GENERAL NON LINEAR PERTURBATION MODEL OF PHASE NOISE IN LC OSCILLATORS DISSERTATION Presented in Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy in the Graduate School of The Ohio State University By Jayanta Mukherjee, MS ***** The Ohio State University 2006 Dissertation Committee: Patrick Roblin, Adviser Steven Bibyk Joanne Degroat Approved by Adviser Graduate Program in Electrical and Computer Engineering
130
Embed
GENERAL NON LINEAR PERTURBATION MODEL OF PHASE …roblin/lsna/dissertation.pdf · GENERAL NON LINEAR PERTURBATION MODEL OF PHASE NOISE IN LC OSCILLATORS DISSERTATION Presented in
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
GENERAL NON LINEAR PERTURBATION MODEL OFPHASE NOISE IN LC OSCILLATORS
DISSERTATION
Presented in Partial Fulfillment of the Requirements for
the Degree Doctor of Philosophy in the
Graduate School of The Ohio State University
By
Jayanta Mukherjee, MS
* * * * *
The Ohio State University
2006
Dissertation Committee:
Patrick Roblin, Adviser
Steven Bibyk
Joanne Degroat
Approved by
AdviserGraduate Program in
Electrical and ComputerEngineering
c�
Copyright by
Jayanta Mukherjee
2006
ABSTRACT
We present a general circuit-based model of LC oscillator phase noise applicable to
both white noise and 1/f noise. Using Kurokawa theory, differential equations governing
the relationship between amplitude and phase noise at the tank are derived and solved.
Closed form equations are obtained for the IEEE oscillator phase noise for both white and
1/f noise. These solutions introduce new parameters which take into account the correlation
between the amplitude noise and phase noise and link them to the oscillator circuit oper-
ating point. These relations are then used to obtain the final expression for Voltage noise
power density across the output oscillator terminals assuming the noise can be modeled by
stationary Gaussian processes. For white noise, general conditions under which the phase
noise relaxes to closed-form Lorentzian spectra are derived for two practical limiting cases.
Further, the buffer noise in oscillators is examined. The forward contribution of the buffer
to the white noise floor for large offset is expressed in terms of the buffer noise parameters.
The backward contribution of the buffer to the �������� oscillator noise is also quantified. To
model flicker noise, the Kurokawa theory is extended by modeling each 1/f noise pertur-
bation in the oscillator as a small-signal dc perturbation of the oscillator operating point.
A trap level model of flicker noise is used for the analysis. A rigorous asymptotic analysis
is reported to demonstrate how the expressions obtained for a single trap can be applied
for a continuum of trap. Again the derivations take into account the correlation existing
between amplitude and phase noise at the LC tank. Conditions under which the resulting
ii
flicker noise relaxes to an �������� phase noise distribution are derived. The proposed model
is then applied to a practical differential oscillator. A novel method of analysis splitting
the noise contribution of the various transistors into modes is introduced to calculate the
Kurokawa noise parameters. The modes that contribute the most to white noise and flicker
noise are identified. Further the tail noise contribution is analyzed and shown to be mostly
up-converted noise. The combined white and flicker noise model exhibits the presence of a
number of corner frequencies whose values depend upon the relative strengths of the vari-
ous noise components. The proposed model is compared with a popular harmonic balance
simulator and a reasonable agreement is obtained in the respective range of validity of the
simulator and theory. The analytical theory presented which relies on measurable circuit
parameters provide valuable insight for oscillator performance optimization as is discussed
in the paper. Next we design a practical TSMC 0.18 um differential LC oscillator. The
model is then applied on a practical TSMC 0.18 um LC oscillator. In this case also the
model matches very well with simulation results. We further analyze the impact of vary-
ing the supply voltage of this oscillator using loadline techniques. Finally we compare
simulation results with experimental results and see that there is a very close match be-
tween simulation and experimental results. From this we conclude that our model matches
experimental results as well.
iii
This is dedicated to my parents, Sharbani, Ananta, Basanta and Uma
iv
ACKNOWLEDGMENTS
First of all I would like to thank my adviser Dr Patrick Roblin in helping me complete
the research. He provided some crucial insights especially at the end for completing the
research. Many of the ideas in this research were his. Thanks are also due to him for giving
me the Texas Instruments Fellowship. I would also like to thank Dr Bibyk and Dr Degroat
for their support and encouragement.
Prof Wan Rone Liou of National Taiwan Ocean University helped us a lot in arranging
the funding and the tools for the TSMC 0.18 um chip fabrication. Without his help the
experimental part of this work would not have been possible. I should definitely thank Dr
Young Gi Kim for his design and layout insights for the oscillator. Sook Joo Doo of the
Non Linear RF Lab helped me a lot in the testing of the chip.
I also owe a lot of gratitude to my wife. In a sense the whole research came alive only
after she came to the United States. She help me in being more comfortable in the US and
supporting me during turbulent times. The contribution of this provident and extremely
intelligent lady cannot be overemphasized.
I also would like to thank my parents who very patiently guided me through the whole
process, never being too demanding yet encouraging me to complete the work.
v
VITA
May 25, 1977 � � ��� � ��� � � ��� � ����� ��� ��� ��� ����� � ��� � � ��� Born - India
Mukherjee, Jayanta; Roblin, Patrick “Novel modeling methodology of LC differential os-cillator noise”. Microwave and Optical Technology Letters , April 2006, pp. 805 - 808
Mukherjee J, Roblin, P, Bibyk S “General non linear perturbation model for flicker noisein LC oscillator”. Circuits and Systems, 2005. 48th Midwest Symposium on , August 7-10,2005 Page(s):595 - 59
Mukherjee J, Roblin, P, Bibyk S “Accurate circuit based non linear perturbation model ofphase noise in LC oscillators”. Circuits and Systems, 2005. 48th Midwest Symposium on ,August 7-10, 2005 Page(s):754 - 757
Instructional Publications
Mukherjee, J.; Parry, J.; Dai, W.; Roblin, P.; Bibyk, S.; Lee, J. “RFIC loadpull simulationsimplementing best practice RF and mixed-signal design using an integrated agilent and
vi
cadence EDA tool”. Microelectronic Systems Education, 2003. Proceedings. 2003 IEEEInternational Conference on, 1-2 June 2003 Page(s):76 - 77
F. Derivation of 5,.�A�B� for white noise case . . . . . . . . . . . . . . . . . . . . . 101
G. Derivation of the Phase Noise with Buffer Noise . . . . . . . . . . . . . . . . . 102
H. Fourier Transform of �C��D(� with exponential of exponential . . . . . . . . . . . 106
I. Derivation of the Fourier Transform of the More General Exponential of anEponential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
I.1 Obtaining the Fourier Transform of the form �C��D(�:E�FHG��A�I�4DJ� . . . . . . . 110I.2 Fourier Transform of �K�0D(� for a more general case . . . . . . . . . . . . 111
7.1 Response in dBV of the tank voltage at �H�MLN�O� to a single tone excitationinjected in the circuit with 1KHz offset ��� from the various harmonics PQ�R�. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
7.2 Response in dBV of the tank voltage at �H�MLN�O� to a single tone excitationinjected in the circuit with 1 MHz offset ��� from the various harmonicsP��S� . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
8.1 Table compares the pre and post simulation results. . . . . . . . . . . . . . 73
9.1 Table compares the simulation and experimental results. Note the closeagreement between simulation and experiment thereby validating the pro-posed model since the model results are close to the simulation results. . . . 79
4.2 Impact of buffer white noise on the phase noise of a differential oscillator . 37
5.1 The figure shows that +-@ has a linear relationship with time . . . . . . . . . 45
5.2 Comparison of analytic expression for AM+PM and AM with exact nu-merical solution and simulator results for Mode E. . . . . . . . . . . . . . 46
5.3 Comparison of analytic expression for AM+PM and AM with exact nu-merical solution and simulator results for Mode A. . . . . . . . . . . . . . 47
5.4 Comparison of analytic expression for AM+PM and AM with exact nu-merical solution and simulator results for Mode C. . . . . . . . . . . . . . 48
5.5 Comparison of analytic expression for AM+PM and AM with exact numer-ical solution and simulator results for all 3 modes summed. In other wordsthis is a comparison of the analytic model with simulation results for flickernoise as a whole. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.6 Figure shows how the value of �X��� changes with Z[� . Also it computes thederivative of �\��� for various ]RZ^� . . . . . . . . . . . . . . . . . . . . . . . 49
7.2 Various modes of a differential oscillator . . . . . . . . . . . . . . . . . . . 59
xiii
7.3 Impact of the tank voltage at ���_L`��� of a 1 nA perturbation current atvarious offsets �O� for Mode A through E . . . . . . . . . . . . . . . . . . 61
7.4 Comparison of 1/f modes A, C and E . . . . . . . . . . . . . . . . . . . . . 63
7.5 Comparison of model and simulator results for the AM and AM+PM 1/fnoise for all modes summed (A+C+E) . . . . . . . . . . . . . . . . . . . . 64
7.6 Comparison of the Voltage noise densities when white and flicker noisesummed (circle, square, star) with flicker noise (plain line) and white noise(dashed line, dashed dotted line, dotted line) for three different white noiselevels. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
8.1 A TSMC 0.18 um differential oscillator with buffer stage . . . . . . . . . . 68
8.5 The figure shows the cross section of a multi metal layer process. Thoughonly a single metal layer has been shown, there may be more than one. Theinter metal layer is filled with 5?Z>a � dielectric. . . . . . . . . . . . . . . . . 71
9.2 Comparison of the white noise simulation results and the proposed model.Note that due to the low value of b c , the plot for the two different values ofd 1 overlap. The match between simulation results and the model is within3 dB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
xiv
9.3 Comparison of the combined white and flicker noise simulation results andthe proposed model. The model matches very well in the white noise re-gion. In the flicker noise region, the deviation with simulation results isaround 5 dB. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
9.7 Die Micrograph of the oscillator having an area of 1.4 o 1.4 dpd � . . . . . . . 81
9.8 Experimental Results: Shows the Phase Noise Spectrum with a span of1 MHz. The phase noise computed at 100 KHz offset from the centerfrequency is -87dBc/Hz . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
9.9 Experimental Results: Shows the Phase Noise Spectrum with a span of 10MHz. The phase noise computed at 1 MHz offset from the center frequencyis -108dBc/Hz . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
10.1 Comparison of Phase Noise obtained at various values of q . We can seethat at q g`lSr d the lowest Phase Noise is obtained for �s�tgi��u�uHvxwzy . . 88
Here �K�0/!� is the phase in the oscillation and may vary with time. �s�0/!� is the amplitude
of oscillation. In most systems, the amplitude of oscillation can be sufficiently limited,
as a result of which the variation in �s�0/!� can be sufficiently suppressed. Nevertheless the
presence of both amplitude and phase fluctuations cause sidebands in the output voltageeJ�>�����0/!� spectrum of the oscillator. Fig 2.2 shows the difference in the output spectrum of
an ideal oscillator with that of a real oscillator.
9
outV
Activedevice
Tank Circuit
Figure 2.1: A simple oscillator
2.2 Phase Noise
Phase noise is a quantity similar to the Signal to Noise Ratio (SNR). However instead
of measuring the ratio of the noise to the signal power in the same frequency range, Phase
Here F is an empirical parameter, k the Boltzmann constant, T the absolute temperature, ¯Å×the average power dissipated in the resistive part of the tank, �.� is the oscillation frequency,
13
ÚÜÛ is the loaded Q factor of the tank with all loadings accounted for. As we can see from
the equation it is empirical with a goal of accounting for all possible regions of the phase
noise spectrum without giving any logic for how such regions might have arrived.
2.3.2 Linear Time Invariant Model of Phase Noise
The Time Invariant model of phase noise assumes a linear transformation of device
noise to Phase Noise. The device noise is assumed to be completely shaped by the oscillator
tank. A simple derivation of the above equation can be made as follows. The impedance of
L
f∆
in
)0
G2
L
Active Device
C −Gm(V
Figure 2.5: Equivalent one-port model of an LC oscillator
the tank part can be approximated as,â �A���KL§���B�?ã �"äÛ � ���Læå¡lçÚäÛ Á\èèjé (2.7)
14
The active device modeled as Æ�"$µ-�8e(�)� cancels the tank conductance "�Û during steady
state. Hence during steady state, the net impedance of the combination is,â �8���B�?g {S�>���!������L����B�Z # |¤���xL����B� gÊÆ�å �"äÛ � ���lçÚäÛ:�s� (2.8)
Now both the tank as well as the active device will contribute noise. The equivalent noise
source arising from "$Û can be represented as # 9¾ÁQ³ = ÒRÔÖÕ "äÛ . If we then combine the equiv-
alent noise contributed by the active device, then the total equivalent current source can be
represented as # 9¾Á�³ = ÒRÓ$ÔÖÕ "äÛ , where F is fitting factor determined a posteriori . Hence the
total phase noise can be calculated as ,
�ÅÍS�s�_Î<gÊ��u¡�2��FH��Ø { �| � #&·8²� Ý gÊ� u:�2��F��ëêìí 1� � á â �����B� á ��� # 9¾ÁQ³1� �Ìe �� î�ïðgn��u:�2�¥FH�ÐÏÑ3ÒRÓ$ÔÖÕ¯K× �(Ø �K�l�Úä�s� Ý � Þß (2.9)
Here # 9¾ÁQ³ g ÒRÓ$ÔÖÕ . This phenomenon can be described by Fig 2.6. We can see that the
noise spectra gets shaped by the tank impedance thereby giving the characteristic 1³ 9 shape
of the output voltage spectrum. We also note that the expression does not contain any 1³)´region. To account for both of these,the Eqn 2.9 is modified and the � 7à ´ term included to
arrive at the final form given by Eqn 2.6.
In [8] [9] a simple explanation of how this 1³)´ noise arises. Since the active device is
basically a nonlinear device, its output can be represented as
eJ�>���g�ñ?14e # |�L§ñ � e �# | L§ñ e # |Then let us consider an input to the device given by
e # |¤��/!�Igm�Y�(E�F�G:�K�!/L§��|CE�FHG¤��|j/15
Z(
Output
ω
ω
Noise Power Spectral Density
0ω
0ω
ω+∆ω)
Spectrum
Figure 2.6: Noise shaping in the oscillator
The output will exhibit the following important terms
Hence if the frequency component �C| arises due to the flicker noise in the active device, it
will be up converted to ���C�Iôõ��|H� frequencies at the output of the oscillator. This in turn
will then be shaped by the tank impedance in the same way the white noise was shaped as
described above giving a 1³ ´ region in the phase noise spectrum. The theory though simple
assumes a linear behavior of oscillator which is not true.
2.3.3 Linear Time Variant Model of Phase noise
Hajimiri and Lee [2], [5], [4] proposed a linear time variant model of oscillator . They
introduce a new quantity known as the Impulse Sensitivity Function(ISF). Let us consider
the lossless circuit of Fig 2.7 where we apply an impulse ö¡�0/!� . Suppose initially the
oscillation in the lossless circuit was ideal. Now on applying the impulse, the oscillation
changes as shown in the same figure 2.8. We also note that applying the impulse at different
instants produces different effects. If the impulse is applied while the oscillation amplitude
is at its peak, then the oscillation returns to its original state except undergoing a small
amplitude change. On the other hand if the impulse is applied while the amplitude is at
its lowest point, the the oscillation undergoes a noticeable phase change. So the impulse
response of this LC circuit may be written as÷ @¡�0/� !D(�Ig T.�����4DJ�ø µ�°�¶ r��0/�ÆND(� (2.10)
where r��0/!� is the unit step function and T���ù\� is the impulse sensitivity function. Dividing
by ø µ�°�¶ , the maximum charge displacement across the capacitor, makes the function T���ù\�independent of signal amplitude. T.�0ù\� is a dimensionless, frequency and amplitude inde-
pendent function periodic in lSú . As its name suggests, it encodes information about the
sensitivity of the oscillator to an impulse injected at phase ���2D . In the LC oscillator exam-
ple, T.��ù\� has its maximum value near the zero crossings of the oscillation, and a zero value
17
i(t)
δ(
i(t)
t
C
L
τ)t−
Figure 2.7: LC oscillator excited by a current pulse
at maxima of the oscillation waveform. Once the ISF has been determined(by whatever
means), we may compute the excess phase through use of the superposition integral. This
computation is valid here since superposition is linked to linearity, not time invariance
���0/!�?g`¸�ûü û ÷ @¤��/� 4D(�WZ)��D(�!]HDýg �ø µK°W¶ ¸ �ü û T.�����4DJ��Z)��D(��]RD (2.11)
This computation can be visualized with the help of the equivalent block diagram shown
in Fig 2.9. To cast this equation in a more practically useful form, we note that the ISF is
periodic and therefore expressible as a Fourier series
T.�����4DJ��gÿþ �l L û�|��\1 þ |?E�FHG��0P����2D$L���|R� (2.12)
18
∆
τ
τ
τ
t
t
V
∆V
Vout
maxV
maxV
Vout
t
Figure 2.8: Time variant perturbation of the oscillator signal by an impulse showing howperturbation at different instants produces different effects on the oscillation
where the coefficients þ | are real and � | is the phase of the nth harmonic of the ISF. We will
ignore � | in all that follows because we will be assuming that noise components are uncor-
related , so their relative phase is irrelevant. The value of this decomposition is that, like
many functions associated with physical phenomena, the series typically converges rapidly,
so that it is often well approximated by just the first few terms of the series. Substituting the
Fourier expansion in to Eqn 2.11, and exchanging summation and integration, one obtains
�K��/!�.g �ø µ�°�¶ � þ �l ¸ �ü û Z2�0D(��]HDsL û�|��\1 þ |.¸ �ü û Z2�0DJ�:E�FHGj�0P����4DJ�!]HDÖ® (2.13)
The corresponding sequence of mathematical operations is shown graphically in Fig 2.10.
Now, suppose, we inject a sinusoidal current whose frequency is near an integer multipled of the oscillation frequency, so that
Z)��/!�.g���µýE�FHGj� � d ���KL����B�W/[� (2.14)
19
ωIntegrationIdeal V(t)ϕ( t)
t)0
0 ϕ( t)cos[ t+ ]ψ(t)q max
i(t)
Γ(ω
Phase Modulation
Figure 2.9: The equivalent block diagram of the process
where ���� �K� . Substituting Eqn 2.13 into Eqn 2.14 and noting that there is a negligible
net contribution to the integral by terms other than when P�g d , one obtains the following
We note from the above equation that on lowering the dc value of T ie, T?UWV , the corner
frequency is reduced and hence the the power spectral density curve has a more thinner cen-
tral lobe. From equation 2.22 we get a design idea ie, if a differential topology is adopted
Tank
gnd!
vdd!
Tail Capacitor improves symmetryand helps the ISF to be identical oneither arms
The symmetrical nature of the circuithelps in making very smalldcΓ
Figure 2.12: A Differential oscillator has low value of TKU�Vthen due to symmetry, the dc components of the two arms of the differential topology (as
shown in Fig 2.12) cancel each other significantly thereby resulting in a lower value ofTQUWV . Further, we also note that in order to maintain this symmetry, the noise currents from
the tail current source has to be kept very low, since the noise currents coming from the
tail capacitor are random and hence, can change T�¶ which depends on the time instant at
23
which the noise current is applied. To reduce this noise current from the tail capacitor, a
tail capacitance can be used as shown in Fig 2.12.
Limitations of the LTV Theory
The LTV theory though providing a good tool for explaining the phase noise spectrum
in oscillators especially the �j��� region, suffers from a number of shortcomings
1. It assumes oscillators are inherently linear time variant, but does not give a concrete
reason for this
2. It is based on the parameter T.�0ù\� which is very difficult to determine
3. It does not give an in depth view of oscillator design
2.3.4 Perturbation Models
Kaertner [24] and Demir [11] [1] obtain differential equations for describing the fre-
quency and amplitude response of oscillators through perturbation techniques. Due to the
involved nature of their work, we shall not go into the details. However, the salient features
od both their works is as follows.
1. They obtain differential equations describing the amplitude and phase deviations of
the oscillator in terms of Taylor series expansions, assuming the underlying device
noise can be completely described stochastically.
2. The stochastic differential equations so obtained are solved to obtain the final expres-
sion of phase noise.
3. Since flicker noise is difficult to characterize in time domain, they obtain approxi-
mates series solutions.
24
4. The models depend on complex parameters and have no circuit focus.
5. They require special tools and efficient algorithms to evaluate the model parameters.
So in conclusion, as we stated in the previous chapter, in this work we intend to solve
some of the problems with the previous models. As we shall see in the succeeding chapters,
though our model follows the general philosophy of Kaertner [24] and Demir [11] and [1],
it does have a strong circuit focus. The parameters involved in the model will be easy to
measure. Further, our approximation that flicker noise is equivalent to variation of device
bias point will lead to a closed form solution for Phase Noise arising due to flicker noise.
2.4 Simulation Tools
The results obtained from our model will be compared with simulation results. The
primary simulation tool we use is Agilent ADS. This is a powerful tool and performs a
number of simulations like Harmonic Balance(HB), S-Parameter Analysis(SP), DC and
Transient. While good for simulation, it has limited layout drawing capabilities.
On the other hand a tool like Cadence Design System allows us to draw the layout of
a circuit. The TSMC oscillator described in Chapter 8 was laid out using the Virtuoso
tool in Cadence. It has the ability to perform a number of standard analyses like HB, SP,
DC. However the simulations are time intensive and inconvenient. So in summary, for the
initial simulations of the TSMC oscillator, ADS was used and then Cadence was used for
the layout.
25
CHAPTER 3
NON LINEAR PERTURBATION ANALYSIS OF A OSCILLATOR
i = i + i INLNINYLY
Figure 3.1: An Admittance model of an oscillator
Fig 3.1 shows a negative resistance oscillator model. The basic oscillator has been
divided into a linear frequency sensitive part (having admittance �ÛX���B� ) and a non linear
or device part (which is both frequency and amplitude sensitive and has admittance given
by ���������� ��B� ). In a conventional LC oscillator the linear part usually represents the tank.
For all derivations henceforth we shall consider the tank to be a parallel +-��� circuit. In
steady state, in the absence of noise and other perturbing signals, the operating point ( �s� ,��� ) is given by, ��Û\�������QL������������� !�����?g¹u26
Z^�Y�0/!� represents an equivalent noise source which arises due to noise sources in both the
linear and the non linear parts of the oscillator circuit and which will be extracted in the
example shown later. It can be shown (Appendix A) that the phase and amplitude deviations
( � and öj� ) obey the following Langevin equations (3.1) & (3.2), when the oscillator is
linearized about its operating point voltage amplitude ( �$� ) and frequency( �K� )[16].
��� �MöS�tL ]ÈöS�]H/ á � '� �����3� á � gmZ^�?1!* '� �������QLtZ^� � " '� ������� (3.1)��� � ñKöj�§L á � '� �A����� á � ]Ö�]R/ ®�g�Z>�I1!" '� �������KÆÇZ^� � * '� ������� (3.2)
with, ñægm" '� ��������" '��� ���Y�� ��K�)�QL�* '� �A�K�)�!* '��� ���Y�� ��K�)��xg¹* '� �������!" '��� �0�Y� ��K�)�KÆt" '� �A�K�)�!* '��� ���Y�� !�����and where Z^�I1 and Z^� � are given by,
Z>�I1 g lÕ ¸ �� ü � Z>�ýE�F�G����I/�L ���0/!�!�!]H/ (3.3)Z>� � g lÕ ¸ �� ü � Z>�ýG4�¥�Q�A�I/QL��K�0/!�!�!]H/ (3.4)
Note that the processes öj� and � are similar to Ornstein-Uhlenbech processes [14] except
that they are correlated. The term ñ accounts for the correlation between � and öS� (i.e.,
if ñ�gÙu there is no correlation). The terms � '� �A���3� and � '��� �0�Y�� !���)� are given by (more
general treatment than in [16] which uses a frequency independent ����������3� ),� '� �A�����óg " '� �A�����QLhåÈ* '� �A�K�)�.g"! ��Û! �$##### èjé L%! �(���! � ##### = é } èjé (3.5)
+Y;8=?�0DJ� and +<@¡��D(� represent the autocorrelation functions of the amplitude deviation öj�and the phase deviation � respectively.
The first term in Eqn 3.7 which is proportional to � �� is the PM noise. The second term
involving +<;8=?��D(� is the AM noise term. Equation 3.7 is derived neglecting a third contri-
bution arising from the correlation between phase and amplitude noises. This is justified as
the PM noise will be verified below to be dominant over AM noise: �$�0/�/ þ so that it can
be assumed to be dominant as well over this third contribution.
We shall first consider white noise sources. Flicker noise analysis will be discussed
later. As shown in Appendix C,
If 5 #&% �>�X� á g á 1:á � �5 #&%J7 �>���?g¹5 #&%:9 �8���?g`l á 1:á �As is shown in Appendices D and E the expressions for +Ü;>=C�0DJ� & +Y@¡��D(� are derived from
the Langevin equations (3.1) & (3.2) to be given by:
+�;>=?��D(�Ig á21:á ���� � �3+¡ ��WÆ.4 á D á � (3.8)
+Y@¡�0DJ� g Æ á21:á �� �� á � '� ������� á � ؤ�BL ñ �� � Ý á D á Æ ñ � á 1:á �� � � �,+¡ $��Æ04 á D á � (3.9)
where we introduce the frequency 4 :
4�g �Y�5�á � '� ������� á �28
+<,.�0DJ�Ig �l � � �� L á 1:á ��Y� � �,+¡ s�[Æ04 á D á �>®�E�FHG������4D(���,+¡ � Æ á21:á �� �� á � '� ������� á � ؤ��L ñ �� � Ý á D á�651 ü87:9 ;�9 Æõ��<ç®(3.10)
According to the Wiener Khintchin theorem, the power spectral density of the voltage
across the tank can now be obtained by taking the inverse Fourier Transform of +s,.�0DJ� :5�,I���B�?g>= ü 1 ÍS+<,.�0DJ��Î
Note that 5Q, is directly observable on a spectrum analyzer. 5M@¡���B� which is derived in
Appendix E is the IEEE definition of phase noise [13] requires a phase detector. For ñægmu(no correlation) a closed form solution under the form of a Lorentzian (defined below)
is obtained for 5Q, . No exact analytic solution is available for the Fourier Transform of+<,.�0DJ� when we account for the correlation between the amplitude and phase ( ñ non zero).
However an approximate analytical Lorentzian solution can be obtained for two practical
limit conditions:
5�,ç} ·>·�? ���B� g � �� � d 1d � 1 L¹���kÆh���3� � ®@ ACB DEGFIHKJ «2L�¬ L þ � d 1QLM4� d 1QLM4:� � L¹�A�NÆæ�K�3� � ®@ ACB DN FOHPJ «2L�¬ (3.11)
with þ g 9 ² 9 9= é c and where we define d 1 as
d 1 g QR S d �41 for ���T/U/�4 or for ñ g¹u and all �d �41ÅoÊÂ0��L 6 b c < � Ä for 4V/�/ d �41 and ���XWT4with d �41Ig á21:á �� �� á � '� ������� á �
The asymptotic result for ����/T4 is derived in Appendix H using the stationary phase
approximation. If 4 is very large the second approximation involving the correlation factor�8ñK����� � is the relevant choice.
29
For intermediate value of 4 the Fourier transform of +ä,.�0D(� can readily be calculated
numerically. This is illustrated in Figure 3.2 for 4¤�¤�8lSúM� = 3.5 MHz. As is shown in Figure
3.2 5�,ç} ·>·�? ���B� is seen to relax for ���Y/�/Y4 to the limiting Lorentzian with ñ g u . The
voltage noise spectrum is no longer strictly a Lorentzian and an inflexion point is introduced
in the PM voltage noise density at the frequency 4:�¤�>lSú� .
104 105 106 107 108−120
−110
−100
−90
−80
−70
−60
Offset Frequency (Hz)
Sv N
oise
Spe
ctru
m (d
BV
)
η2/(2π)
Exact numerical solution: PM noiseLorenzian using m
01(1+α2/β2)
Lorenzian using m01
Figure 3.2: Comparison of PM voltage noise spectrum with various Lorentzian approxi-mations.
The expressions of 5Q, obtained for ñÊg u (no correlation) are consistent with other
published works [11][24]. A Lorentzian spectrum ensures that the total power of the os-
cillator remains finite. A �j������� spectrum of noise spectral density for all frequencies on
the other hand implies infinite oscillator power. Note that for large values of ���§g��zÆ ���compared to d 1 , the spectrum can be approximated as5�,ç} ·>·Z? ���B�Cg �Y�� d 1��� � L þ � d 1QLM4:���� �
30
Integrating the phase noise over frequency it can be shown that,
¸ ûü û 5�,I���B��]W�pg ¸ û� 5�,ç} ·8·�? �A�B�!]W�pg � �� L þlwhich is the same as the power of a noiseless oscillator if AM white noise is neglected
( þ g�u ).
104 105 106 107 108−120
−110
−100
−90
−80
−70
−60
−50
Offset Frequency (Hz)
Sv N
oise
Spe
ctru
m (d
BV
)
AM
AM+PM
PM
Simulator
PM noiseAM noisePM+AM noiseSimulator results
Figure 3.3: Comparison of AM and PM white noise spectrum in a differential oscillator
Figure 3.3 compares the AM, PM and (AM+PM) white noise component of 5K,ç} ·>·Z? for
a differential oscillator (introduced in Section V). On a logarithmic graph the approximate
PM/AM Lorentzian spectra have corner frequencies given by d 14�¤�8lSúM�\[ u¡�^] MHz and� d 1QLM4:�4�:�>lSúM�_[ Ò � l MHz respectively.
Since � �� (PM noise) exceeds þ (AM noise) (by 20dB in Fig 3.3) d 1!�¤�>ljúM� is the corner
frequency of the total (AM+PM) white noise spectrum.
31
The inflexion point at 4 observed in the PM noise spectrum is not easily observable in
the total (AM+PM) noise spectrum due to the AM noise contribution. Agreement of the
theory with the circuit simulator is verified to be within 0.6 dB at high frequency offsets
( ���T/�/ d 1 ) for the circuit considered.
Both d 1 and d 1�L`4 are proportional to ��� á � '� ������� á � . For a parallel tank,á � '� �A����� á [ la�
and is proportional to the tank Ú . This shows that at large offsets 5K,.���B� is proportional to�j��ÚÜ� , thus agreeing with Leeson's model. However the equation derived above presents
the voltage noise in terms of easily measurable parameters � '� �A����� and � '��� ����� !����� in
conventional harmonic balance simulation of oscillators.
The presence of the additional terms ��ñC����� in d 1 provides greater accuracy in the
expression for the Voltage noise density. The ratio ñK��� has to be as low as possible for
reducing phase noise.ñ� g � "�b� *cb� �d�Ö�¿"�b ��� *cb��� �� " b� * b� �Qot�¿" b ��� * b��� � g á �Ub� á¡á ��b��� á E�FHG �á � b� áÖá � b��� á G!���e� g�E�F�¢*�where � is the angle between the complex vector � b� and � b��� . It results that when � is 90 �the noise correlation is minimized: ñC���hgÊu . When � is 0 � the noise the noise correlation
is maximized: ñK���xgIf . This well known result was first inferred by Kurokawa [16] from
the inspection of 5Q@¡���B� . One of the contribution of the present work for white noise is to
introduce the correlation factor ñC���tg�E�F�¢*� and to quantify the impact of the correlation
on the overall phase noise spectra 5M,.���B� of the oscillator. In the circuit considered the cor-
relation term is verified in Fig. 3.2 to bring a shift on the order of ô 3.8 dB for frequencies
below 4:�¤�>ljúM� .Kurokawa [16] provided a graphical interpretation of the correlation factor. In the limit
where � b� g � bÛ , (active devices contributing minimally to the tank Q factor), � is the an-
gle at the oscillator operating point ���I�� 2�Y��� between the locus of the device admittance
32
line ���������� ��K�)� as a function of � and the locus of the circuit admittance line �Q��A�B� as
function of � . Therefore in the limit where � b� gÿ� bÛ holds the oscillator noise is respec-
tively maximized or minimized when the circuit and device admittance lines are tangent or
perpendicular to one another.
For an ideal high Q parallel tank ( "gb� gm"�b Û g¹u ) we have, bc gm* '��� �0�Y� !�����!��" '�!� ����� !����� .In other words to make the ratio small the variation of device conductance and susceptance
with respect to amplitude should be very high and low respectively. This is the case with an
active device operating below its �ç� since such a device has a high conductance sensitivity
with respect to amplitude. In Chapter 7 we will show how the white noise theory can be
applied to a differential oscillator.
3.1 Obtaining the circuit parameters hjilkCm�nao and hqpsrtkvuwnPx�m�naoFig 3.4 shows the circuit used for obtaining the values of �X�K�A����� and �����Y�0�Y� ��K�)� .
A signal �xG!����A�K�4/!� is injected at terminal 3 of the element. The problem associated with
measuring ���!���0�� !�B� is that the value of �\��� changes with amplitude and frequency. Hence
the value of �\��� at any particular configuration has to be measured by measuring the input
voltage and current at that particular configuration. The circuit element in Fig 3.4 does just
that. The S parameter matrix of this element is given as follows,
5zy*{¤�Kf�=}|�~���¡=:Û g Ï���Ñ u � u uu u � uu u u u� u u uÞ����ß 5��:�d��~�~.y(~K~*�*{}~��_�¤�C~�×äg Ï���Ñ u � u u� u u uu u u u� u u u
Þ����ß(3.12)
So a signal injected at a particular frequency � and amplitude � will pass from terminal 3
to 2, then pass through the tank, then pass through the device part. Due to the non linearities
of the tank, the signal gets distorted. On the return path only the fundamental signal passes
33
from terminal 1 to 4 while other signals are sent to 2. This way by collecting the signal
(usually current) at terminal 4 the input and output characteristics of the non liner device
can be obtained and so �\�!� can be easily obtained. By changing � about the operating
point, � '��� ������ !����� can be obtained. Similarly by changing � about the operating point,! �����B� ! � can be computed, which in turn leads us to � '� ������� .
Z0 A sin(w t)0Circuit Element
34
2 1
S
Device Part
Tank Part
Z0
+
−
Figure 3.4: Schematic of the circuit element used for measuring �X�������3� and �(�!������� !�����
34
CHAPTER 4
BUFFER NOISE
Most oscillator circuits have some kind of 2-port buffers to stabilize the load impedance.
This can range from a simple pad attenuator to a differential output buffer stage like in the
circuit to be considered in Section V. The impact of a 2-port buffer on the oscillator is
usually not discussed in details in the literature. Also most simulators neglect the presence
of a noise floor introduced by the buffer in the output and shows the noise decreasing
infinitely with increasing frequency offset.
−
+
Buffer
+
+n
inV VLYYIN out
L
Load
n,Rv+
in
v
inZ
−
LR
−
−
iN,o
iN,buf
Oscillator
Figure 4.1: Mechanism of Buffer Noise Addition
Consider the noisy 2-port equivalent circuit for the buffer circuit shown in Fig 4.1
which features the usual input-referred noise current ZW| and noise voltage {ç| .35
Using the results given in Appendix G the Norton equivalent noise current Z2�} ? � ³ in-
jected by the buffer network in the oscillator circuit is given by
Z^�M} ? � ³ g�Z8|�L {j|â # | L��(�4{j|ç} ~ �with â # |Üg �� # | g¹yç1[1KÆ yç1 � y � 1+YÛäL�y �[�and �(��g �y � 1 � yç1[1â # | Æ�� �It results that the noise current power density Z � �M} ? � ³ injected by the buffer in the oscillator
is: Z � �M} ? � ³ g Z �� L á � # |�L§��� á � { �| L á �(� á � { �|ç} ~ �where �*� is the correlation admittance Z[Väg ���\{S| . In these expressions, Z>� is the uncor-
related component and Z[V the correlated noise component of the total buffer noise currentZ8|$g�Z8� L§Z>V . The load noise { �|ç} ~ � and the buffer noises can themselves be expressed as:
{ �| g ÒÈÔ Õ +�|: { �|ç} ~ � g ÒÈÔ Õ +YÛ and Z �� g ÒRÔ Õ "-�È�Thus an additional component Z � �M} ? � ³ gets added to Z � �M} � g á 1:á � in the total Z � � and the
definition of d 1 changes to d 1?g � á21:á ��L á Z^�} ? � ³ á ���� �� á � '� �A����� á � Ø:��L ñM�� � ÝThe equivalent voltage noise source {H| appearing at the input contributes directly to the
input-referred noise floor. The total noise at the output of the buffer including the buffer
noise is then:
5�,��Z�v�!���B�?g�� �� 6 5�,ç} ·>·�? ���B��L { ��} ? � ³ < g á � � á � Ø 5�,ç} ·>·Z? ���B�XL { �| L { �|ç} ~ � ####yç1[1y � 1 ####
� Ý (4.1)
36
104 105 106 107 108 109−120
−110
−100
−90
−80
−70
−60
−50
Offset Frequency (Hz)
Sv N
oise
Spe
ctru
m (d
BV
)
AM+PM + Buffer
AM+PM
BufferAM+PM+Buffer
SimulatorPM + AM + Buffer noisePM+AM noiseSimulator resultsBuffer noise
Figure 4.2: Impact of buffer white noise on the phase noise of a differential oscillator
and where � � is the voltage gain provided by the buffer:
� � g y � 1!+YÛâ 1[1��8+YÛäL§y �[� �KÆÇyç1 � y � 1The derivation of {¡��M} ? � ³ is shown in Appendix refappd. { �| (i.e, +�| ) is usually the leading
term contributing to the noise floor. The impact of the noise floor is illustrated in Fig. 4.2
for a buffer noise floor of -130 dBV with two different white noise strength of1 �j�W�j�ýg� �^�$o ��u ü 1Z� A � /Hz (unusually strong white noise) and
1 �j�>lH�Igml¡� � oæ��u ü � A � /Hz (normal
weaker white noise). The corner frequency between white noise and noise floor is given
by: � � yzg �l&���� � �� d 1{ ��M} ? � ³�³Note that � � for white noise of usual strength
1 �j�>l�� is on the order 4¤�:�>lSúM� and therefored 1 should indeed be used instead of d �41 in the frequency range ( ��Wÿ� � ) where white
37
noise dominates. Note also that of the 6 dB offset between model (plain line) and simulator
(+) results for1 �j�8lH� , 4 dB are due to the new ����L§E�F�¢��*��� correlation factor.
38
CHAPTER 5
EXTENDING THE MODEL TO FLICKER NOISE
In order to extend the Kurokawa analysis to 1/f noise we can study the effect of a
variation of the DC current Z[� g%���ÅL�öSZ>� in any oscillator component, upon the active
admittance �\��� . Using a linearization scheme similar to the one we employed in the white
noise case we get,
"ä�������Y�� !��� 2Z^���óg "Ü�!�������� !��� ������QL ! "Ü�!�! � öS�§L ! "ä���! � ö �zL ! "Ü�!�! Z^� öSZ>�*Y�������Y�� !��� 2Z^���óg *Y�������Y�� ��K�� ������QL�! *Y���! � öj�§L�! *Y���! � ö �zL�! *Y���! Z^� öjZ^�With the additional derivative term at the end, the master equations become,á � '� ������� á ��Y� ]ÈöS�]R/ L��MöS� g *�öSZ>�<�0/!� (5.1)á � '� ������� á � ]Ö�]H/ L�ñ�öS� g Æ��<öjZ^�Y�0/!� (5.2)
öjZ^���0/!� will be defined by its power density 5; #&% }1�� ³ ���B�?g¹5Q� á � á as is developed in the next
section.
5.1 Solving the differential equations and obtaining the expression forthe voltage noise density
We use the autocorrelation function of the charge trapping model of the flicker noise
to find the final noise voltage density as shown in [6]. The goal is to obtain a stationary
model of flicker noise. We first assume the autocorrelation function is WSS and equal to
(Ornstein-Uhlenbeck assumption [14]),+ #&% �0DJ�.g Ô 1 üP��9 ;�9 (5.3)
Taking the Fourier Transform of + #&% �0D(� gives the Noise density as,5 # % �A�B�Cg lG� Ô� � LT� � � (5.4)
This autocorrelation corresponds to a random telegraph noise which has a Lorentzian dis-
tribution. A superposition of many of these processes with time constants DS�6±[°��¡���¤�Cgi�j�G� gD�×��,+¡ M�����¤� which are spatially varying with position � in the oxide (MOS) or wide-bandgap
(HFET), will result in a noise process with a �j�H� distribution:5 #&% }1�� ³ g ¸ U º:»[¼� + #6% ��D(�!]�� gnÆ ¸ � 7� é + #&% �0DJ� ]P��g l���  ¢ � � ü 1 � ��J� � ÆN¢ � � ü 1 � ���1 � Ä�[ Ô ú��� g 5� for ��1�W��XWX�(�where �(�BgÊ�j�jD ×g[�f , ��1?gi���SD��6±^°��Ö��]Rµ�°W¶ç��[nu and 5Ng Ô úM��� . The superposition is valid
since we assume the different traps behave independently.
We take the Fourier transform of Eqns 5.1 & 5.2 and express the RHS of each equation
in a simplified form. á � '� ���B� á ���� å�� öj���A�B��LM�ÿöj���A�B� g * öjZ^�����B� (5.5)
40
á � '� ���B� á � å��B�����B�XL§ñÙöj���A�B� g Æ�� öSZ>�<�A�B� (5.6)
As shown in Appendix I using the method of stationary phase the Inverse Fourier Trans-
form of +Y@j}10�6±[°Z�¡�0DJ� for large frequency offset �s�Çg�� Æë�C� is given by (using a single side
band representation)
5�, 7 �°¯ »�±v² ³´³´µ �A�B�42
g �Y�� vl � P � Ô ���� � � Ô � L���� � � L P Ô ��� � � Ô L���� � ®ÜL v 'l¤� Ô � Æ Ô �  �Ô L���� � Æ �Ô � L���� � Äg �Y�� vl���� � Ô 1QL����.�� Ô L���� � ����� � L§��� � � L v 'l �� Ô L§��� � ����� � L���� � � (5.15)
This expression gives the noise spectral density at large frequency offset for a single trap
(i.e for a particular value of � ). Note that Appendix I can be applied to derive the PM
component because the terms ø 1 , ø � , and ø which can be identified for a single trap in+<,ç}10�6±[°Z� verify ø 1QL ø � L ø g�u thus satisfying the required property � À ø À gmu .To obtain the tank voltage 1/f noise spectrum we need first to consider all the traps at
once when calculating the phase noise +ä@+Y@�}1�� ³ �0D(�IgnÆ ¸ � 7� é +Y@�}10�6±[°��Ö��D(� ]K��
before we calculate +-,ç}1�� ³ and 5�,ç}1�� ³ . It turns out that a closed form expression for+Y@�}1�� ³ �0D(� can be obtained using among other thing the exponential integral function ¶$Z)��ù\� .The expression is given below,
L Ô 1KÆ Ô l Ô » �,+¡ M��Æ Ã Ô DJ���¡Æ.¸?�[�0DQ���X1�Æ Ã Ô �QL�¸?�^�0D���(�.Æ Ã Ô � �Æz�3+: M� Ã Ô DJ���ÖÆ.¸?�[��D���X1QL Ã Ô �QLM¸?�[�0DQ���J�CL Ã Ô � �½¼LëÔ 1KÆ Ô l Ô �3+: äÆ Ã Ô D · �¥� #####
�X1KÆ ª Ô �J�.Æ�ª Ô ##### ÆÇ�¥� #####�X1QL ª Ô �J�KLXª Ô ##### º(® (5.16)
where v ' ' g¹vë�G� . This analytic solution + @j}1�� ³ ��D(� is however quite useful for verification
purpose as we can then obtain the exact voltage noise density by calculating numerically
its inverse Fourier transform. Similarly an exact expression for +$;8=�}1�� ³ �0DJ� can be obtained
43
as,
+ ;8=�}1�� ³ �0DJ�óg Æ v ' ' 'Ò� �,+¡ ���Æ ª Ô á D á �Ô
· �¥� #####�X1�Æ ª Ô �J�IÆTª Ô ##### ÆÇ��� #####
�X1L ª Ô �J�CLXª Ô ##### ºÆ �Ô �» �3+¡ M��ÆäÃ Ô á D á � ¸I�>�4����1KÆ Ã Ô � á D á �KƸ?�[���(�IÆ Ã Ô � á D á �Æz�,+¡ M�WÃ Ô á D á �¾¸I�^�4����1�L Ã Ô � á D á �KÆ�¸?�>���J�CL Ã Ô � á D á �Æt��¸?�^���X1 á D á �KÆ�¸?�^���(� á D á �4�)Î��(5.17)
where v ' ' ' g¹v ' �G� . An analytic expression for 5 ,ç}1�� ³ valid for large frequency offsets can
also be obtained using the method of stationary phase. This is due to the fact that the key
assumption made in Appendix I namely � À ø À gmu still holds when summing (integrating)
over all the traps. This follows from the dependence of +ä@ on time. It can be shown that a
linear relationship between +Ü@ and time will ensure that the superposition of the 5�, values
of a number of traps is the same as that obtained by superposing the noise contributions of
the traps first and then obtaining 5M, . As seen from Figure 5.1 the relationship is indeed
linear. It results by applying the results of Appendix I that for large enough frequency
offsets the voltage noise density of flicker noise is simply obtained by averaging over all
values of � the expression obtained for the Voltage Noise Density of a single trap noise:
5�,ç}1�� ³ } ·>·�? ���B�g �� ¸ � 7� é 5�, 7 �°¯ »�± �A�B� ��Æ�]P�\��g · �Y�� v ' 'l��È��� � Ô 1QL§���.�Ô L§��� � L v ' ' 'l¾� �Ô L���� � ºxo ¸ � 7� é �� � L���� � �WÆY]P���g · � �� v ' 'l��È��� � Ô 1QL§��� �Ô L§��� � L v ' ' 'l¾� �Ô L���� � ºxo ��s� o � ¢ � � ü 1 Ø �J���� Ý Æk¢ � � ü 1 Ø �X1��� Ý ®(5.18)
44
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x 10−5
1.8292
1.8292
1.8293
1.8293
1.8294
1.8294
1.8295
1.8295
1.8296
1.8296
1.8296x 105
time (s)
Rφ(e
xact
)
Figure 5.1: The figure shows that +-@ has a linear relationship with time
where, v ' ' gmvë�G� and v ' ' ' g¹v ' �G� . In the limit of �(�Åg�f and �X1Ig¹u this reduces to
5 ,ç}1�� ³ } ·8·�? �A�B� g � �� 5 ¦ �l � Ô 1�L���� ���� ��4 � L���� � � ®-L � �� 5?* �l á � '� �A����� á ¥ ����Y�Z4 � L§��� � � (5.19)
Figure 5.2 shows the analytic expressions obtained for AM+PM 1/f noise (plain line)
and for AM 1/f noise (dotted line). The test circuit used is the differential oscillator to
be discussed in Section V. The voltage noise spectra reported are for 1/f noise originating
in the tail transistor (later referred at Mode E). Also shown in Figure 5.2 are the results
obtained for AM+PM 1/f noise (dashed line) for the uncorrelated case ( ñ g u ). Clearly
correlation can play an important role for mode E as it leads to a 30 dB decrease in noise
in the �j�ç�O� region. Figure 5.3 and 5.4 show the same results for Mode A and Mode C.
Figure
At low offset (here below 1 Hz) the stationary phase approximation fails as indicated
by the exact numerical results (circles). A simple estimate of the ceiling voltage noise
Figure 5.2: Comparison of analytic expression for AM+PM and AM with exact numericalsolution and simulator results for Mode E.
density 5 ,ç}1�� ³ �0E��������¥�¤�È� and associated corner frequency ��� 1�� ³ ��E������¥���¤�H� can be obtained by
enforcing power conservation and assuming a ������� spectrum up to the corner frequency:
5�,ç}1�� ³ ��E����¥�����:�R�óg u:�¿] o �Y�ª Ó with Ó g � �� 5 ¦ � Ô 1l¾4 ����R1�� ³ ��E����¥�����:�R�óg �lSú Ø Ó5�,ç}1�� ³ ��E������¥���¤�H�jÝ 7´Agreement with circuit simulation is only of 7 dB for mode E. Better agreement will
be obtained for other dominant 1/f modes (A and C) in Section V when correlation brings
a weaker correction.
5.2 Obtaining the values of À(ÁÃÂ�Äpsr kvuwnPx�m�nao and ÅÆÁÃÂ�Äpsr kvuwnPx�m�naoThe test setup described in chapter 3 is used to perturb the oscillator about its oscillat-
ing point. Only a single variable is changed at a time (1 dimensional perturbation). The
Figure 5.3: Comparison of analytic expression for AM+PM and AM with exact numericalsolution and simulator results for Mode A.
corresponding values of �X��� are noted. A parabolic least square fit of the derivative is then
made to obtain accurately the derivatives. Figure 5.6 shows how the derivative converges
using this method.
47
10−2 100 102 104 106 108−250
−200
−150
−100
−50
0
50
Offset Frequency (Hz)
Vol
tage
Noi
se D
ensi
ty (d
BV
/Hz)
Mode C
Correlated PM+AM 1/f noiseCorrelated: AM 1/f noiseUncorrelated PM+AM 1/f noiseADS simulation 1/f noiseAtan model
Figure 5.4: Comparison of analytic expression for AM+PM and AM with exact numericalsolution and simulator results for Mode C.
101 102 103 104 105 106 107 108−250
−200
−150
−100
−50
0
50
Offset Frequency (Hz)
Vol
tage
Noi
se D
ensi
ty (d
BV
/Hz)
All 3 1/f Modes (A+C+E)
AM 1/f noise
Simulator
Correlated PM+AM 1/f noiseCorrelated: AM 1/f noiseUncorrelated PM+AM 1/f noiseExact numerical resultsSimulator results for 1/f noise
Figure 5.5: Comparison of analytic expression for AM+PM and AM with exact numericalsolution and simulator results for all 3 modes summed. In other words this is a comparisonof the analytic model with simulation results for flicker noise as a whole.
48
−1 0 1
x 10−7
−1.08
−1.06
−1.04
−1.02
−1
−0.98
−0.96
−0.94
−0.92x 10−6
iN
real
(Yin
)
−1 0 1
x 10−7
6.3885
6.389
6.3895
6.39
6.3905
6.391
6.3915
6.392
6.3925
6.393
6.3935x 10−4
iN
imag
(Yin
)
0 1 2
x 10−7
0.6827
0.6827
0.6827
0.6827
0.6827
0.6827
0.6827
0.6827
0.6827
diN
real
(dY
in/d
iN)
0 1 2
x 10−7
−1.8847
−1.8847
−1.8847
−1.8847
−1.8847
−1.8847
−1.8847
−1.8847
−1.8847
−1.8847
diN
imag
(dY
in/d
iN)
Figure 5.6: Figure shows how the value of �X�!� changes with Z[� . Also it computes thederivative of �\��� for various ]RZ[�
49
CHAPTER 6
APPLYING THE MODEL ON A VAN DER POL OSCILLATOR
B
A
L’
L
I(v)CLRN
i (t)
YL
YIN
Figure 6.1: Van Der Pol oscillator
The Van Der Pol oscillator is shown in Fig 6.1. The non linear current �\��{¡� is given as
�\��{¡�?gÊÆ {+ �Ö�_ÆÈÇ 1É { � �The circuit has been divided into 2 parts (a nonlinear part and a linear part ) by the line
LL'. ��Û\�A�B� represents the admittance of the linear part , while �Q���Y�0����B� represents the
50
admittance of the non linear part. Now let us calculate the values �Û��A�K�)� and �������0�Y�� !������(Û��A�B� g Z��j��L �å����g Z��j� Â0�_Æ �� � ��� ÄÊ �(ÛX�A�����óg Z����3� � ��Æ �� �� ��� ® g¹u
Hence at the frequency of oscillation, the admittance of the linear part is 0. The
Kurokawa coefficient � 'Û �A����� is obtained from the equation above as,
� 'Û �������IgmZ^la�From which we get " ' Û �������Igmu and * 'Û �������Ig¹la� .
To compute the admittance �X�������� ��B� , we proceed as follows. Let us inject a voltageengm�xG4�¥�Q�A�K�!/!� into the nonlinear part. Then
�\��{¡�?gÊÆ �+ ؤ�_ÆËÇ 1�� �É G4�¥� � �A�K�!/!� Ý G!����A���4/!�We can assume that since the system is time invariant , the harmonics of �\��{¡� will be
periodic with a fundamental frequency �I� . Taking a Fourier expansion of �\��{¡� we get,
�\��{¡�?g>���CL��S1¡G4�¥������4/!�QL�� � G4�¥��>lj���4/!�QL��¥� � �where, ��� , �j1 , � � ... are the coefficients of the harmonics terms of �\��{¡� . Then �����Y�0�� !�B� is
given by, �(�!���0�� !�B�?g �j1�The coefficient of the fundamental component �H1 is given by
g Æ lç�+  �l Æ Ç 1� � � ÄÊ �(�����0�� !�B� g Æ �+ � �_Æ Ç 1��Y�Ò ® (6.1)
We can see that �\���Y�0�� !�B� is independent of � . Now the Kurokawa coefficient � '��� �0�Y� ��K�)�can be obtained from Eqns 6.1 as,
� '�!� ���$ !�B� g Ç 1��l�+Ê � '�!� ����� !�����óg Ç 1��Y�l�+From the result of �<� obtained later we will see that � '��� ������ !���)� can be expressed as
����������� !����� g ÆY�(Û��A�����Ê ����������� !����� g uÊ Æ �+ � �_Æ Ç 1[�Y�Ò ® g¹uÊ �Y� g lª Ç 1Now in our model
ñ g " ' Û ��������" '��� ����� !�����QL�* 'Û �A�K�)�!* '��� ���Y�� ��K�)�g u� g * 'Û �������!" '��� �0�Y� ��K�)�KÆt" 'Û ��������* '��� �0�Y�� !���)�g la� ª Ç 1+á 1:á g Ì ÒRÔÖÕ+52
g ��u üPÍ PX�-� ª wzy(� � G4GvÎdÏO�:Ð(��Y� g lª Ç 1á � 'Û ������� á g la�d 1óg á21:á �� �� á � 'Û �A�K�)� á � ؤ��L ñ �� � Ýg ÔÈÕlÇ 1Ò +�� �þ g á21:á ��Y� �g ÔÖÕ�d 1L�4 g �����á � 'Û ������� á �g �+��5�,��A�B�óg �Y��l � d 1d � 1 L���� � ®<L þl � d 1QL d �� d 1QL d � � � L���� � ®g lÇ 1 Ø ÒÈÔÖÕ.Ç 1�+��Ü�� ÔÖÕ.Ç 14� � L¹� Ò +�� � � � ��� � Ý L ÔÖÕla� Ø Ò +��Ü�j� ÔÖÕ.Ç 1�L Ò ���� ÔÖÕ.Ç 1�L Ò ��� � L¹� Ò +�� � � � ��� � Ý@ AÑB D× (6.3)
We note that if we consider the deviation of amplitude of oscillation to be negligible ,
then the S portion of Eqn 6.3 is 0 and for large values of ��� ,
5�,.���B� ã �l ÔÖÕ+�� � ��� � (6.4)
We assume the values of � , � , + and Ç 1 are , ��g��u:Ò(w , � g � Ò ��� Ó , +�g ÉGÓGÓaÔandÇ 1�g É
We have implemented this circuit in both ADS and Spectre . Fig 6.2 shows the
comparison of the results obtained in various models. Fig 6.3 shows the circuit used to
obtain the the phase noise in Cadence while Fig 6.4 shows the one used in ADS. We see
that our model matches well with the simulation results obtained using Spectre(less than
0.035 dB) and ADS(pnfm)(less than 0.15 dB) at high frequencies. At low frequencies (less
than 100 Hz), the shape of the PSD in our model is a Lorentzian. By not approximating
53
10−10 10−5 100 105 1010 1015 1020−350
−300
−250
−200
−150
−100
−50
0
50
100
freq (Hz)
Noi
se V
olta
ge P
SD
(dB
)
Proposed ModelADS(pnmx)spectreRF(Cadence)
Figure 6.2: Comparison of the our model with ADS and Spectre
to a �j�H�\� model we ensure that the total power in the oscillator signal is finite. The noise
PSD integrated over all frequencies i.e.the total power comes out to be= 9é� ( �Y� is the steady
state amplitude of oscillation) which is the same as the power of a noiseless oscillator.
54
Figure 6.3: Cadence implementation circuit
Figure 6.4: ADS implementation circuit
55
CHAPTER 7
COMPARISON OF THE PROPOSED MODEL WITH APRACTICAL DIFFERENTIAL OSCILLATOR
Let us consider the differential oscillator shown in Fig 7.1. The tank is selected to form
the linear part of the oscillator while the remaining part of the circuit forms the non-linear
part.
iN2
O/P
A
Tank
Buffer StageStage
Bi
ip12
in1
Oscillator
2p2i
2n2
2
Figure 7.1: A differential oscillator with buffer stage
7.1 Mode Analysis of the oscillator
The Kurokawa analysis is typically used in harmonic balance simulators for calculating
the oscillator operating point ( �� , ��� ) using TQ��A�K�)�_oxTQ�������Y�� ��K�)�Bg�� . Consequently the
56
various non-linear device and linear tank admittance are readily available for applying the
circuit based noise model presented above. However a method for calculating the equiva-
lent noise source appearing across the tank is needed to apply the noise theory developed
in the previous sections for the white noise analysis. The methodology introduced for this
purpose will also facilitate the 1/f analysis by reducing the number of independent modes
(noise sources) considered from 5 to 3.
Fig 7.1 shows the schematic of a simple differential oscillator having four 'core' tran-
sistors. Z �� 1 and Z �� � represent the noise produced by PMOS transistors while Z �|S1 and Z �| �represent the noise produced by the NMOS transistors. Note that for white noise the gate
and drain noise currents (see for example [22])
Z ��,�>|ç} U g ÒÈÔ Õ½Õ |��(×aÖçµ and Z ��:�>|S} Ë g ÒÈÔ Õ ö:|��(×aÖ Ëof opposite NMOS and PMOS transistors (n1 & n2 or p1 & n2) appear in parallel due to
the circuit topology such that we have for example Z �� 1 g Z �� 1[} U L Z �� � } Ë . For the rest of the
analysis we defineá Z´� 1 á � g á Z×� � á � g Ç and
á Z8|ç1 á � g á Z8| � á � g&Ø . For 1/f noise we similarly
have Ç g��>5}� 1[} UCL�5}� � } Ë �4� á � á and Ø_gÊ�>5\|ç1[} UKL 5\| � } Ë �!� á � á .The first step for the mode analysis is to split the instantaneous noise currents of the
four transistors into uncorrelated unit mode currents.
Let É be the matrix which converts the transistor noise currents into these uncorrelated
mode currents. i.e. Z8µ��^U ² gmÉ Z8�6±[°W| · where Z8µK�>U ² and Z8�6±^°�| · are given by,
Z8µ��^U ² g Ï���Ñ Z8µÚÙZ8µ_ÛZ8µ_ÜZ8µ_ÝÞ ���ß and Z8�6±^°�| · g Ï���Ñ Z×� 7Z×� 9Z�| 7Z�| 9
Þ ���ß (7.1)
57
Since Z8µ��^U ² consists of uncorrelated mode currents having unit power density, we have,
á Z�µK�>U ² á � gmZ�µK�^U ² ��Z �µK�^U ² g¹É Z8�6±[°W| · ��Z ��6±[°W| · É � g Ï���Ñ � u u uu � u uu u � uu u u �Þ ���ß (7.2)
It can be shown that the matrix É , which leads to the above relation, is given by
where, Ç |$g Ç Ì lÇ L l Ø jØ2|$gIØ Ì lÇ L l Ø þ g �à l¤� Ç L�Ø3�Taking the inverse of É , we obtain,
É ü 1 g Ï���ÑlÞ Þ ß ßÞ Æ Þ ß Æ ßÞ Þ Æ Õ Æ ÕÞ Æ Þ Æ Õ Õ
Þ����ßwhere, Þ , ß and Õ are given by,
Þ g Ì Ç Øl¤� Ç L�Ø3� ß g Çà l¤� Ç L�Ø3� Õ g Øà l:� Ç LMØ3�Now we can express Z8�6±[°W| · in terms of Z8µK�^U ² as,
Z8�6±[°W| · g¹É ü 1 Z8µ��^U ²We have now expressed Z>�6±[°W| · in terms of a certain number of uncorrelated unity strength
mode currents. Schematically these modes can be represented as shown in Fig 7.2.
As seen from Tables 7.1 & 7.2 when the noise currents are substituted by a single tone
perturbation, each mode contributes differently to the noise current injected across the tank
by either up-conversion or direct transfer. For modes A and C the largest contribution to the
current Z^� across the tank at �ç�çLO��� is coming from the up-conversion of the input currents
58
Tank
Mode A
2 µ imA
µ i
Tank
Tank Tank
.
Mode C
Mode B
Mode D
mA
2 µ
ν i mC2 ν imC
(ν−γ)
mCiγ
(ν−γ)imC
γ imC 2
imC
(ν−γ)
mD mD
i mD
γ imD
+
(γ−ν)
iν
imD
γ imC
ν imCiν
imA
2µ i mA
µ i
µ
mAi
imA
µ imA2 µ mA
2 µimB
µ imBµ imB
µ−2
mB
imB
µ i mB µ i
voutSiin
S
iγmD
0
0
0
0
−
Figure 7.2: Various modes of a differential oscillator
Z8µ�=d�Ñ� at ��� . For modes B and D the largest contributions to the current Z!� injected across
the tank at �ç�ÖLë��� is coming from the transfer of the input currents at Z[µ� �[f at �S�ÖLë��� . It
results that modes B and D are most significant for white noise and modes A and C are most
significant for ���H� noise. The tail current leakage of each of the individual mode noises is
verified to be usually insignificant in a differential oscillator (the tail current source acts as
an open for mode A, B, C and D).
59
The resultant current across the tank for modes B and D are l Þ and � ß Æ Õ � respectively.
The effect of individual modes can be accounted by adding their relative contributions to+Y@ and +<;8= respectively i.e for white noise,
+Y@Üg�+Y@ Û L�+Y@:Ý+�;>= g¹+Y;8= Û L§+Y;>=}Ý
The resulting voltage noise density 5M, obtained will be the same as the one we obtained
previously reported in Section 3 ifá21:á � is replaced now byá21:á � g á 1 á � L á 1 f á �
.
Input Noise offset Mode A Mode B Mode C Mode D Mode E��� -96 -316 -84 -299 -89�S� + ��� -700 -95 -700 -91 -700l��S� + ��� -700 -700 -700 -700 -700É �S� + ��� -700 -700 -700 -700 -700
Table 7.1: Response in dBV of the tank voltage at �H��Ln��� to a single tone excitationinjected in the circuit with 1KHz offset �O� from the various harmonics P��� .
If we consider the total ZW� flowing across the tank due to modes B and D, we get,
5 #&% ���B�?g Ò Þ � Lm� ß Æ Õ � � gàÇ LMØl g á 1:á �For white noise this results is the same as the noise current density across the tank given
in [3]. For 1/f noise we need to calculate "�#6%�'��� �0�Y� ��K�� ������ and *$#&%('�!� ���Y�� !��� ����)� for mode
A and C. One can select ö #&% to be l Þ and ß for mode A and C respectively, For mode
60
Input Noise offset Mode A Mode B Mode C Mode D Mode E��� -96 -322 -83 -302 -89�j� + �O� -700 -95 -700 -92 -700l��S� + ��� -700 -700 -700 -700 -700É �S� + ��� -700 -700 -700 -700 -700
Table 7.2: Response in dBV of the tank voltage at �H��Ln��� to a single tone excitationinjected in the circuit with 1 MHz offset �O� from the various harmonics P��� .
C if we calculate "ä#6%('��� �0�Y� ��K�� ������ and *$#&%�'��� ����� !��� ������ using ö #&% g ßd �Ùg � nA for
the two top current sources in Figure 7.2 then the two bottom current sources are given byÕ d �Ðg Õ � ß ö #&% g Ç �aØ�ö #&% g Ç ��Ø�o�� nA. Note that for 1/f noise Ç �aØ g`5¹���H5\| is frequency
independent.
100 101 102 103 104 105 106 107 108−140
−130
−120
−110
−100
−90
−80
Offset Frequency (Hz)
Tank
Vol
tage
at f 0+∆
f (d
BV
)
Mode A: Input noise at ∆f
Mode B: Input noise at f0+∆
f
Mode C: Input noise at ∆f
Mode D: Input noise at f0+∆
f
Mode E: Input noise at ∆f
Figure 7.3: Impact of the tank voltage at �R�ML��� of a 1 nA perturbation current at variousoffsets ��� for Mode A through E
61
Note that our model neglects the frequency dependence of the oscillator on the pertur-
bating current ö #6% . As can be verified in Figure 7.3, this is a reasonable assumption up to 1
MHz offset for the differential oscillator considered in this paper.
7.2 Impact of Tail Noise
For completeness we need to consider the impact of tail noise on the overall voltage
noise density. We refer to the tail noise analysis as Mode E. However unlike the previous
modes, the noise current of the tail current is directly the mode current. Tables 7.1 and 7.2
show the effect of a single frequency tone with 1KHz and 1MHz offset relative to dc, ���and lj��� respectively. The results are similar to those obtained for cases A and C. In other
words the up-converted noise dominates over the transfered noise.
7.3 Obtaining the circuit parameters hjilkCm�nao and hqpsrtkvuwnPx�m�naoThe operating point �8��� 2���3�`[ �8l¡� Ò ]G]\áUâ0ã� ���¥���cä á � for the oscillator was obtained
using conventional harmonic balance simulation. An accuracy of 0.1 Hz and 8 Þ V was
achieved. The convergence accuracy of the voltage and current was set to �-o ��u ü 1Zå A and
V respectively in the harmonic balance simulation. 8 harmonics were used in all oscillator
simulations.
By changing � about the operating point, � '��� ����� !����� can be obtained. Similarly by
changing � about the operating point, ! �����B� ! � can be computed, which in turn leads us to� '� ������� . A quadratic convergence of the derivatives calculated was observed for shrinking
derivative intervals. A least square fit was further used to remove any residual numerical
noise. The maximum error bound (.e.g.,á � � '�!� ����� !�����!�:� '��� ����� !����� á ) in the derivatives
calculated was of lsoN��u üPæ % in relative magnitude.
62
100 102 104 106 108−250
−200
−150
−100
−50
0
50
Offset Frequency (Hz)
Vol
tage
Noi
se D
ensi
ty (d
BV
/Hz)
Modes A, C and E
mode A (model)mode C (model)mode E (model)mode A (simulator)mode C (simulator)mode E (simulator)
Figure 7.4: Comparison of 1/f modes A, C and E
7.4 Comparing Simulation results with proposed model for CombinedWhite and 1/f Noise
In this section we consider both white and flicker noise together. As we have already
seen when studying white noise and flicker noise to properly evaluate the combined effect
of multiple independent noise processes on the oscillator output voltage spectral density
we must first sum their respective phase and amplitude autocorrelations
+Y@j} �&�8�!�0D(� g +�@j}1�� ³ ��D(�L�+Y@�} çdè # � ² ��D(�+�;>=�} �&�8�!�0D(� g +�;>=�}1�� ³ ��D(�QL�+�;>=�} ç}è # � ² ��D(�before calculating the voltage spectral density 5�,ç} �&�>� . The commonly used approach which
consists of summing the voltage spectral density of white noise and flicker noise
Correlated PM+AM 1/f noiseCorrelated: AM 1/f noiseUncorrelated PM+AM 1/f noiseExact numerical resultsSimulator results for 1/f noise
Figure 7.5: Comparison of model and simulator results for the AM and AM+PM 1/f noisefor all modes summed (A+C+E)
is an approximation. As we shall demonstrate below, the validity of this approximation will
depend on the relative strengths of the �j�H� noise and white noise processes.
Consider the two corner frequencies ���aç}è # � ² �0E��������¥�¤�R�-g d 14�¤�>ljúM� and ��� 1�� ³ ��E������¥���¤�H�at which the voltage spectral density reaches its ceiling value when respectively considering
white noise or flicker noise separately.
When the corner frequency ����ç}è # � ² �0E��������¥�¤�R� of white noise is larger than the corner
frequency ��� 1�� ³ ��E����¥�����:�R� of flicker noise then no �j����� region will be present in the noise
spectrum since the ceiling dictated by power conservation has already been reached. This
is numerically verified to take place in Figure 7.6 in the presence of a strong white noise1 � � É ��g � �^� oÇ��u ü � A � /Hz when the total voltage density (circles) follows the white noise
Lorentzian spectra and not the flicker noise spectra. Clearly in this strong white noise case
64
100 101 102 103 104 105 106 107−180
−160
−140
−120
−100
−80
−60
−40
−20
0
Offset Frequency (Hz)
Vol
tage
Noi
se D
ensi
ty (d
BV
/Hz)
e2(3)
e2(2)
e2(1)
1/f
1/f noise (analytic)White noise e2(1) (analytic)White noise e2(2) (analytic)White noise e2(3) (analytic)1/f + White e2(1) (numerical)1/f + White e2(2) (numerical)1/f + White e2(3) (numerical)
Figure 7.6: Comparison of the Voltage noise densities when white and flicker noisesummed (circle, square, star) with flicker noise (plain line) and white noise (dashed line,dashed dotted line, dotted line) for three different white noise levels.
the usual summation of the voltage spectral density of white noise and flicker noise would
lead to incorrect noise spectra prediction.
On the other hand if the corner frequency ���açdè # � ² ��E����¥�����:�È� is smaller than the corner
frequency ���R1�� ³ �0E����¥���¥�¤�R� then the �j����� region will be observed in the noise spectrum.
This is the case in Figure 7.6 in the presence of the weaker white noise1 � ������g 1 � � É �4�¡��uaé
and1 �S�>l��4�¤� É om� u å � when the total (white + flicker) voltage densities (square and star)
follow first the ������� flicker noise spectra at low offset frequencies before switching at
high offset frequencies to the �j������� noise spectra. In such a case the corner frequency
between the �j�ç�O�\� and �j����� regions is given by (assuming mode A dominating):
�O� � g �lSú Ô 1[} =4 � 5X= ¦ �=l d 165
The usual summation of the voltage spectral density of white noise and flicker noise is then
an excellent approximation. In summary the simple rules described above for combining
noise processes permit us to predict the total noise spectrum from our analytic models
without resorting to the time consuming numerical analysis.
66
CHAPTER 8
DESIGN OF A TSMC 0.18UM LC OSCILLATOR
In this chapter we study the design of a LC oscillator using a TSMC 0.18 um, 6 metal
layer 2 poly process. In the next chapter we shall apply our model to this oscillator. This
way our model can be validated on a real differential LC oscillator. The general design of
the LC oscillator follows from that of the differential oscillator described in the previous
chapters using ideal BSIM3 transistors.
8.1 Basic Topology
The circuit diagram of the LC oscillator is shown in Figure 8.1. The circuit exhibits
a simple double differential architecture. Though differential oscillators are not good as
far as noise performance is considered, having this configuration gives us more flexibility
in varying the parameters of the circuit. An optimized circuit (lowest noise) is considered
for fabrication. As seen from Figure 8.1, the oscillator has a core consisting of 2 NMOS
and 2 PMOS transistors. The current bias of the oscillator is obtained through a current
mirror. The center frequency of the oscillator is found to be �R�ýg l¡�^� Ò�ê "äwzy . A buffer
stage consisting of a differential topology is connected to the oscillator core. The harmonic
balance and the phase noise simulation results are shown in Figures 8.2 and 8.3. Power
consumption without parasitics is found to be 91 mW. The layout of the circuit is shown
67
Figure 8.1: A TSMC 0.18 um differential oscillator with buffer stage
in Figure 8.4.
8.2 Post Layout Simulation
Usually post layout simulation involves extracting the RC parasitics of the oscillator
circuit and then re-simulating with the parasitics included. However, the problem with this
approach is that only RC parasitics are included. The inductive parasitics are not included.
A micro-strip approximation of the metal and dielectric layers can be used to find the
inductive as well as the RC parasitics as shown below.
A multi metal layer process like TSMC 0.18 um consists of 6 metal layers inter-spaced
by 5CZ^a � dielectric layers (Figure 8.5). If the thickness of the intermediate layers is ne-
glected (thickness ã u¡�^]G]çr d ), the top-layer metal (thickness ã u:� êGê r d ) can be approxi-
mated as the micro-strip line. The bottom layer metal is grounded and acts as the ground
68
Figure 8.2: Pre Layout Oscillator output harmonics
Figure 8.3: Pre Layout Oscillator Phase Noise
plane. Finally the dielectric layer between the top and bottom layer metal (ignoring the
intermediate metal layers) acts as the dielectric layer of a micro-strip line. The thickness of
this dielectric layer is around 6.5 um.
Once the above approximation has been made, the layout is carried out in such a way so
that most of the inter device connections are made through the top layer interconnects. The
other metal layers are kept to a minimum. The bottom layer is filled in extensively and is
grounded. These features can be seen in the layout Fig 8.4. The circuit for simulating the
69
Figure 8.4: Oscillator Layout
circuit post layout is shown in Figure 8.6. ADS is used for the micro-strip like simulation.
8.3 Design Flow
1. Identify the architecture (in this case double differential).
70
SiO Layers
Layer MetalIntermediate
(Microstrip Line)Top Layer Metal
(Ground Plane)Bottom Layer Metal
2
(Dielectric)
Active Region
Figure 8.5: The figure shows the cross section of a multi metal layer process. Though onlya single metal layer has been shown, there may be more than one. The inter metal layer isfilled with 5?Z>a � dielectric.
2. See that the oscillator oscillates. Optimize the design to obtain the lowest noise
(optimized design). Design done using ADS.
3. Obtain the layout of the circuit using Virtuoso editor in Cadence.
4. Translate the design back again to ADS, modeling the device interconnects as micro-
strip lines.
5. Re-simulate and see the performance.
6. Perform final layout optimization.
7. Obtain the gds files for the design.
71
Figure 8.6: Post Layout Circuit simulation setup
8.4 Post Layout Simulations Results
The Phase Noise of the optimized circuit is shown in Figure 8.7. The center frequency
shifts to 2.397 GHz. The post layout harmonics are shown in Figure 8.8. We then do a
supply voltage sweep to see the impact on the Fundamental tone as shown in Fig 8.9. The
net power dissipation of the circuit is 104 mW while total wafer area is 867 x 987 sq um.
72
Figure 8.7: Post Layout Phase Noise simulation of the oscillator
Pre Simulation Post SimulationCenter Frequency 2.549 GHz 2.397 GHz
Phase Noise -100 dBc/Hz at 100 KHz offset -88 dBc/Hz at 100 KHz offsetPower of Fundamental 4.976 dBm -8.262 dBm
Total Power Consumption 91 mW 104 mW
Table 8.1: Table compares the pre and post simulation results.
Figure 8.8: Post Layout HB simulation of the oscillator output
73
Figure 8.9: Impact of supply voltage sweep on the oscillator output
74
CHAPTER 9
APPLYING THE MODEL ON THE TSMC 0.18 UM OSCILLATOR
Figure 9.1 shows the circuit setup for obtaining the parameters necessary for applying
the model to the circuit described in chapter 8. As in the case of the ideal differential
Figure 9.1: A TSMC 0.18 um differential oscillator with buffer stage-setup for parameterextraction
75
oscillator, a least square method was used to obtain the values of the various derivatives
used in the model. The oscillation operating point was perturbed by injecting signals with
slightly different amplitude keeping the frequency same as the operating frequency and vice
versa. Similarly to obtain the derivatives wrt dc bias also, a perturbation of the oscillator
operating point was made by injecting dc currents across the transistors according to the
schemes for the modes A, C and E. Figure 9.2 shows the result of applying the model for
100 101 102 103 104 105 106 107−140
−120
−100
−80
−60
−40
−20
0
Frequency
Sv S
pect
rum
eta2 = 3766664376.109
Exact S
V (numerical−)
m1=|e|2(1+α2/β2)/A
02|dY
T/dω|2
m1=|e|2/A
02|dY
T/dω|2
Simulation Results
Figure 9.2: Comparison of the white noise simulation results and the proposed model. Notethat due to the low value of bc , the plot for the two different values of d 1 overlap. The matchbetween simulation results and the model is within 3 dB
76
the white noise case. The flicker noise sources of the transistors are turned off by setting
the NOIMOD flag in the BSIM3 model of the TSMC 0.18 um transistors to 1 and the KF
flag to 0.
The flicker noise parameters are obtained from the BSIM3 model of the transistors. It
follows that for the parameter NOIMOD=1, the flicker noise is given by,
Flicker Noise g v ³ � ° ³U ·�Å�>¶j� �² ³3³ � ² ³ (9.1)�Å�>¶ is obtained from the TSMC test run parameters available from the MOSIS webpage.� ² ³�³ is given by,
� ² ³3³ g �zL§öç� ² ³�³g �zLMë ����L �qì� Ûaí | L ��îq Û¾ç:| L ��î�ì� Ûaí | q Û�ç¡|(9.2)
The terms mentioned in the above equations can be obtained from the BSIM3 model spec-
ifications for TSMC 0.18 um process. Figure 9.3 compares the simulation results and the
proposed model for both white and flicker noise combined.
9.1 Loadline analysis of the TSMC oscillator
The only parameter that we can control in TSMC 0.18 um oscillator designed, is the
supply voltage. Reducing the supply voltage will reduce the Phase Noise of the oscillator
as shown in Figure 9.4. This should either result in reduction ofá21:á � or a reduction of ñC��� .
Reduction of ñC��� in turn means an increase of the angle � as described in Chapter 3. The
loadline plots for two different supply voltages are shown in Fig 9.5 and 9.6. As we can
see the angle is actually lower for supply voltage equal to �je . Hence, the main factor in
the lower phase noise must be the lower value ofá21:á � .
77
100 102 104 106 108−180
−160
−140
−120
−100
−80
−60
−40
−20
0
Offset Frequency (Hz)
Vol
tage
Noi
se D
ensi
ty (d
BV
/Hz)
1/f noise (analytic)White noise (analytic)1/f + White (analytic)Simulation Results
Figure 9.3: Comparison of the combined white and flicker noise simulation results and theproposed model. The model matches very well in the white noise region. In the flickernoise region, the deviation with simulation results is around 5 dB.
9.2 Experimental Results
Figure 9.7 shows the die micrograph of the fabricated oscillator. The chip area is
1.4 o 1.4 dpd � . Figure 9.8 shows the output of the oscillator at supply voltage 2V. The
Phase Noise can be computed at an offset of 100 KHz from the center frequency to be
-87 dBc/Hz. Figure 9.9 shows the same graph with a wider frequency span. The Phase
Noise can be computed from this graph to be -108 dBc/Hz at 1MHz offset. These results
very nicely match with the simulated results. Comparing the results obtained with high
end commercial oscillators like Maxim MAX2753 (which uses a SiGe Bipolar process),
which has a Phase Noise of -104 dBc/Hz at 100 KHz offset in the ISM band, we see that
the performance of the oscillator needs to be improved. Also the power consumption of
78
Figure 9.4: Phase Noise at e\fQfkgn�jeSimulation Experimental
Center Frequency 2.397 GHz 2.320 GHzPhase Noise -88 dBc/Hz at 100 KHz offset -87 dBc/Hz at 100 KHz offset
Power of Fundamental -8.262 dBm -3.898 dBmTotal Power Consumption 104 mW 80 mW
Table 9.1: Table compares the simulation and experimental results. Note the close agree-ment between simulation and experiment thereby validating the proposed model since themodel results are close to the simulation results.
80 mW is a bit too high compared to MAX2753 which has a power consumption of only
10 mW. The best Phase Noise performance obtained in TSMC 0.18 um at the ISM band is
Figure 9.7: Die Micrograph of the oscillator having an area of 1.4 o 1.4 dpd � .
Figure 9.8: Experimental Results: Shows the Phase Noise Spectrum with a span of 1 MHz.The phase noise computed at 100 KHz offset from the center frequency is -87dBc/Hz
81
Figure 9.9: Experimental Results: Shows the Phase Noise Spectrum with a span of 10MHz. The phase noise computed at 1 MHz offset from the center frequency is -108dBc/Hz
82
CHAPTER 10
CONCLUDING REMARKS
10.1 Summary
In this work we have attempted to find simple analytical solutions of the Phase Noise
spectrum of an LC oscillator. Analytical models are inherently approximate. The phase
noise spectrum of an oscillator is complex. There are minute noisy features, peaks and
troughs in the spectrum which are difficult to model. Nevertheless, a model by definition is
an idealized representation of a phenomenon. The purpose of a model is to aid the designer
in gaining insights and ease in understanding the phenomenon. It is similar to using a map.
It is not the real thing but helps us in visualizing the earth. The primary goal of our model
is simplicity. Our model incorporates a strong circuit focus. A circuit focus in the model
by itself makes the designer aware of the intricacies of the model and how it applies to his
design. We have tried to amalgamate the best of the world of models which on one hand
despite their mathematical beauty incorporate complex parameters which in turn require
their own algorithms to compute; and the world of pure circuit based heuristical models
which while simple are not entirely accurate and based on incorrect assumptions.
Our approach follows a non linear perturbation method. Perturbation methods allow us
to simplify a complex phenomenon by linearizing it about a reference point. In the case
of the oscillator, its operating point provides us the reference. By following the Kurokawa
83
theory of oscillations, we obtained the linearized equations. The other important aspect of
this thesis was the incorporation of correlations existing between the amplitude and phase
deviations of the oscillator from the reference. The proposed model takes into account for
the first time, correlations existing between the amplitude and phase voltage-noises at the
tank (embodied by the ñ and � factors).
As we had stated earlier, while white noise has been studied closely, flicker noise anal-
ysis left many holes. The key assumption for the flicker noise part was that the perturbation
in oscillation due to flicker noise is equivalent to the perturbation due to device bias current
variation. We succeeded in finding asymptotic expression for the voltage noise spectra.
These analytic expressions were verified to hold for a wide range of frequencies using nu-
merical analysis relying on the exact solution +ä@¡�0DJ� . Numerical(exact) solutions were
obtained which helped show the range of validity of the asymptotic analytic expressions.
The results were also compared with HP ADS simulation results and close agreement was
obtained in all cases as shown in the previous chapters. A mode theory of noise was devel-
oped to facilitate the calculation of the various Kurokawa noise parameters needed. This
mode theory also provides valuable insights in the various noise up-conversion/transfer
processes. Rules for combining various uncorrelated noise (e.g., white and flicker) were
presented and verified with numerical simulation
While simulation results are a convenient shortcut for designers, they have some limi-
tations. For one thing, simulators are black boxes where the designer has no idea of what is
going on. Further, the absence of the correct Lorentzian spectrum for phase noise showing
conservation of power tells us that simulator results may not always be correct. A simple
circuit focused model comes in handy then.
84
Another important contribution was the impact of buffer noise. Through a simple two
port model we demonstrated how the buffer noise contributes partly to the device and tank
noise and also produces the noise floor. This simple but important part of the phase noise
spectrum is often ignored. Finally the design experience of the TSMC 0.18 um oscillator
gave us a good example of the validity of this model on an actual oscillator. The applica-
bility of the proposed model on this oscillator proves its universality.
The analytic phase noise theory presented here is based on an extension of the Kurokawa
oscillator theory and has therefore the advantage of relying on circuit impedances and their
derivatives. These terms are easily obtained by harmonic balance simulations and could be
obtained experimentally from measurements of the device impedance.
10.2 Design Insights
The circuit based-approach used, provides deep insights in the noise processes. Specif-
ically for white noise the analytic model points towards the need to effectively reduce the
amplitude and phase noise correlation ( E�F�¢��kg ñK��� ). According to the analytic models
derived in this work, Leeson formula for large offset frequencies should be updated to be
with ¯K×�g������HlH��� �� "ÜÛ , the resonator Ú g �C�5ðÑÏ Í�ñ 'ò Îç�¤�>lGá�óH��[ ���3ô �Gá�ó and using for
example "$Ûzg&õ_�çÍ�ñló:Î for the resonator + buffer conductance. Although more accurate
formulas are given within this dissertation this simplified expression should be useful to
85
circuit designers to identify device and circuit parameters which are critical for phase noise
optimization.
The above equation can be used to find a lower bound on the Phase Noise wrt transistor
width i.e. the best Phase Noise performance that can be achieved by modifying the transis-
tor width. As we can see from the above equation, with increasing bias stability the phase
noise due to the flicker noise will reduce. Bias stability depends on transistor width, the
higher the width, the lower is the deviation in admittance due to change in bias current.
However, increase of width leads to higher white noise current. Hence there has to be a
compromise in the width that a transistor can have so as to achieve the lowest possible
phase noise. This is basically an optimization problem which we would like to quantify by
an example. The flicker noise current of a MOSFET is given by,
Z � ³ í # V À ² ± �>���óg vqn���Å�>¶j� Ö �µ (10.1)
When the dc bias current of the MOSFET's is constant as is the case in the differential
oscillator topology shown, Ö �µ gml Þ |a�Å�>¶Ö�8qÇ���.�ö��UFrom which the expression for flicker noise current reduces to,
Z � ³ í # V À ² ± �8���?g l Þ |�vt��U� � �This clearly shows that the flicker noise current is independent of width of the transistor.
On the other hand due to the dependence of the bias derivatives terms " # b %�!� and * # b %��� on q ,
the Phase Noise due to flicker noise current reduces by a factor q � (Ò ranging from 1 to 2)
with change in q for the same flicker noise current.
Further, the white noise current of a MOSFET is given by,Z �çdè # � ² �>�X� g �É ÔÖÕ ÖSµ§g �É ÔÈÕ Ì l Þ |a�Å�>¶q � ��U
86
(10.2)
which shows that the white noise current increases by a factor ª q for increasing q .
Hence the total Phase Noise can be written as,
¯$É g ÷@�AÑB�Dconstant (noise floor)
L É ª q@ AÑB Dwhite noise
L +q �@ ACB Dflicker noise
where, É g Ç çdè # � ² " �Û ����LtE�F�¢ � ���� � � ��� � ª qx|��>±^µ�°víand +`g Ç ³ í # V À ² ± �8" # b %��� ÆNE�Fç¢*��* # b %��� ���3q �| �>±^µ�°öí q �� � � ��� �The minimum value of PN is obtained by taking,]��8¯äÉz�]¡q gmuwhich gives, qzµ # |sg¨� l,Ò(+É � � 99 ±Ñø 7 � (10.3)
It can be shown that q µ # | is indeed a minima becauseU 9 ¢°ù � £Uöú 9 �8q g qzµ # |R� is always
greater than 0. We simplify our calculations by applying noise only across the PMOS
transistors and thereby assume that any change in "���� or *Y��� is due to change in the qof the PMOS transistors only. The Relation between the term ��" # b %��� ƧE�Fç¢*��* # b %��� ��� and qcan therefore be given by,
�8" # b %��� ÆNE�Fç¢*��* # b %��� � � g¹v q �|��>±^µ�°víq �The terms " # b %��� and * # b %��� only belong to those for Mode C since Þ g Õ g u for Øýg u .Hence plugging values in Equation 10.3 and taking Òxg � we get qNµ # |Og�l¡� ��]�r d . This
is indeed what is observed as shown in Fig 10.1 where the lowest phase noise occurs atq g¹lçr d .
87
Figure 10.1: Comparison of Phase Noise obtained at various values of q . We can see thatat q g`lSr d the lowest Phase Noise is obtained for �s�Çgi� uHuHvxwzy .10.3 Recommendations
Flicker noise involves memory effects. This arises because the strength of the flicker
noise depends on the history of trap distribution. Hence a model which incorporates these
memory effects will be highly desirable. We applied the white noise part of our theory on
the Van Der Pol oscillator. However the flicker noise part could not be applied because
of the absence of a connection between the dc bias of the device and the flicker noise
produced. In the BSIM3 transistors that we discussed this connection is inherent in the
model which is why we were able to get the values of � # b %��� . This is turn provides us two
avenues of further research. Is it possible to somehow produce this connection in the Van
Der Pol oscillator? Conversely, is it possible to obtain a model of the flicker noise part
without the dc bias connection?
Our derivation focused principally on PM phase noise which is the dominant term com-
pared to the AM noise which was also derived. There exist however the possibility for a
third type of combined AM-PM noise for strong correlated amplitude and phase noises.
This noise would be suppressed if the buffer acts an amplitude limiter. [2]. Note that the
88
expressions obtained for the IEEE phase noise definition +ä@¡���B� remain themselves unaf-
fected by this approximation.
Yet another work that remains to be done is to obtain a mechanism to translate the model
to a simulator. Most of the parameter extraction work described in this dissertation was
done manually. HP ADS provides for user defined models which can effectively automate
the entire process. Special AEL codes need to be written for this purpose.
Though the model has been developed only for LC oscillators, it may be possible to
generalize it for ring oscillators. The ring oscillator circuit may be divided into 2 parts as
we had done for the LC oscillator case. However since for the ring oscillator the admittance
of both the parts will have amplitude dependence we need to incorporate a derivative term
for the admittance of the ”tank” part with respect to amplitude and then obtain a new set of
equations. This way the model can be generalized for any oscillator.
89
APPENDIX A
DERIVATIONS OF MASTER EQUATIONS
Let us consider an admittance model of an oscillator as shown in Fig 3.1. The circuit
behaves as an open at resonance and a short for the harmonics. The current Z4��� (t) flowing
through the non linear part can be expressed as
Z>���Y��/!�Ig¹+ 1 � �s�0/!� 1 û ¢ è �ýü¤@ ¢ � £×£ �(�������Y�� 5î # �W� + large harmonics (A.1)
The voltage {\�0/!� is given by,
{\�0/!�Igm+ 1 �����/!� 1 û ¢ è �ýü¤@ ¢ � £ý£ � + small harmonics (A.2)
We note that at steady state when UW=U[� = UW@U[� = 0, � # g�� and Z^�tgmu"ÜÛX�������QL�"Ü���Y�0�Y�� !�����?gmu (A.7)*YÛX�������QL§*Y���Y�0�Y�� !�����?gmu (A.8)
Now multiplying both sides of Eqn A.6 first by E�FHGj�A�I/�L�K��/!�4� and then by G!����A�I/�L�K��/!�4�and then integrating each of those equations in time over one time period of oscillation we
get
Z>�I1óg lÕ ¸ �� ü � Z^�ýE�FHGj�A�I/QL��K�0/!�!�¤]H/g ����/!� ØH"äÛ����B��L§"ä�������� ��B��L�" '� ���B� ]¡�]H/ L * '�� ]R�]H/ Ý (A.9)Z>� � g lÕ ¸ �� ü � Z^�ýG!����A�I/QL��K��/!�4�¤]H/g Æ��s�0/!� ØR*YÛX���B�QL§*Y���Y�0�� !�B��L§* '� ���B� ]Ö�]R/ Æ " '� ���B�� ]È�]H/CÝ (A.10)
The equations above can be simplified by noting that
The above 2 equations can be simplified to,]ÈöS�]H/ á � '� �����3� á � L§�Y�5� öj� g Z^�?1!* '� �������QLtZ^� � " '� �A�K�)� (A.17)���YØ á � '� �A����� á � ]Ö�]R/ L�ñÿöS� Ý g Z^�?1!" '� �A�����KÆNZ^� � * '� ������� (A.18)
which are the required master equations.
92
APPENDIX B
DERIVATIONS OF ���`k�����x���¾oThe autocorrelation function of the oscillation for a time interval Dýg�/ � Æz/21 ( as shown
This derivation assumes öj� and � are uncorrelated. This assumption is often justified on
the basis that the buffer acts as a limiter which suppresses the öS� fluctuation at its output.
We will keep track of öj� in our work to monitor that its contribution remains negligible. If�Q1?g¹���0/!� and � � g¹���0/�LtD(� are jointly normal with zero mean and -1 = - � = - , then
DERIVATION OF � p Â�� k�� o AND � p  � k�� oWe have, Z^�?1 g lÕ ¸ �� Z^�OE�FHG ���I/�L �K��/!�!]H/Z^� � g lÕ ¸ �� Z^�OG4���Q���I/XL ���0/!�!]H/
The autocorrelation function of ZW�I1 is given by,
+ # %J7 �0/21� !/ � � g ÒÕ � ¸ � 7� 7 ü � ¸ � 9� 9 ü � ¶ý� Z>�Y��/214��Z^���0/ � �[�@ AÑB D9 ² 9 9 ; ¢ � 9 ü � 7 £ E�F�G����I/21�L��K�0/214�!�:E�FHG ���I/ � L ���0/ � �!�!]H/21�]H/ �From which, we get,
+ #&%(7 �0/213 4/ � �Ig Q!!!!R!!!!S u for
á / � ÆN/21 á / Õ¥ 9 ² 9 9� 9 ¸ �� E�FHG � ���I/XL ���0/!�!�!]H/@ ACB D¡ 9 g � 9 ² 9 9� for /)1KÆ Õ W§/ � W§/21The above expression gives us the value of + # %(7 ��/21� 4/ � � as a function of D . The plot of+ #&%J7 �0/21� !/ � � versus D is shown in Figure C.1. The total area under the graph is given by,
¸ ûü û + #6%(7 �0/21) 4/ � �!]HD g l á 1:á �This is the same area if + #6%(7 ��/213 4/ � �Ðg l á 1:á � ö:��D(� . Hence in the limit Õ " f , the
expression of + #&%(7 ��/213 4/ � � becomes
+ #&%(7 ��/21� 4/ � �?g`l á21:á � ö:��D(�95
i
−T0
T
22|e|
R (τ)N1
T
Figure C.1: + #6%(7 �0/213 !/ � � versus D
96
APPENDIX D
DERIVATION OF �$#�%�k�����x&�'��oFrom the Master equation (3.1) we get,
Z>�I1!* '� L§Z>� � " '� g��Y���Möj�§L ]ÈöS�]H/ á � '� á �Referring to Papoulis' book [20] we can obtain the following relation from the above,
assuming non correlation between the noise components Z!�I1 and Z^� �* '� +<�I13�0/213 4/ � �óg ��� �M+<�I1�;8=?�0/21) 4/ � �QL ]È+<�?1�;>=?��/21� 4/ � �]R/ � á � '� á �Ê l�* '� á21:á � ö:��/ � ÆN/21!�óg ��� �M+<�I1�;8=?�0/21) 4/ � �QL ]È+<�?1�;>=?��/21� 4/ � �]R/ � á � '� á � (D.1)
Solving for +-�I1�;8= gives us the following result,
+<�?1�;>=I��/21� 4/ � �óg l * '� á21:á �á � '� á � �3+¡ � Æ ��� �Dá � '� á � ® r��0D(�QL þ 1¡�3+: � Æ �Y� �QDá � '� á � ®where, Dýg�/ � Æk/21 . Using the boundary condition
+Y�I1�;>=C�0/213 Ælf��?g¹uwe get þ 1?gmu . Hence,
+Y�I1�;>=C�0/213 4/ � � g l * '� á21:á �á � '� á � �,+¡ � Æ ��� �QDá � '� á � ® r��0/ � ÆN/21!� (D.2)
97
Similarly we get the following expression for +ä� � ;8=+<� � ;8=I�0/21) 4/ � �óg l " '� á21:á �á � '� á � �3+¡ � Æ ��� �Dá � '� á � ®�r��0/ � ÆN/214� (D.3)
Again from the master equation 3.1 we get,
+<�?1�;>=I��/21� 4/ � ��* '� L§+<� � ;8=?��/213 4/ � �!" '� g �Y� �M+Y;>=K�0/213 !/ � �QL ]È+Y;8=?�0/213 !/ � �!]H/21 á � '� á �Ê ]È+Y;>=]R/21 L �Y� �M+Y;>=á � '� á � g l á21:á �á � '� á � �,+¡ � Æ �Y� ���0/ � ÆN/214�á � '� á � ®�rK��/ � ÆN/21!�(D.4)
From Eqn E.1 we get the amplitude variation spectral density as,
5X;8=?���B�?g l á � '� �A�K�)� á � á21:á �� �� � � LÇ� � á � '� �����3� á ¥using
á Z^�?1����B� á ��g á Z^� � ���B� á �Og�l á 1:á � . The derivation of this expression is shown in Ap-
pendix F
From the equations E.2, we get the following expression for theá � á �
5X@¡���B� g á21:á �� � á � '� ������� á ¥ � l á � '� �����3� á �� �� L§ñ � á öS� á � ®�g l á21:á �� � � �� � ���� á � '��� �0�Y�� !����� á ��LÇ�.� á � '� ������� á �� �� � � LÇ� � á � '� �A����� á ¥ ®g l á21:á �� �� á � '� ������� á � � � Ø ��L ñM�� � Ý Æ á � '� ������� á �� � á � '� ������� á ¥ L§� �� � � o l á21:á �2ñ��� �� � � (E.5)
99
For the uncorrelated öS� and � case, the above equation reduces to
5X@:�A�B�?g l á 1:á �� � � �� á � '� ������� á � (E.6)
By taking the inverse Fourier transform of the equations (E.5) & (E.6) we obtain for the
correlated case,
+Y@¡��D(�IgÊÆ á21:á �� �� á � '� �A�K�)� á � Ø ��L ñM�� � Ý á D á Æ ñM� á21:á �� � � 1 ü ( Ù é* + b¡ �.- é « * 9 9 ;�9 (E.7)
Here we have used the relation ñ � LU� � g á � '��� �0�Y�� !����� á � á � '� ������� á � . For the uncorrelated
case the equation reduces to,
+Y@¡��D(�IgÊÆ á21:á �� �� á � '� �A�K�)� á � á D á (E.8)
The inverse Fourier Transform of the expressions above have been computed indirectly,
since direct computation is difficult. It can be easily shown that,
=n�0/ á D á � g Æ l�/� �=n�0/I�,+¡ M��Æ21 á D á �[� g l3/411 � LN� �
Hence, conversely,
= ü 1  �� � Ä g Æ �l á D á= ü 1 � �1 � LN� � ® g �l�1 �3+¡ M��Æ21 á D á �
These results have been used in computing the inverse Fourier Transforms above.
100
APPENDIX F
DERIVATION OF �2�`kÑm\o FOR WHITE NOISE CASE
Given the value derived for +-@:�0D(� and +Y;8= and using -Q�Bg¹+Y@¡��uR� we obtain,
+<,��0DJ� g �l � � �� L á21:á ���� � �3+¡ zØ:Æ �Y� � á D áá � '� ������� á � Ý ®�E�F�Gj�����4DJ�o �,+¡ � Æ á21:á �� �� á � '� ������� á � ؤ��L ñ��� � Ý á D á Æ ñ�� á21:á �� � � Ø 1 ü ( Ù é* + b¡ � - é « * 9 9 ;�9 Æ�� Ý ® (F.1)
Now taking the Fourier Transform of the above equation and proceeding along the Method
of stationary phase (shown in Appendix H to obtain the Fourier Transform of an exponential
of an exponential, we obtain,
5�,ç} ·>·Z? �A�B� g � �� � d 1d � 1 L¹�A�kÆk����� � ®@ AÑB DEaFOHKJ «2L¥¬ L þ � d 1QL d �� d 1QL d � � � L¹���hÆk����� � ®@ ACB DN FOHPJ «2L�¬ (F.2)
where, d 1ýg 9 ² 9 9= 9é 9 b¡ ¢ èjéC£ 9 9 and d � g = é c9 b¡ ¢ èjéC£ 9 9 þ g 9 ² 9 9= é c . Since �A��L��K���W� is very high
compared to d � 1 , those terms containing it in the denominator have been neglected.
101
APPENDIX G
DERIVATION OF THE PHASE NOISE WITH BUFFER NOISE
−
+ + +
+
+
+
(a)
2i2
vn2i1
v1v2v−
−
−−
n1i
n
1vni
−1i 2iv
(b)
Figure G.1: Conventional (a) and modified (b) 2-port noise model equivalent circuits
Consider the two circuits shown in Fig G.1.
Let the Z-parameters of the circuit be yR1[13 )yç1 � )y � 13 and y �[� . For the circuit of Fig G.1(a)
{ � g �0Z�1KÆNZ8|R�!y � 1QL§Z � y �[�Ê { � g ZW1!y � 1LtZ � y �[� ÆNZ8|Hy � 1 (G.2)
102
Similarly for the circuit of Fig G.1 (b) the voltages {:1 and { � can be given as,
{R1 g ��ZW1CÆÇZ�|ç1!��yç1[1LtZ � yç1 �Ê {R1 g Z�1!yç1[1QL§Z � yç1 � ÆÇZ�|ç1�yç1[1 (G.3)
Again
{ � Æk{S| � g �0Z�1�ÆÇZ8|S14�!y � 1QL§Z � y �[�Ê { � g ZW1�y � 1LtZ � y �[� L§{j| � ÆÇZ8|S1�y � 1 (G.4)
Comparing Eqns G.1 and G.2 we get,
Z8|S1.g�Z�|�L {S|yç1[1and from Eqns G.3 and G.4, {j| � g y � 1yç1[1 {j|Alternatively, the voltages {�| and Z8| can also be given in terms of Z^|S1 and {j| � as follows,
DERIVATION OF THE FOURIER TRANSFORM OF THE MOREGENERAL EXPONENTIAL OF AN EPONENTIAL
We will try to obtain an expression for the Fourier transform of expressions of the form
�C�0DJ�?gm�,+¡ � Æ�¯ Ø ø 1 á D á L |� # � � ø #4 # � 1 ü87 ½ 9 ;�9 Æ��j� Ý ®where, |� # � � ø # g ø 1 and ¯$/�uNow taking the Fourier transform of �C�0/!� we get,
Ó �A�B� g =z� �C��D(�[�g ¸Oûü û �,+¡ � ÆY¯ Ø ø 1 á D á L |�# � � ø #4 # 6 1 ü87 ½ 9 ;�9 Æõ� < Ý ® 1 ü # è ; ]RD
g ¸ �ü û �,+¡ � ÆY¯�ØJÆ ø 1WD�L |�# � � ø #4 # � 1 7 ½ ; Æ��j� Ý ® 1 ü # è ; ]RD
L ¸ û� �,+¡ � ÆY¯ Ø ø 1WD�L |�# � � ø #4 # 6v1 ü87 ½ ; Æõ��< Ý ® 1 ü # è ; ]RDg ¸ �ü û
I.1 Obtaining the Fourier Transform of the form >Uk65qo@?�A B}kÑm�n&5qoThe results obtained in the preceding section can be used to compute the Fourier trans-
form of functions having the form �C�0DJ�:E�FHG ���C�4D(�=z� �C��D(�:E�FHG��A�K�!D(�W�Ùg ¸ ûü û �C��D(�:E�F�G������4DJ�:�,+¡ M��Æ�Z��IDJ�!]HDg �l � ¸Oûü û �C��D(�· �,+¡ M�0Z����4D(�Lt�,+¡ M��Æ�Z����2D(�l º(® �3+¡ M��Æ�Z��ID(��]HDg �l  ¸ ûü û �C��D(�:�3+: M�WÆ�Z2���kÆh���3�WDJ�!]HD$L ¸ ûü û �C�0DJ�:�,+¡ ��Æ�Z)�A� LN���3�WD(��]HD¡Äg �l � l�¯���kÆh����� � |�# � � 4¡�# ø #4 �# Lm���hÆk����� � L l�¯�A�xLN����� � |� # � � 4¡�# ø #4 �# L¹�A�zLÇ���3� � ®
(I.5)
At RF frequencies ���hL��K����� becomes very large and hence the second term becomes
quite small compared to the first and hence can be neglected. So the final expression be-
[1] Demir A. “Phase Noise And Timing Jitter in Oscillators With Colored-NoiseSources”. IEEE Transactions on circuits and systems-1, 49(12):1782–1791, Dec2002.
[2] Hajimiri A and Lee T.H. “A General Theory of Phase Noise in Electrical Oscillators”.IEEE Journal of Solid-State Circuits, 33(2), February 1998.
[3] Hajimiri A and Lee T.H. “Design Issues in CMOS Differential LC Oscillators”. IEEEJournal of Solid-State Circuits, 34(5), May 1999.
[4] Hajimiri A and Lee T.H. “The Design Of Low Noise Oscillators”. Kluwer AcademicPublishers, first edition, 1999.
[5] Hajimiri A and Lee T.H. “Oscillator Phase Noise: A Tutorial”. IEEE Journal ofSolid-State Circuits, 35(3), March 2000.
[6] Van Der Ziel A. Noise Sources, Characterization, Measurement. Prentice-Hall, Inc.,Englewood Cliffs, N.J., first edition, 1970.
[7] Razavi B. “Analysis, modeling and simulation of phase noise in monolitic voltage-controlled oscillators”. Proc IEEE Custom Integrated Circuit Conf, May 1995.
[8] Razavi B. “A Study of Phase Noise in CMOS Oscillators”. IEEE Journal of SolidState Circuits, 31(3):331–343, March 1996.
[9] Razavi B. “RF Microelectronics”. Prentice Hall, first edition, 1998.
[10] Leeson D.B. “A Simple Model of Feedback Oscillator Noise Spectrum”. Proc IEEE,54:329–330, February 1966.
[11] Roychowdhury J Demir A, Mehrotra A. “Phase Noise in Oscillators: A UnifyingTheory and Numerical Methods for characterization”. IEEE Transactions on circuitsand systems-1, 47(4):655–674, May 2000.
[12] Hafner E. The effect of noise in oscillators. Proc IEEE, 54(2):179, Feb 1937.
113
[13] E.S. Ferre-Pikal et al. Draft Revision of IEEE STD 1139-1988 Standard Definitionsof Physical Quantities for Fundamental Frequency Theory and Applications. Proc.IEEE International Frequency Control Symposium, pages 338–357, 28-30 May 1997.
[14] Risken H. “The Fokker-Planck Equation”. Springer-Verlag, 2nd edition, 1989.
[15] Bernamont J. Proc Physics Society, 49, 1937.
[16] K Kurokawa. Microwave Solid State Oscillator Circuits. Addison-Wesley, Secondedition, 1968.
[17] McWhorter A. L. “1/f noise and related surface effects in germanium”. R.L.E. 295and Lincoln Lab Tech Rep,80 MIT, March 1955.
[18] Brian Nguyen Limketkai. “PhD Thesis:Oscillator Modelling and Phase Noise”.Electrical Engineering and Computer Science Dept, UC Berkeley, 2004.