James and A. Henderson, “High-Frequency Behaviour of Microstrip Open-Circuit Terminations”, IEE Journal on Microwaves, Optics and Acoustics, vol. 3, n. 5, pp. 205–218, Sept 1979.
[9] E. Charbon, G. Holmlund, A. Sangiovanni-Vincentelli and B. Donecker,
“A Performance-Driven Router for RF and Microwave Analog Circuit
Design”, Memorandum UCB/ERL M94/40, UCB, May 1994.
[10] U. Choudhury and A. Sangiovanni-Vincentelli, “Constraint Generation
for Routing Analog Circuits”, in Proc. IEEE/ACM DAC, pp. 561–566,
June 1990.
[11] E. Charbon, E. Malavasi and A. Sangiovanni-Vincentelli, “Generalized Constraint Generation for Analog Circuit Design”, in Proc. IEEE
ICCAD, pp. 408–414, November 1993.
[12] N. J. Nilsson, Problem-Solving Methods in Artificial Intelligence,
McGraw-Hill, 1971.
[13] G. W. Clow, “A Global Routing Algorithm for General Cells”, in Proc.
IEEE/ACM DAC, pp. 45–51, 1984.