WHITE PAPER

MODELING AND SIMULATION OF JITTER IN PLL

FREQUENCY SYNTHESIZERS
TABLE OF CONTENTS
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

Frequency synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

Phase-domain noise model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

Jitter definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

Cyclostationary processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

PM jitter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

FM jitter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

Metrics for jitter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

Phase noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

Sources of jitter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

Thresholds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

Oscillator phase noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

Model of PLL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

Modeling PM jitter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

Modeling FM jitter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

Efficiency of models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

Characterizing jitter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

Characterizing PM jitter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

Characterizing FM jitter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

Simulation and analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
TABLE OF FIGURES
Figure 1 Block diagram of a frequency synthesizer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

Figure 2 Linear time-invariant phase-domain model of Figure 1 synthesizer . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

Figure 3 Long-term jitter (Ji ) for a PLL as a function of the number of cycles . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

Figure 4 Block diagram of VCO behavioral model with jitter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

Figure 5 Effects of jitter on reference signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

Figure 6 A limiter is applied to driven-block output to suppress noise outside of active transitions . . . . . . . . . 13

Figure 7 Closed-loop (CL) versus open-loop (OL) noise at VCO output when only the reference
oscillator exhibits jitter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

Figure 8 Closed-loop (CL) versus open-loop (OL) noise at VCO output when only the VCO exhibits jitter . . . . . 16

Figure 9 Closed-loop (CL) versus open-loop (OL) noise at VCO output when only the PFD/CP,
FDM , and FDN exhibit jitter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

Figure 10 Closed-loop versus open-loop noise performance of PLL individual components . . . . . . . . . . . . . . . . . 16

Figure 11 Long-term jitter (Ji ) of closed-loop PLL as a function of observation interval . . . . . . . . . . . . . . . . . . . . 16

LIST OF TABLES
Table 1: Characteristics of PM and FM jitter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

Table 2: Characteristics of VCO output relative to output of FDN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

TABLE OF LISTINGS
Listing I: Frequency divider that models PM jitter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

Listing II: PFD/CP model with PM jitter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

Listing III: Fixed-frequency oscillator with FM jitter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

Listing VI: VCO model with FM jitter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

Listing V: Quadrature differential VCO model that includes FM jitter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

Listing VI: Fixed-frequency oscillator with FM and PM jitter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

Listing VII: PFD/CP without jitter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

Listing VIII: VCO with FDN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

Listing IX: Limiter used to characterize PM jitter in binary signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

Listing X: MATLAB script used for computing S(m ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

Listing XI: Spectre netlist for PLL synthesizer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

 Modeling and Simulation of Jitter in PLL
Frequency Synthesizers
Ken Kundert

Abstract A methodology is presented for predicting more on the practical aspects. It presents all the informa-
the jitter performance of a Phase-Locked Loop (PLL) tion a designer would need to predict the noise and jitter of
using simulation that is both accurate and efficient. a PLL synthesizer. The jitter extraction methodology is
The methodology begins by characterizing the noise based on the commercially available SpectreRF simulator
behavior of the blocks that make up the PLL using [17,18] and presents behavioral models for Verilog-A, a
transistor-level simulation. For each block, the jitter is standard, non-proprietary analog behavioral modeling lan-
extracted and provided as a parameter to behavioral guage [6, 14]. Both SpectreRF and Verilog-A are options
models for inclusion in a high-level simulation of the to the Spectre circuit simulator [12], available from
entire PLL. This approach is efficient enough to be Cadence Design Systems.1
applied to PLLs acting as frequency synthesizers with
large divide ratios. II. FREQUENCY SYNTHESIS
The block diagram of a PLL operating as a frequency syn-
I. INTRODUCTION thesizer is shown in Figure 1 [8]. 2 It consists of a reference
Phase-locked loops (PLLs) are used in wireless receivers
to implement a variety of functions, such as frequency FD
f in
OSC
synthesis, clock recovery, and demodulation. One of the f ref 1/M PFD CP LF
major concerns in the design of PLLs is noise or jitter per- f fb
formance. Jitter from the PLL directly acts to degrade the FD VCO
noise floor and selectivity of a transceiver. 1/N
f out
Demir proposed an approach for simulating PLLs whereby
a PLL is described using behavioral models simulated at a
high level [1,2]. The models are written such that they Fig. 1. The block diagram of a frequency synthesizer.
include jitter in an efficient way. He also devised a power-
ful new simulation algorithm that is capable of character- oscillator (OSC), a phase/frequency detector (PFD), a
izing the circuit-level noise behavior of blocks that make charge pump (CP), a loop filter (LF), a voltage-controlled
up a PLL that is based on solving a set of nonlinear sto- oscillator (VCO), and two frequency dividers (FDs). The
chastic differential equations [3, 4]. Finally, he gave for- PLL is a feedback loop that, when in lock, forces ffb to be
mulas that can be used to convert the results of the noise equal to fin. Given a reference frequency fref, the frequency
simulations on the individual blocks into values for the jit- at the output of the PLL is
ter parameters for the corresponding behavioral models N
[5]. This approach provides accurate and efficient predic- f out = ----- f ref . (1)
M
tion of PLL jitter behavior once the noise behavior of the
blocks has been characterized. However, it requires the use
of an experimental simulator that is not readily available. 1. SpectreRF is currently the only commercial simulator that is suit-
able for characterizing the jitter of the blocks that make up a PLL. SPICE
This paper presents the relevant ideas of Demir, but while and its descendants are not suitable because they only perform noise
he focussed on presenting the conceptual aspects of mod- analysis about a DC operating point and so do not take into account the
time-varying nature of these circuits. Harmonic balance simulators do
eling and simulating jitter in PLLs, this paper concentrates perform noise analysis about a periodic operating point, which is a criti-
cal prerequisite, but they have convergence, accuracy, and performance
Unpublished Cadence Confidential manuscript. Last updated on Decem- problems with blocks such as the PFD/CP, FD and VCO that are strongly
ber 2, 1998. nonlinear.
This work was supported by the Defense Advanced Research Projects 2. Frequency synthesis is used as an example, but the concepts pre-
Agency under the MAFET program. sented are easily applied to other applications, such as clock recovery and
K. Kundert of Cadence Design Systems, San Jose, California can be FM demodulation. In addition, they are also applicable to other types of
reached via e-mail at [email protected]. PLL-based synthesis, such as fractional-N synthesis.

1
2 KUNDERT

By choosing the frequency divide ratios and the reference As , T fwd 0 because of the VCO and the low-
frequency appropriately, the synthesizer generates an out- pass filter, and so Tin, T det, T div 0 and T vco 1 . At
put signal at the desired frequency that inherits much of high frequencies, the noise of the PLL is that of the VCO.
the stability of the reference oscillator. In RF transceivers, Clearly this must be so because the low-pass LF blocks
this architecture is used to generate the local oscillator any feedback at high frequencies.
(LO) at a programmable frequency, which tunes the trans- As 0 , T fwd because of the 1 j term from
ceiver to the desired channel. the VCO. So at DC, Tin, T div N and Tvco 0 . At low
frequencies, the noise of the PLL is contributed by the
A. Phase-Domain Noise Model
OSC, PFD/CP, FD M and FD N , and the noise from the
If the signals around the loop are interpreted as phase, then VCO is diminished by the gain of the loop.
the small-signal noise behavior of the loop can be explored
Consider further the asymptotic behavior of the loop and
by linearizing the components and evaluating the transfer
the VCO noise at low offset frequencies ( 0 ) . Oscil-
functions. Figure 2 shows this phase-domain model.
lator phase noise in the VCO results in the power spectral
density S being proportional to 1/2, or S 1 2
in PFD/CP LF VCO vco vco
K 2K out (neglecting flicker noise). If the LF is chosen such that
+
det
----------- H ( ) vco
-------------------- H ( ) 1 , then T fwd 1 , and noise contribution from
2 j
the VCO to the output, T vco 2 S
vco , is finite and nonzero. If
det vco the LF is chosen such that H ( ) 1 , as it typically is
FDN when a true charge pump is employed, then T fwd 1 2

1
----
N
and the noise contribution to the output from the VCO
goes to zero at low frequencies.
div
III. JITTER DEFINITIONS
Fig. 2. Linear time-invariant phase-domain model of the Jitter is an uncertainty or randomness in the timing of
synthesizer shown in Figure 1. The out-board frequency divider
events. In the case of a synthesizer, the events of interest
is removed for simplicity. The s represent various sources of
noise.
are the transitions in the output signal. One models jitter in
a signal by starting with a noise-free signal v(t) and dis-
placing time with a stochastic process j(t). The noisy sig-
Define
nal becomes
K det 2K vco K det K vco H ( )
T fwd = ---------- H ( ) ------------------ = ----------------------------------- (2) vn ( t ) = v ( t + j ( t ) ) . (7)
2 j j
For simplicity, j is assumed to be a zero-mean Gaussian
as the forward gain of the loop. Then the transfer function
process, but it may be non-stationary. In addition, v will be
from the various noise sources to the output are
assumed to be T-periodic.
out Tfwd NT fwd
T in = --------- = ---------------------------- = --------------------- (3) There are two types of blocks that are used in the construc-
in 1 + T fwd N N + T fwd
tion of a PLL, driven blocks and autonomous blocks. Each
out N type exhibits a different type of jitter. Driven blocks, such
T vco = ---------- = --------------------- . (4) as the PFD, CP, and FD exhibit phase modulation, or PM
vco N + T fwd
jitter. Autonomous blocks, such as the OSC and VCO,
And by inspection, exhibit frequency modulation, or FM jitter. Table I pre-
out views the basic characteristics of PM and FM jitter, which
T div = --------- = T in (5) are discussed more fully after cyclostationary noise is
div
introduced.
and
out 2Tin A. Cyclostationary Processes
T det = --------- = -------------- . (6) A process is T-cyclostationary if its mean and autocorrela-
det K det
tion function (and hence, its variance) is bounded and T-
On this last transfer function, we have simply referred det periodic. Cyclostationary processes result from circuits
to the input by dividing through by the gain of the phase that are driven by a large periodic signal. Such a signal
detector. acts to modulate the noise generated by bias dependent
noise sources, such as shot noise sources present in for-
ward-biased semiconductor junctions or the thermal noise

DRAFT
MODELING AND SIMULATION OF JITTER IN PLL FREQUENCY SYNTHESIZERS 3

TABLE I small if j PM T K where T is the period of v and K is

CHARACTERISTICS OF PM AND FM JITTER. the highest significant harmonic of v.
Jitter Type Circuits J C. FM Jitter

driven var ( n v, t c ) FM jitter is exhibited by autonomous systems, such as

PM synchronous ----------------------------
- oscillators, that generate a stream of spontaneous output
(PFD/CP, FD) v ( t c ) transitions. In the PLL, the OSC and VCO exhibit FM jit-
autonomous ter. FM jitter is characterized by a randomness in the time
FM accumulating aT since the last output transition, thus the uncertainty of
(OSC, VCO)
when a transition occurs accumulates with every transi-
generated by nonlinear resistors. It also modulates the tion. Thus, compared with a jitter free signal, the fre-
transfer characteristics of the circuit from the noise source quency of a signal exhibiting FM jitter fluctuates
to the output. In a PLL synthesizer in steady-state, the randomly, and the phase drifts without bound in the form
OSC, PFD, CP, VCO, and FD all have a large periodic sig- of a random walk.
nal present and so exhibit cyclostationary noise.
One can construct a signal that exhibits FM jitter from a T-
Define T to be a T-cyclostationary process. If it is sam- cyclostationary process T using
pled every T seconds, the resulting process, {(kT)} where t
k = 0, 1, 2, 3, ..., is stationary. j FM(t) = 0 T ( ) d , (9)
The variance of flicker noise processes is unbounded and
so they are neither stationary nor cyclostationary. Flicker v n ( t ) = v ( t + j FM(t) ) . (10)
noise is not explicitly considered in this paper. It is not
conceptually difficult to include, but it serves to compli- While T is cyclostationary and so has bounded variance,
cate the presentation and flicker noise models tend to be from (9) it is clear that the variance of jFM, and hence the
expensive to implement in the time domain [3]. phase difference between v(t) and vn(t), can be unbounded.
If T is a white Gaussian T-cyclostationary process, then
B. PM Jitter v n(t) exhibits simple FM jitter. In this case, the process
PM jitter is a synchronous jitter exhibited by driven sys- {j FM(kT)} that results from sampling j FM every T seconds
tems. In the PLL, the PFD, CP, and FDs, all exhibit PM jit- is a discrete Wiener process and the phase difference
ter. In these components, an output event occurs as a direct between v(kT) and vn(kT) is a random walk [9].
result of, and some time after, an input event. PM jitter is a The essential characteristic of simple FM jitter is that the
random fluctuation in the delay between the input and the incremental jitter that accumulates over each cycle is inde-
output events. PM jitter is so named because it is a modu- pendent or uncorrelated. Autonomous circuits exhibit sim-
lation of the phase of the signal by a random process with ple FM jitter if they are broadband and if the noise sources
zero mean and bounded variance. Thus, the frequency of are Gaussian and small. The sources are considered small
the output signal exactly comensurate with that of the if the circuit responds linearly to the noise, though at the
input, but the phase of the output signal fluctuating ran- same time the circuit may be responding nonlinearly to the
domly with respect to that of the input. oscillation signal. An autonomous circuit is considered
PM jitter is modeled using (7), except j is replaced by j PM, broadband if there are no secondary resonant responses
v n ( t ) = v ( t + j PM ( t ) ) (8) close in frequency to the primary resonance.

where j PM = T and T is T-cyclostationary. If jPM is fur- D. Metrics for Jitter

ther restricted to be a white Gaussian T-cyclostationary Define ti be the sequence of times at which positive going
process, then vn(t) exhibits simple PM jitter. The essential zero crossings, henceforth referred to as transitions, occur
characteristic of simple PM jitter is that the jitter in each in vn(t). Define Tk = tk+1 tk and J(k) to be the standard
event is independent or uncorrelated. Driven circuits deviation of Tk,
exhibit simple PM jitter if they are broadband and if the
noise sources are Gaussian and small. The sources are J(k) = ( T k ) = var ( T k ) . (11)
considered small if the circuit responds linearly to the J(k) is referred to as period jitter. It is a measure of jitter
noise, though at the same time the circuit may be respond- over a single cycle and has units of time. This definition is
ing nonlinearly to the periodic drive signal. valid for both PM and FM jitter, but does not distinguish
As will be shown in Section IV, if j PM is white and small, between them.
then from (29), the variation in vn is also white. j PM is

CADENCE CONFIDENTIAL
4 KUNDERT

A metric that does distinguish between PM and FM jitter

is the long-term jitter Ji(k), the standard deviation of the FM jitter
from OSC
length of i adjacent periods from reference point k, FM jitter
from VCO
J i(k) = ( t i + k t k ) = var(t i + k t k) . (12) log(Ji)
PM Jitter
The period jitter J(k) is related to the long-term jitter Ji(k) J
with J(k) = J1(k). In steady state, the jitter is independent 0 fLT
log(i)
of k, and so the period and long-term jitter are denoted J
and Ji. Fig. 3. Long-term jitter (Ji) for a PLL as a function of the
To fully characterize an arbitrary jitter process, Ji must be number of cycles.
known for all i. However, for simple PM and FM jitter one
can determine Ji for all i simply by knowing J. of signal. One common way is S(f), the power spectral
density of . Related, but much less common, is
For simple PM jitter,
S (f) = ( 2f ) 2 S (f) , (20)
J i(k) = var(t i + k t k) , (13)
the power spectral density of = d/dt. S(f) and S(f) are
J i(k) = var(( i + k )T + j PM ( t i + k ) ( kT + j PM ( t k ) )) (14) conceptually useful, but are not directly measurable
because both and are not observable quantities. The
Ji = 2var(j PM) , (15)
other common measure is Sv(f), the power spectral density
where jPM (t) is T-cyclostationary and so jPM = jPM(tk) is of v, which is measurable with a spectrum analyzer. Often,
independent of k. The factor of 2 stems from the length of especially when measuring oscillator phase noise, S v (f)
an interval including the independent variation from two close to f c is given as a function of offset frequency and
transitions. From (15), Ji is independent of i, and so normalized relative to the power in the first harmonic of v,
J i = J for i = 1, 2, . (16) V1 (defined below in (25)),
S v(f c + f m)
For simple FM jitter, each transition is relative to the pre- L ( f m ) = ------------------------
- , (21)
2 V1 2
vious transition, and the variation in the length of each
period is independent, so the variance in the time of each where fm = f fc.
transition accumulates,
Ji = i J for i = 0, 1, 2, , (17) Relating S(f) and L(fm): Assume that (t) is small and
expand (19) using the Taylor series,
J = ( j FM(t k) ) = var(j FM(t k)) . (18) dv(t) ( t )
v n ( t ) = v ( t ) + ----------- ----------- . (22)
dt 2f c
The jitter behavior for PLLs is a combination of simple
PM and simple FM jitter [13]. If the PLL has a closed-loop Model the noise due to jitter using additive noise,
bandwidth of f L , and L = 1/2f L , then for i such that
v n(t) = v ( t ) + n v ( t ) (23)
iT L , jitter from the VCO dominates and the PLL
exhibits simple FM jitter. Similarly, if the reference oscil- dv(t) ( t )
n v ( t ) = ----------- ----------- . (24)
lator exhibits simple FM jitter, then for large i (low fre- dt 2f c
quencies), FM jitter again dominates. Between these two
extremes, the PLL exhibits PM jitter. The amount of which Since v(t) is periodic, it can be written as a Fourier series,

depends on the characteristics of the loop and the level of
PM jitter exhibited by the FDs and the PFD/CP. The
behavior of a typical PLL is shown in Figure 3.
v(t) = Vke
j2 k fc t
(25)
k =
IV. PHASE NOISE where j = 1 . Differentiating (25) and substituting into
Recall that v is a T-periodic function and reformulate j as (24),
phase, then (7) is rewritten as
(t )
(t )
v n(t) = v t + -----------
2f c
(19) n v ( t ) = -----------
2f c j2kf c V k e
j2kf c t
(26)
k =
where (t) = 2 fc j(t) is in radians and fc = 1/T. There are
several common ways of describing the noise in this type

DRAFT
MODELING AND SIMULATION OF JITTER IN PLL FREQUENCY SYNTHESIZERS 5

var ( n v, t c )
var ( j PM(t c) ) = ------------------------
-. (32)


nv ( t ) = ( t ) jkV k e
2kf c t
. (27) v ( tc )2
k = The jitter is computed from (31) or (32) using (16),
Computing the power spectral density of nv gives J = 2var ( j PM(t c) ) . (33)

Sn ( f ) =
v kV k 2 S ( f kf c ) (28) B. Oscillator Phase Noise
FM jitter is strongly related to oscillator phase noise. Both
k =
are different ways of describing the same underlying phe-
nomenon. In other words, all free running oscillators

S n ( fc + fm ) 1 exhibit a behavior that is generically referred to as oscilla-

kV k 2
L ( f m ) = --------------------------- --------- S ( f m ( k 1 )f c )
v
- = --- tor phase noise. Any noise in an autonomous system will
2 V1 2 2 V1 cause the phase to drift freely because there is no reference
k =
signal with which to lock. When the phase fluctuations are
(29)
measured in terms of time deviation over a period, it is
Assume that S(f) drops rapidly with increasing frequency, referred to as FM jitter. If it is measured in terms of noise
then for f m f c signal amplitude as a function of frequency, it is referred
1 to as oscillator phase noise.
L ( f m ) --- S ( f m ) . (30)
2
Relating J and S: In order to determine the period jitter J
Recall that several assumptions were made in developing
of vn(t) for a noisy oscillator, assume that it exhibits simple
this approximation. In particular, that (t) is small and that
FM jitter so that in (9) is a white Gaussian noise process
S at low frequencies is much larger than at frequencies
(this excludes flicker noise) with a PSD of
near fc and higher. The first approximation is not valid for
S ( f ) = a , (34)
phase noise of a free running oscillator for small f m
because (t) is undergoing a random walk, and so is and an autocorrelation function of
unbounded. R (t 1, t 2) = a ( t 1 t 2 ) , (35)
V. SOURCES OF JITTER where is a Dirac delta function. Then
t
A. Thresholds
In systems where signals are continuous valued, an event
j FM(t) = 0 T ( ) d (36)

is usually defined as a signal crossing a threshold in a par-

is a Wiener process [9], which has an autocorrelation func-
ticular direction. The threshold crossings of a noiseless
tion of
periodic signal, v(t), are precisely evenly spaced. However,
when noise is added to the signal, vn(t) = v(t) + nv(t), each R j (t 1, t 2) = amin ( t 1, t 2 ) . (37)
FM
threshold crossing is displaced slightly. Thus, a threshold
The variance of j FM over one period T is
converts additive noise to PM jitter.
var(j FM(t + T) j FM(t)) (38)
The amount of displacement in time is determined by the
amplitude of the noise signal, nv(t), and the slew rate of the = E [ ( j FM(t + T) j FM(t) ) 2 ]
periodic signal, dv(tc)/dt, as the threshold is crossed. If the = E [ j FM(t + T) 2 2j FM(t + T)j FM(t) + j FM(t) 2 ]
noise nv is stationary, then
= E [ j FM(t + T) 2 ] 2E [ j FM(t + T)j FM(t) ] + E [ j FM(t) 2 ]
var ( n v )
var ( j PM(t c) ) -----------------
- (31) = R j (t + T, t + T) 2R j (t + T, t) + R j (t, t)
v ( tc ) 2 FM FM FM
= a ( t + T ) 2at + at
where tc is the time of a threshold crossing.
= aT
Generally nv is not stationary, but cyclostationary. It is
only important to know when the noisy periodic signal Finally, jitter is the standard deviation of the variation in
vn(t) crosses the threshold, so the statistics of nv are only the period, and so
significant at the time when vn(t) crosses the threshold, J = aT . (39)

CADENCE CONFIDENTIAL
6 KUNDERT

We now have a way of relating the jitter of the oscillator to time when existing activity occurs. Thus, models with jit-
the PSD of . However, what is needed is to relate the jit- ter can run as efficiently as those without.
ter to either S or L. To do so, consider simple FM jitter
written in terms of phase, A. Modeling PM Jitter
t A feature of Verilog-A allows especially simple modeling
FM(t) = 2f c j FM(t) = 2f c T ( ) d 0 (40) of PM jitter. The transition() function, which is used to
model signal transitions between discrete levels, provides
and so from (34) and (40) the PSD of FM(t) is a delay argument that can be dithered on every transition.
The delay argument must not be negative, so a fixed delay
( 2f c ) 2 af c2
S (f ) = a -----------------
- - .
= ------- (41) that is greater than the maximum expected deviation of the
FM ( 2f ) 2 f2 jitter must be included. This approach is suitable for any
Thus, a is the noise power in FM where f = fc, or model that exhibits PM jitter and generates discrete-valued
outputs. It is used in the Verilog-A divider module shown
a = S (f c) . (42)
FM in Listing I, which models PM jitter with (8) where j PM is
Given S and fc, a is found with (41)3 and then J is found LISTING I
with (39). FREQUENCY DIVIDER THAT MODELS PM JITTER.

Relating S(f) and L(fm): The phase variation in a simple // Frequency Divider with Jitter
FM jitter process is not bounded, and so (29) cannot be include discipline.h
used to predict L(fm) of an oscillator. Demir shows that for include constants.h
a free-running oscillator exhibiting simple FM jitter module divider (out, in);
1 af c2 input in; output out; electrical in, out;
L ( f m ) = --- ----------------------------
-, (43)
2 a22f 4 + f 2
c m parameter real Vlo=1, Vhi=1;
parameter integer ratio=2 from [2:inf);
which is a Lorentzian with corner frequency of
parameter integer dir=1 from [1:1] exclude 0;
2
J // dir=1 for positive edge trigger
f corner = -----2- f c f c . (44) // dir=1 for negative edge trigger
T parameter real tt=1n from (0:inf);
At frequencies above the corner, parameter real td=0 from (0:inf);
parameter real jitter=0 from [0:td/5);
af c2 1
L ( f m ) = --------
- = --- S (f m) , (45) parameter real ttol=1p from (0:td/5);
2f m 2 2 FM // recommend ttol << jitter
integer count, n, seed;
which agrees with (30) and Vendelin [19]. Thus, in free-
real dt;
running oscillators at frequencies above the corner, L(fm)
and S(fm) are easily related.4 analog begin
@(initial_step) seed = 311;
VI. MODEL OF PLL // Phase / frequency detector state machine
The basic behavioral models for the blocks that make up a @(cross(V(in) (Vhi + Vlo)/2, dir, ttol)) begin
PLL are well known and so will not be discussed here in // count input transitions
any depth [1,2]. Instead, only the techniques for adding jit- count = count + 1;
if (count >= ratio)
ter to the models are discussed.
count = 0;
Jitter is modeled in an AHDL by dithering the time at n = (2count >= ratio);
which events occur. This is efficient because it does not // add jitter
create any additional activity, rather it simply changes the dt = 0.707jitter$dist_normal(seed,0,1);
end

3. In general, (42) is not used when computing a because S is not

// Charge pump
accurately known at fc. Instead, f is chosen where S is accurately known, V(out) <+ transition(n ? Vhi : Vlo, td+dt,tt);
and (41) is used. end
4. When using (45) and (41) to determine a, choose f well above the endmodule
corner frequency of (44) to avoid ambiguity and well below fc to avoid
the noise from other sources that occur at these frequencies. a stationary white discrete-time Gaussian random process.

DRAFT
MODELING AND SIMULATION OF JITTER IN PLL FREQUENCY SYNTHESIZERS 7

It is also used in Listing II, which models a simple PFD/ that could cause problems for the simulator. The jitter is
CP. embodied in dt, which varies randomly from transition to
LISTING II transition. The jitter J represents the variation in a period,
PFD/CP MODEL WITH PM JITTER. and dt the variation of a single transition, so from (16)
dt = ( j PM(t c) ) = 2J 0.707J , (46)
// Phase-Frequency Detector & Charge Pump
`include discipline.h which compensates for the fact that both ends of each
`include constants.h interval are varying. To avoid nonnegative delays, td must
always be larger than dt.
module pfd_cp (out, ref, vco);
input ref, vco; output out; electrical ref, vco, out; PFD/CP Model: The model for a phase/frequency detec-
parameter real Iout=100u; tor combined with a charge pump is given in Listing II. It
parameter integer dir=1 from [1:1] exclude 0; implements a finite-state machine with a three-level out-
// dir=1 for positive edge trigger put, 1, 0 and +1. On every transition of the VCO input in
// dir=1 for negative edge trigger direction dir, the output is incremented. On every transition
parameter real tt=1n from (0:inf); of the reference input in the direction dir, the output is dec-
parameter real td=0 from (0:inf); remented. If both the VCO and reference inputs are at the
parameter real jitter=0 from [0:td/5); same frequency, then the average value of the output is
parameter real ttol=1p from (0:td/5); proportional to the phase difference between the two, with
// recommend ttol << jitter the average being negative if the reference transition leads
integer state, seed; the VCO transition and positive otherwise [8]. As before,
real dt; the time of the output transitions are randomly dithered by
analog begin dt to model jitter. The output is modeled as an ideal current
@(initial_step) seed = 716; source and a finite transition time models the dead band in
the CP.
@(cross(V(ref), dir, ttol)) begin
if (state > 1) state = state 1;
B. Modeling FM jitter
dt = 0.707jitter$dist_normal(seed,0,1);
end
OSC Model: The delay argument of the transition() func-
@(cross(V(vco), dir, ttol)) begin tion cannot be used to model FM jitter because of the
if (state < 1) state = state + 1; cumulative nature of this type of jitter. When modeling a
dt = 0.707jitter$dist_normal(seed,0,1); fixed frequency oscillator, the timer() function is used as
end shown in Listing III. At every output transition, the next
I(out) <+ transition(Ioutstate, td + dt, tt); transition is scheduled using the timer() function to be
end T K + J K in the future, where is a unit-variance
endmodule zero-mean random process and K is the number of output
transitions per period. Typically, K = 2.
Frequency Divider Model: The model, given in Listing I,
operates by counting input transitions. This is done in the VCO Model: A VCO generates a sine or square wave
@cross block. The cross function triggers the @ block at whose frequency is proportional to the input signal level.
the precise moment when its first argument crosses zero in VCO models, given in Listings IV and V, are constructed
the direction specified by the second argument. Thus, the using three serial operations, as shown in Figure 4. First,
@ block is triggered when the input crosses the threshold the input signal is scaled to compute the desired output
in the user specified direction. The body of the @ block frequency. Then, the frequency is integrated to compute
increments the count, resets it to zero when it reaches ratio, the output phase. Finally, the phase is used to generate the
then determines if count is above or below its midpoint (n desired output signal. The phase is computed with idtmod,
is zero if the count is below the midpoint). It also gener- a function that provides integration followed by a modulus
ates a new random dither dT that is used later. Outside the operation. This serves to keep the phase bounded, which
@ block is code that executes continuously. It processes n prevents a loss of numerical precision that would other-
to create the output. The value of the ?: operator is Vhi if n wise occur when the phase became large after a long
is 1 and Vlo if n is 0. Finally, the transition function adds a period of time. Output transitions are generated when the
finite transition time of tt and a delay of td + dt. The finite phase passes /2 and /2.
transition time removes the discontinuities from the signal

CADENCE CONFIDENTIAL
8 KUNDERT

LISTING III Therefore,

FIXED FREQUENCY OSCILLATOR WITH FM JITTER. J i
i = -------- (49)
// Fixed-Frequency Oscillator with Jitter K
include discipline.h where is a zero-mean unit-variance Gaussian random
include constants.h process. The dithered frequency is
module osc (out); 1
------ fc
output out; electrical out;
1 1 K
f i = ---- ----------------- = ----------------- = -------------------------- (50)
K + i i 1 + K i f c
parameter real freq=1 from (0:inf); 1 + --------
parameter real Vlo=1, Vhi=1;
parameter real tt=0.01/freq from (0:inf); Let T i = K i , then
parameter real jitter=0 from [0:0.1/freq);
fc
integer n, seed; f i = ----------------------- . (51)
1 + T i f c
real next, dT;
analog begin Finally, var(i) = J2/K, and so i = J i K and
@(initial_step) begin T i = KJ i .
seed = 286; The @cross statement is used to determine the exact time
next = 0.5/freq + $realtime;
when the phase crosses the thresholds, indicating the
end
beginning of a new interval. At this point, a new random
@(timer(next)) begin trial i is generated.
n = !n;
dT = jitter$dist_normal(seed,0,1); The final model given in Listing IV. This model can be
next = next + 0.5/freq + 0.707dT; easily modified to fit other needs. Converting it to a model
end that generates sine waves rather than square waves simply
requires replacing the last two lines with one that com-
V(out) <+ transition(n ? Vhi : Vlo, 0, tt);
putes and outputs the sine of the phase. When doing so,
end
endmodule consider reducing the number of jitter updates to one per
period, in which case the factor of 1.414 should be
changed to 1.
Listing V is a Verilog-A model for a quadrature VCO that
Vin k mod 2 exhibits FM jitter. It is an example of how to model an
Vout oscillator with multiple outputs so that the jitter on the out-
J puts is properly correlated.
These models do not include the finite response time of a
Fig. 4. Block diagram of VCO behavioral model that includes real VCO. Those dynamics would be separated out and
jitter. included as part of the model for the LF.
The jitter is modeled as a random variation in the fre- C. Efficiency of the Models
quency of the VCO. However, the jitter is specified as a
Conceptually, a model that includes jitter should be just as
variation in the period, thus it is necessary to relate the
efficient as one that does not because jitter does not
variation in the period to the variation in the frequency.
increase the activity of the models, it only affects the tim-
Assume that without jitter, the period is divided into K
ing of particular events. However, if jitter causes two
equal intervals of duration = T / K = 1 / K f c . The fre-
events that would normally occur at the same time to be
quency deviation will be updated every interval and held
displaced so that they are no longer coincident, then a cir-
constant during the intervals. With jitter, the duration of an
cuit simulator will have to use more time points to resolve
interval is
the distinct events and so will run more slowly. For this
i = + i . (47) reason, it is desirable to combine jitter sources to the
is a random variable with variance degree possible.
var(T) J2 To make the HDL models even faster, rewrite them in
var ( ) = --------------- = ----- . (48) either Verilog-HDL or VHDL. Be sure to set the time reso-
K K
lution to be sufficiently small to prevent the discrete nature

DRAFT
MODELING AND SIMULATION OF JITTER IN PLL FREQUENCY SYNTHESIZERS 9

LISTING IV LISTING V
VCO MODEL THAT INCLUDES FM JITTER. QUADRATURE DIFFERENTIAL VCO MODEL THAT INCLUDES FM
JITTER.
// Voltage Controlled Oscillator with Jitter
// Quadrature Differential VCO with Jitter
include discipline.h
include constants.h include discipline.h
include constants.h
module vco (out, in);
module quadVco (PIout,NIout, PQout,NQout, Pin,Nin);
input in; output out; electrical out, in;
electrical PIout, NIout, PQout, NQout, Pin, Nin;
parameter real Vmin=0;
output PIout, NIout, PQout, NQout;
parameter real Vmax=Vmin+1 from (Vmin:inf);
input Pin, Nin;
parameter real Fmin=1 from (0:inf);
parameter real Fmax=2Fmin from (Fmin:inf); parameter real Vmin=0;
parameter real Vlo=1, Vhi=1; parameter real Vmax=Vmin+1 from (Vmin:inf);
parameter real tt=0.01/Fmax from (0:inf); parameter real Fmin=1 from (0:inf);
parameter real jitter=0 from [0:0.25/Fmax); parameter real Fmax=2*Fmin from (Fmin:inf);
parameter real ttol=1u/Fmax from (0:1/Fmax); parameter real Vlo=1, Vhi=1;
parameter real jitter=0 from [0:0.25/Fmax);
real freq, phase, dT;
parameter real ttol=1u/Fmax from (0:1/Fmax);
integer n, seed;
parameter real tt=0.01/Fmax;
analog begin
real freq, phase, dT;
@(initial_step) seed = 561;
integer i, q, seed;
// compute the freq from the input voltage
analog begin
freq = (V(in) Vmin)(Fmax Fmin)
@(initial_step) seed = 133;
/ (Vmax Vmin) + Fmin;
// compute the freq from the input voltage
// bound the frequency (this is optional)
freq = (V(Pin,Nin) - Vmin) * (Fmax - Fmin)
if (freq > Fmax) freq = Fmax;
/ (Vmax - Vmin) + Fmin;
if (freq < Fmin) freq = Fmin;
// bound the frequency (this is optional)
// add the phase noise
if (freq > Fmax) freq = Fmax;
freq = freq/(1 + dTfreq);
if (freq < Fmin) freq = Fmin;
// phase is the integral of the freq modulo 2
// add the phase noise
phase = 2M_PIidtmod(freq, 0.0, 1.0, 0.5);
freq = freq/(1 + dTfreq);
// update jitter twice per period
// phase is the integral of the freq modulo 2
// 1.414=sqrt(K), K=2 jitter updates/period
phase = 2*M_PI*idtmod(freq, 0.0, 1.0, 0.5);
@(cross(phase + M_PI/2, +1, ttol) or
cross(phase M_PI/2, +1, ttol)) begin // update jitter where phase crosses /2
dT = 1.414jitter$dist_normal(seed,0, 1); // 2=sqrt(K), K=4 jitter updates per period
n = (phase >= M_PI/2) && (phase < M_PI/2); @(cross(phase 3*M_PI/4, +1, ttol) or
end cross(phase M_PI/4, +1, ttol) or
cross(phase + M_PI/4, +1, ttol) or
// generate the output
cross(phase + 3*M_PI/4, +1, ttol)) begin
V(out) <+ transition(n ? Vhi : Vlo, 0, tt);
dT = 2*jitter*$dist_normal(seed,0,1);
end
i = (phase >= 3*M_PI/4) && (phase < M_PI/4);
endmodule
q = (phase >= M_PI/4) && (phase < 3*M_PI/4);
of time in these simulators from adding an appreciable end
amount of jitter. // generate the I and Q outputs
V(PIout) <+ transition(i ? Vhi : Vlo, 0, tt);
Including PM Jitter into OSC: From the discussion of the V(NIout) <+ transition(i ? Vlo : Vhi, 0, tt);
phase-domain model of the synthesizer, it is clear that one V(PQout) <+ transition(q ? Vhi : Vlo, 0, tt);
can easily combine the output-referred noise of FD M and V(NQout) <+ transition(q ? Vlo : Vhi, 0, tt);
FDN and the input-referred noise of the PFD/CP with the end
output noise of OSC. A modified fixed-frequency oscilla- endmodule
tor model that supports two jitter parameters and the
the divide ratio on jitter in the next section). The fmJitter
divide ratio M is given in Listing VI (more on the effect of
parameter is used to model the FM jitter of the reference

CADENCE CONFIDENTIAL
10 KUNDERT

LISTING VI LISTING VII

FIXED-FREQUENCY OSCILLATOR WITH FM AND PM JITTER. PFD/CP WITHOUT JITTER.

// Fixed-Frequency Oscillator with Jitter // Phase-Frequency Detector & Charge Pump

include discipline.h `include discipline.h
include constants.h `include constants.h
module osc (out); module pfd_cp (out, ref, vco);
output out; electrical out; input ref, vco; output out; electrical ref, vco, out;
parameter real freq=1 from (0:inf); parameter real Iout=100u;
parameter real ratio=1 from (0:inf); parameter integer dir=1 from [1:1] exclude 0;
parameter real Vlo=1, Vhi=1; // dir = 1 for positive edge trigger
parameter real tt=0.01ratio/freq from (0:inf); // dir = 1 for negative edge trigger
parameter real fmJitter=0 from [0:0.1/freq); parameter real tt=1n from (0:inf);
parameter real pmJitter=0 from [0:0.1ratio/freq); parameter real ttol=1p from (0:inf);
integer n, fmSeed, pmSeed; integer state;
real next, dT, dt, fmSD, pmSD;
analog begin
analog begin @(cross(V(ref), dir, ttol)) begin
@(initial_step) begin if (state > 1) state = state 1;
fmSeed = 286; end
pmSeed = 459; @(cross(V(vco), dir, ttol)) begin
fmSD = fmJittersqrt(ratio/2); if (state < 1) state = state + 1;
pmSD = pmJittersqrt(0.5); end
next = 0.5/freq + $realtime;
I(out) <+ transition(Iout state, 0, tt);
end
end
@(timer(next + dt)) begin endmodule
n = !n;
dT = fmSD$dist_normal(fmSeed,0,1); ing a logfile containing the length of each period. The log-
dt = pmSD$dist_normal(pmSeed,0,1); file is used in Section VIII when determining S VCO, the
next = next + 0.5ratio/freq + dT; power spectral density of the phase of the VCO output.
end
Recall that the PM jitter of FDM and FDN has already been
V(out) <+ transition(n ? Vhi : Vlo, 0, tt); included as part of OSC, so the divider model incorporated
end into the VCO is noiseless and the jitter at the output of the
endmodule noiseless divider results only from the VCO jitter. Since
oscillator, and the pmJitter parameter is used to model the the divider outputs one pulse for every N at its input, the
PM jitter of FDM, FDN and PFD/CP. PM jitter is modeled variance in the output period is the sum of the variance in
in the oscillator without using a nonzero delay in the tran- N input periods. Thus, the jitter at the output is N times
sition function. This is a more efficient approach because larger than the jitter at the input, or
it avoids generating two unnecessary events per period. To J FD = NJ VCO . (52)
get full benefit from this optimization, a modified PFD/CP
given in Listing VII is used. This model runs more effi- Thus, to merge the divider into the VCO, the VCO gain
ciently by removing support for jitter and the td parameter. must be reduced by a factor of N, the jitter increased by a
factor of N , and the divider model removed.
Merging the VCO and FDN: If the output of the VCO is After simulation, it is necessary to refer the computed
not used to drive circuitry external to the synthesizer, if the results, which are from the output of the divider, to the out-
divider exhibits simple PM jitter, and if the VCO exhibits put of VCO, which is the true output of the PLL. Clearly
simple FM jitter, then it is possible to include the fre- the jitter can be computed from (52).
quency division aspect of the FDN as part of the VCO by
To determine the affect of the divider on S (), square
simply adjusting the VCO gain and jitter. If the divide
both sides of (52) and apply (39)
ratio of FD N is large, the simulation runs much faster
a FD TFD
because the high VCO output frequency is never gener- a VCO T VCO = ------------------- . (53)
ated. The Verilog-A model for the merged VCO and FDN N
is given in Listing VIII. It also includes code for generat- TVCO = TFD / N, and so

DRAFT
MODELING AND SIMULATION OF JITTER IN PLL FREQUENCY SYNTHESIZERS 11

LISTING VIII aVCO = aFD (54)

VCO WITH FDN.
From (41),
// Voltage Controlled Oscillator with Jitter f2 f2
S VCO -----------
2
- = S FD -------
2
- (55)
include discipline.h f VCO f FD
include constants.h
Finally, fVCO = N fFD, and so
module vco (out, in);
SVCO = N 2 SFD. (56)
input in; output out; electrical out, in;
Once FDN is incorporated into the VCO, the VCO output
parameter real Vmin=0;
signal is no longer observable, however the characteristics
parameter real Vmax=Vmin+1 from (Vmin:inf);
parameter real Fmin=1 from (0:inf);
of the VCO output are easily derived from (52) and (56),
parameter real Fmax=2Fmin from (Fmin:inf); which are summarized in Table II.
parameter real ratio=1 from (0:inf); TABLE II
parameter real Vlo=1, Vhi=1; CHARACTERISTICS OF VCO OUTPUT RELATIVE TO THE OUTPUT
parameter real tt=0.01ratio/Fmax from (0:inf); OF FD N ASSUMING THE VCO EXHIBITS SIMPLE FM JITTER AND
parameter real jitter=0 from [0:0.25ratio/Fmax); THE FDN IS NOISEFREE.
parameter real ttol=1uratio/Fmax from (0:ratio/Fmax);
parameter real outStart=inf from (1/Fmin:inf); Frequency Jitter Phase Noise
real freq, phase, dT, delta, prev, Vout;
integer n, seed, fp; J FD
f VCO = Nf FD J VCO = --------- S = N 2 S
VCO FD
analog begin N
@(initial_step) begin
seed = 561; It is interesting to note that while the frequency at the out-
delta = jitter sqrt(2ratio); put of FD N is N times smaller than at the output of the
fp = $fopen(periods.m); VCO, except for scaling in the amplitude, the spectrum of
Vout = Vlo;
the noise close to the fundamental is to a first degree unaf-
end
fected by the presence of FDN. In particular, the width of
// compute the freq from the input voltage the noise spectrum is unaffected by FDN. This is extremely
freq = (V(in) Vmin)(Fmax Fmin) fortuitous, because it means that the number of cycles we
/ (Vmax Vmin) + Fmin; need to simulate is independent of the divide ratio N. Thus,
// bound the frequency (this is optional) large divide ratios do not affect the total simulation time.
if (freq > Fmax) freq = Fmax;
To understand why FD N does not affect the width of the
if (freq < Fmin) freq = Fmin;
noise spectrum, recall that while we started with a jitter
// apply the frequency divider, add the phase noise that varied continuously with time, j(t) in (7), for either
freq = freq / ratio; efficiency or modeling reasons we eventually sampled it to
freq = freq/(1 + dTfreq); end up with a discrete-time version. The act of sampling
// phase is the integral of the freq modulo 1 the jitter causes the spectrum of the jitter to be replicated
phase = idtmod(freq, 0.0, 1.0, 0.5); at the multiples of the sampling frequency, which adds
// update jitter twice per period aliasing. This aliasing is visible, but not obvious, at high
@(cross(phase 0.25, +1, ttol)) begin frequencies in Figure 8. However, especially with FM jit-
dT = delta $dist_normal(seed, 0, 1); ter, the phase noise amplitude at low frequencies is much
Vout = Vhi; larger than the aliased noise, and so the close-in noise
end spectrum is largely unaffected by the sampling. The affect
@(cross(phase + 0.25, +1, ttol)) begin of FDN is to decimate the sampled jitter by a factor of N,
dT = delta $dist_normal(seed, 0, 1); which is equivalent to sampling the jitter signal, j(t), at the
Vout = Vlo; original sample frequency divided by N. Thus, the replica-
if ($realtime >= outStart)
tion is at a lower frequency, the amplitude is lower, and the
$fstrobe( fp, %0.10e, $realtime prev);
aliasing is greater, but the spectrum is otherwise unaf-
prev = $realtime;
end fected.
V(out) <+ transition(Vout, 0, tt);
end
endmodule

CADENCE CONFIDENTIAL
12 KUNDERT

VII. CHARACTERIZING JITTER

v v
The switching nature of the blocks in a PLL prevents use
of the conventional noise analysis available from S PICE to
characterize the noise of any of the blocks in a PLL with Offset

the possible exception of the LF. The SPICE noise analysis Reference
Difference
operates by linearizing the block about a DC operating
point, which is not sufficient when the block exhibits t t
switching behavior. SpectreRF and the simulator devel- v v
oped by Demir both linearize the circuit about a time-vary-
ing operating point and compute the noise at the output of Offset
the block while taking into account both the effect of the
time-varying operating point on the bias-dependent noise Reference
Difference
sources and the time-varying nature of the transfer func-
tion from the noise source to the output. They differ in that t t
SpectreRF is constrained to operate on periodic circuits. In
addition, SpectreRF outputs noise as a function of fre- Fig. 5. The figures on the left show a reference signal, and one
quency averaged over a period, while Demirs simulator offset due to jitter. In both the top and bottom figures, the offset is
computes the output noise as a function of time and inte- the same (they have the same amount of jitter). The top signals
grated over all frequencies. have a smaller transition time than those on the bottom. The
figures on the right show the difference between the reference
Both simulators linearize about the operating point and signal and the offset signal. For the same amount of jitter, the
compute the noise as a post processing step. Thus, the maximum difference is larger when the transition time is smaller.
noise does not affect the operating point calculation and so If the difference is large enough to cause a nonlinear response,
the simulation will not be accurate if the noise is large the noise simulations will not be accurate. This problem becomes
enough to affect the large-signal behavior of the circuit. significant when the jitter is about the same size of the transition
Generally, the amplitude of the noise sources is quite small time or larger. It limits the number of cascaded thresholding
and so this is not a concern. However, in thresholding cir- stages that can be characterized at one time.
cuits, the noise present when the signal crosses the thresh-
old gets amplified tremendously. When cascading several the range is negligible. Thus, the noise should be at least
thresholding stages, the noise can be amplified to such a 40 dB down and dropping at the highest frequency simu-
degree that it does change the large signal behavior, mak- lated. Finally, integrate the noise power over frequency
ing the simulation inaccurate. This occurs in a FD imple- and apply Wiener-Khinchin Theorem [16] to determine

mented as a ripple counter with a large number of stages.
In such cases it is necessary to break the circuit down and
var ( n v ) = Sn ( f )df ,
v
(57)
only characterize the jitter of one or two stages at a time.
the total noise power [9], and apply (31).
The maximum number of stages that can be characterized
together is greater if the jitter is small relative to the transi- In general, the noise is strongly cyclostationary and so the
tion time of the circuit, as shown in Figure 5. above procedure is insufficient. When the noise is cyclo-
stationary the same procedure is used, except a gating
A. Characterizing PM Jitter function is applied to the output so that only the noise that
SpectreRFs PNoise analysis computes the time-average occurs near the threshold crossing is considered in the jit-
power spectral density of the noise at the output of the ter calculation. SpectreRFs PNoise analysis computes the
block. If this noise is stationary (as opposed to cyclosta- time-average of the noise at the output and it is not possi-
tionary), it is a simple matter to apply (31) to calculate the ble in general to post-process the PNoise results to deter-
jitter. Simply choose a representative set of periodic inputs mine the noise at the time of the threshold crossing.
to the block and use SpectreRFs Periodic Steady State Rather, a limiter is added to the output of the block and
(PSS) analysis to compute the steady-state response. This SpectreRF computes the noise at the output of the limiter.
computes the periodic operating point about which the The limiter, given in Listing IX, is designed to saturate
noise analysis is performed. It also gives dv(tc)/dt, the slew when the output of the block is outside a certain range to
rate of the output at threshold crossing. Apply SpectreRFs prevent any noise at the output from being considered
PNoise analysis to compute the noise power at the output except the noise present near when the signal crosses the
as a function of frequency. Choose the frequency range of threshold. The limiter is shown in Figure 6. The range of
the analysis so that the total noise at frequencies outside the limiter, VL and VH, is chosen such that the noise and

DRAFT
MODELING AND SIMULATION OF JITTER IN PLL FREQUENCY SYNTHESIZERS 13

pendent of T, so to reduce the number of sidebands

v v
DUT l needed, use T as small as possible. If maxsidebands is not
set sufficiently large, then the extracted value of jitter will
Limiter Region
increase as T decreases.
v
Saturated
VH Assume that the amplitude of the noise during each transi-
Output of DUT Active
tion is the same, then
2
VL nl T
Saturated
var ( n l, t c ) ------------------ (58)
t t t1 + t t2

where n l2 is the time average of the noise power at the

vl

Output of Limiter output of the limiter. SpectreRF computes the power spec-
tral density of the noise at the output of the limiter, and by
t
the Parsevals Theorem,
T

T S n ( f )df
nl 1
n l2 = lim ------ n l2 ( t )dt =
T
Noise at Output of Limiter . (59)
t
T 2T l

t t
t1 t2 If we further assume that tt = tt1 = tt2, then
VH V L
Fig. 6. When the device-under-test (DUT) exhibits v ( t c ) ------------------- . (60)
cyclostationary noise, a limiter is applied to output of driven tt
block to suppress noise present outside of the active transitions.
v(t) is the output of the block, vl(t) is the output of the limiter, and And from (32) and (33)
nl(t) is the noise at the output of the limiter. Noise on a waveform 2var ( nl , t c ) 2 n l2 T tt
is denoted by using a thick trace. The output of the limiter is only - ------------------ ------------------- ,
J = -------------------------------
(61)
noisy when it is inside its active region (when it is not limiting). v ( tc ) Kt t V H V L

LISTING IX 2 n l2 Tt t K
LIMITER USED TO CHARACTERIZE PM JITTER IN BINARY SIGNALS. J ---------------------------------
-, (62)
VH VL
// Simple Limiter where K is the number of transitions that occurred during
include discipline.h the period. If K = 2,
include constants.h n l2 Tt t
module limiter (out, in); J ----------------------- . (63)
VH VL
output out; input in; electrical out, in;
In practice, the noise away from the transitions is usually
parameter real Vlo=1, Vhi=1; much smaller than the noise during the transitions. So one
analog begin can usually achieve reasonably accurate results by apply-
// Place time-point at threshold crossings ing (62) or (63) even without using the limiter.
@(cross(V(in) Vlo) or cross(V(in) Vhi));
This general methodology for characterizing the PM jitter
// Determine the output of driven blocks with binary outputs is extended or clari-
if (V(in) < Vlo) fied for important special cases in the next few sections.
V(out) <+ Vlo;
else if (V(in) > Vhi) PM jitter of the PFD/CP: The PFD and CP work together
V(out) <+ Vhi; to generate a three-level discrete-valued signal (it takes the
else values 1, 0, and +1) whose time average is used as the
V(out) <+ V(in);
loop error signal. The average of this signal controls the
end
endmodule
VCO after it has been extracted by the LF.
There are two aspects of the PFD/CP that differ from the
the slew rate is approximately constant while the limiter is assumptions made above. First, the output of the CP is a
active. When running PNoise analysis, assure that the current, so the limiter and the equations given in the previ-
maxsidebands parameter is at least ten times larger than T/ ous section need to be adapted. Second, the output of the
tti for any i. This assures that the narrow noise pulses are CP has three distinct levels rather than the two assumed
adequately resolved by the PNoise analysis. Jitter is inde-

CADENCE CONFIDENTIAL
14 KUNDERT

above. Thus, the CP has a ternary output rather than a simulation should cover an interval long enough to allow
binary output. accurate Fourier analysis at the lowest frequency of inter-
If it is necessary to apply a gating function, care must be est (Fmin ). With deterministic signals, it is sufficient to
taken because of the ternary nature of the output. A simple simulate for K cycles after the PLL settles if Fmin = 1/TK.
limiter would allow the noise associated with the middle However, for these signals, which are stochastic, it is best
value to pass. So the simple limiter should be replaced to simulate for 10K to 100K cycles to allow for enough
with a dead-band limiter. This is a limiter with a dead band averaging to reduce the uncertainty in the result.
in the center of its input range. The dead band rejects noise One should not simply apply an FFT to the output signal
about the equilibrium point associated with the middle of of the VCO/FD N to determine L(f m) for the PLL. The
the three values. result would be quite inaccurate because the FFT samples
the waveform at evenly spaced points, and so misses the
PM Jitter of a FD: With ripple counters, one can only jitter of the transitions. Instead, L(fm ) can be measured
characterize a few stages at a time because of the issue with Spectres Fourier Analyzer, which uses a unique
shown in Figure 5. Thus, a long ripple counter chain has to algorithm that does accurately resolve the jitter [12]. How-
be broken into smaller chains, and characterized individu- ever, it is slow if many frequencies are needed and so is
ally. The total jitter for the ripple counter is then computed not well suited to this application.
by taking the square-root of the sum of the square of the
Unlike L(fm), S(f) can be computed efficiently. The Ver-
jitter on each stage.
ilog-A code for the VCO/FDN given in Listing VIII writes
Unlike in ripple counters, jitter does not accumulate with the length of each period to an output file named peri-
synchronous counters. Jitter in a synchronous counter is ods.m. Writing the periods to the file begins after an initial
independent of the number of stages. Rather, jitter of a delay, specified using outStart, to allow the PLL to reach
synchronous counter is the jitter of its clock along with the steady state. This file is then processed by Matlab from
jitter of the last stage. MathWorks using the script shown in Listing X. This
The input to counters are generally edge sensitive, and so script computes S (f), the power spectral density of ,
are only affected by jitter on either the positive-going or using Welchs method [15]. The frequency range is from
the negative-going clock transitions, but not both. How- fout/2 to fout/nfft. The script computes S(fm) with a resolu-
ever, this should not affect the way in which blocks are tion bandwidth of rbw.5 Normally, S (fm ) is given with a
characterized as long as the behavioral models for the unity resolution bandwidth. To compensate for a non-unity
dividers also have edge-sensitive inputs. Then the behav- resolution bandwidth, broadband signals such as the noise
ioral model of the dividers only react to jitter on the proper should be divided by rbw. Signals with bandwidth less than
edge and ignore jitter on the other edge. rbw, such as the spurs generated by leakage in the CP,
should not be scaled. The script processes the output of
B. Characterizing FM Jitter VCO/FD N. The results of the script must be further pro-
The noise of a free-running oscillator is dominated by cessed using the equations in Table II to remove the effect
phase noise, which is a random shifting of the frequency, of FDN.
and hence the phase, of the oscillation signal over time.
IX. EXAMPLE
The phase of an oscillator is subject to this variation
because it is free running: there is no drive signal with These ideas were applied to model and simulate a PLL act-
which to lock and so no synchronization between the sig- ing as a frequency synthesizer. A synthesizer was chosen
nals generated by the oscillator and any reference signal. with fref = 25 MHz, fout = 2 GHz, and a channel spacing of
200 kHz. As such, M = 125 and N = 10,000.
Oscillator phase noise and FM jitter are different ways of
describing the same underlying phenomenon, and so there The noise of OSC is 95 dBc/Hz at 100 kHz, which corre-
is a direct conversion between phase noise and FM jitter, sponds to 20 ps of FM period jitter. The noise of VCO is
as given in (39). There is no need to invoke the use of 48 dBc/Hz, which gives 6 ps of FM period jitter. The PM
thresholds or gating functions in order to make the conver- period jitter of the PFD/CP and FDs was found to be 2 ns.
sion. The FDs were included into the oscillators, which sup-
presses the high frequency signals at the input and output
VIII. S IMULATION AND ANALYSIS
5. The Hanning window used in the psd() function has a resolution
The synthesizer is simulated using the netlist from Listing bandwidth of 1.5 bins [10]. Assuming broadband signals, Matlab divides
XI and the Verilog-A descriptions in Listings VI-VIII, by 1.5 inside psd() to compensate. In order to resolve narrowband signals,
modifying them as necessary to fit the actual circuit. The the factor of 1.5 is removed by the script, and instead included in the
reported resolution bandwidth.

DRAFT
MODELING AND SIMULATION OF JITTER IN PLL FREQUENCY SYNTHESIZERS 15

LISTING X LISTING XI
MATLAB SCRIPT USED FOR COMPUTING S(fm). THESE RESULTS SPECTRE NETLIST FOR PLL SYNTHESIZER.
MUST BE FURTHER PROCESSED USING TABLE II TO MAP THEM TO
THE OUTPUT OF THE VCO. // PLL-based frequency synthesizer that models jitter
simulator lang=spectre
% Process period data to compute S(fm) ahdl_include osc.va // Listing VI
echo off; ahdl_include pfd_cp.va // Listing VII
nfft=512; % should be power of two ahdl_include vco.va // Listing VIII
winLength=nfft;
overlap=nfft/2; Osc (in) osc
freq=25MHz ratio=125 \
winNBW=1.5; % Noise bandwidth given in bins fmJitter=20ps pmJitter=2ns
PFD (err in fb) pfd_cp Iout=500ua
% Load the data from the file generated by the VCO C1 (err c) capacitor c=3.125nF
load periods.m; R (c 0) resistor r=10k
% output estimates of period and jitter C2 (c 0) capacitor c=625pF
T=mean(periods); VCO (fb err) vco Fmin=1GHz Fmax=3GHz \
J=std(periods); Vmin=4 Vmax=4 \
maxdT = max(abs(periodsT))/T; ratio=10000 jitter=6ps \
fprintf(T = %.3gs, F = %.3gHz\n,T, 1/T); outStart=10ms
fprintf(Jabs = %.3gs, Jrel = %.2g%%\n, J, 100J/T); JitterSim tran stop=60ms
fprintf(max dT = %.2g%%\n, 100maxdT);
fprintf(periods = %d, nfft = %d\n, length(periods), nfft); in err
Osc & VCO &
% compute the cumulative phase of each transition 125 PFD & CP 10,000
phases=2picumsum(periods)/T; c
fb
% compute power spectral density of phase
[Sphi,f]=psd(phases,nfft,1/T,winLength,overlap,linear);
% correct for scaling in PSD due to FFT and window
Sphi=winNBWSphi/nfft;
% plot the results (except at DC) 10
K = length(f);
20
semilogx(f(2:K),10log10(Sphi(2:K)));
title(Power Spectral Density of VCO Phase); 30
S (dB/Hz)

xlabel(Frequency (Hz));
40 OL
ylabel(S phi (dB/Hz));
rbw = winNBW/(Tnfft); 50
RBW=sprintf(Resolution Bandwidth = %.0f Hz (%.0f dB),
60
rbw, 10log10(rbw));
imtext(0.5,0.07, RBW); 70 CL

of the synthesizer. The netlist is shown in Listing XI. The 80

results (compensated for non-unity resolution bandwidth
300 Hz 1 kHz 3 kHz 10 kHz 30 kHz 100 kHz
(28 dB) and for the suppression of the dividers (80 dB))
are shown in Figures 7-10. The simulation took 7.5 min-
Fig. 7. Noise of the closed-loop PLL at the output of the VCO
utes for 450k time-points on a HP 9000/735. The use of a when only the reference oscillator exhibits jitter (CL) versus the
large number of timepoints was motivated by the desire to noise of the reference oscillator mapped up to the VCO
reduce the level of uncertainty in the results. The period frequency when operated open loop (OL).
jitter in the PLL was found to be 9.8 ps at the output of the
VCO. The long-term jitter is shown in Figure 11. There is some contribution from the VCO, but it is dimin-
The low-pass filter LF blocks all high frequency signals ished by the gain of the loop. In this example, noise at the
from reaching the VCO, so the noise of the phase lock middle frequencies is dominated by the PM jitter gener-
loop at high frequencies is the same as the noise generated ated by the PFD/CO and FDs. The measured results agree
by the open-loop VCO alone. At low frequencies, the loop qualitatively with the these expected results. The predicted
gain acts to stabilize the phase of the VCO, and the noise noise is higher than one would expect solely from the
of the PLL is dominated by the phase noise of the OSC. open-loop behavior of each block because of peaking in
the response of the PLL from 5 kHz to 50 kHz. For this

CADENCE CONFIDENTIAL
16 KUNDERT

0 0
VCO-OL
OL 10
10

S (dB/Hz)
S (dB/Hz)

PLL-CL
20
20
30 PFD/CP,FD-OL

30
40
CL OSC-OL
40 50

300 Hz 1 kHz 3 kHz 10 kHz 30 kHz 100 kHz 300 Hz 1 kHz 3 kHz 10 kHz 30 kHz 100 kHz

Fig. 8. Noise of the closed-loop PLL at the output of the VCO Fig. 10. Closed-loop PLL noise performance compared to the
when only the VCO exhibits jitter (CL) versus the noise of the open-loop noise performance of the individual components that
VCO when operated open loop (OL). make up the PLL. The achieved noise is slightly larger than what
is expected from the components due to peaking in the response
of the PLL.

25 OL
30 7ns
VCO-OL
35 5ns
S (dB/Hz)

40
CL 3ns
45
2ns
50 PFD/CP,FD-OL
1.5ns PLL
55 CL
1ns
60 OSC-OL

1 kHz 10 kHz 100kHz 700ps

10k 30k 100k 300k 1M 3M 10M

Fig. 9. Noise of the closed-loop PLL at the output of the VCO
when only the PFD/CP, FDM, and FDN exhibit jitter (CL) versus
the noise of these components mapped up to the VCO frequency Fig. 11. The long-term jitter (Ji) of the closed-loop PLL as a
when operated open loop (OL). function of observation interval, where the interval is measured
in the number of 500 ps periods i at the output of the VCO.
(These results do not agree well with expected results. This may
reason, PLLs used in synthesizers where jitter is important
be due to peaking in the response of the PLL, or it may be due to
are usually overdamped. some yet unidentified problem).

Notes to the Author simple divider model is required (no fractional N ) and one
We have merged the divider with the VCO, which allows us cannot observe the spurs created by PFD/CP
to rapidly simulate synthesizers with large multiplication
ratios and see very close in phase noise. If your synthe-
sizer does not have a large divide ratio, then it may require X. CONCLUSION
long simulation in order to see close in phase noise (in this
A methodology for modeling and simulating the jitter per-
case, try using Verilog or VHDL models rather than Verilog-
A. Also, we are inferring the noise in the VCO output by formance of phase-locked loops was presented. The simu-
inspecting every N -th transition, which in the case of the lation is done at the behavioral level, and so is efficient
example, N=10000. Thus, we cannot know the statistics of enough to be applied in a wide variety of applications. The
the noise far from the carrier. behavioral models are calibrated from circuit-level noise
Other issues from the merged VCO and divider are that a simulations, and so the high-level simulations are accu-
rate. Behavioral models were presented in the Verilog-A

DRAFT
MODELING AND SIMULATION OF JITTER IN PLL FREQUENCY SYNTHESIZERS 17

language, however these same ideas can be used to [7] D. FitzPatrick, I. Miller. Analog Behavioral Modeling with
develop behavioral models in purely event-driven lan- the Verilog-A Language. Kluwer Academic Publishers,
1997.
guages such as Verilog-HDL and VHDL.
[8] F. Gardner. Phaselock Techniques. John Wiley & Sons,
This methodology is flexible enough to be used in a broad 1979.
range of applications where jitter is important. Examples
[9] W. Gardner. Introduction to Random Processes: With Ap-
include, clock generation and recovery, sampling systems,
plications to Signals and Systems. McGraw-Hill, 1989.
over-sampled ADCs, digital modulation and demodulation
systems, and fractional-N frequency synthesis (though it is [10] F. Harris. On the use of windows for harmonic analysis with
the discrete Fourier transform. Proceedings of the IEEE,
not possible to merge the VCO and divider in this case). vol. 66, no. 1, January 1978.
[11] P. R. Gray and R. G. Meyer. Analysis and Design of Analog
ACKNOWLEDGMENTS Integrated Circuits. J. Wiley & Sons, Third Edition, 1993.
I would like to thank Alper Demir and Manolis Terrovitis [12] K. Kundert. The Designers Guide to SPICE and Spectre.
of the University of California in Berkeley for many Kluwer Academic Publishers, 1995.
enlightening conversations about noise and jitter. I would [13] J. McNeill. Jitter in Ring Oscillators. IEEE Journal of Sol-
also like to thank Mark Chapman, Masayuki Takahashi, id-State Circuits, vol. 32, no. 6, June 1997.
and Kimihiro Ogawa of Sony Semiconductor and Rich [14] Verilog-A Language Reference Manual: Analog Extensions
Davis, Frank Hellmich and Randeep Soin of Cadence to Verilog-HDL, version 1.0. Open Verilog International,
Design Systems for their probing questions and insightful 1996. Available from www.ovi.org.
comments, as well as their help in validating these ideas [15] A. Oppenheim, R. Schafer. Digital Signal Processing.
on real frequency synthesizers. Prentice-Hall, 1975.
[16] A. Papoulis. Probability, Random Variables, and Stochas-
REFERENCES tic Processes. McGraw-Hill, 1991.
[1] H. Chang, E. Charbon, U. Choudhury, A. Demir, E. Felt, E. [17] R. Telichevesky, K. Kundert, J. White. Receiver character-
Liu, E. Malavasi, A. Sangiovanni-Vincentelli, and I. Vassil- ization using periodic small-signal analysis. Proceedings of
iou. A Top-Down Constraint-Driven Methodology for Ana- the IEEE Custom Integrated Circuits Conference, May
log Integrated Circuits. Kluwer Academic Publishers, 1996.
1997. [18] R. Telichevesky, K. Kundert, J. White. Efficient AC and
[2] A. Demir, E. Liu, A. Sangiovanni-Vincentelli, and I. Vas- noise analysis of two-tone RF circuits. Proceedings of the
siliou. Behavioral simulation techniques for phase/delay- 33rd Design Automation Conference, June 1996.
locked systems. Proceedings of the IEEE Custom Integrat- [19] G. Vendelin, A. Pavio, U. Rohde. Microwave Circuit De-
ed Circuits Conference, pp. 453-456, May 1994. sign. J. Wiley & Sons, 1990.
[3] A. Demir, E. Liu, and A. Sangiovanni-Vincentelli. Time-
domain non-Monte-Carlo noise simulation for nonlinear STILL TO DO
dynamic circuits with arbitrary excitations. IEEE Transac-
tions on Computer-Aided Design of Integrated Circuits and 1. Include flicker noise from oscillator.
Systems, vol. 15, no. 5, pp. 493-505, May 1996.
2. Model nonlinear dependence of VCO noise, amplitude,
[4] A. Demir, A. Sangiovanni-Vincentelli. Simulation and
and frequency to the control signal.
modeling of phase noise in open-loop oscillators. Proceed-
ings of the IEEE Custom Integrated Circuits Conference, 3. Generate small-signal noise model of PLL for use in
pp. 445-456, May 1996. PNoise analysis (includes correlations).
[5] A. Demir, A. Sangiovanni-Vincentelli. Analysis and Simu- 4. Describe how to model dead-zone in PFD/CP.
lation of Noise in Nonlinear Electronic Circuits and Sys-
tems. Kluwer Academic Publishers, 1997. 5. Describe how to see the spurs created by the PFD/CP
[6] A. Demir, A. Mehrotra, J. Roychowdhury. Phase noise in (keep the frequency divider in the circuit; if the divide
oscillators: a unifying theory and numerical methods for ratio is very large, consider partitioning it into two factors
characterization. Proceedings of the 35th Design Automa- and make one explicit and the other implicit).
tion Conference, June 1998.

CADENCE CONFIDENTIAL
Cadence Design Systems, Inc.
Corporate Headquarters
2655 Seely Avenue
San Jose, CA 95134
800.746.6223
408.943.1234
www.cadence.com

2001 Cadence Design Systems, Inc. All rights reserved. Cadence and the Cadence logo are
registered trademarks and how big can you dream? is a trademark of Cadence Design
Systems, Inc. All others are properties of their holders.
Stock #3487 10/01