Chengzhou Wang
Chengzhou Wang
Chengzhou Wang
by
Chengzhou Wang
Committee in charge:
Professor Lawrence E. Larson, Chair
Professor Peter M. Asbeck
Professor Walter H. Ku
Professor Chung-Kuan Cheng
Professor Bill Hodgkiss
2003
Copyright
Chengzhou Wang, 2003
All rights reserved.
Chair
2003
iii
iv
TABLE OF CONTENTS
iv
Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
I.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
I.1.1 Power Amplifiers in Wireless Communication Systems . . . . . . . . . . .
I.1.2 Power Amplifier Classifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
I.2 Limitations of Sub-micron CMOS Technology . . . . . . . . . . . . . . . . . . . . . . . .
I.2.1 Low Breakdown Voltages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
I.2.2 Low Transconductance-to-current Ratio . . . . . . . . . . . . . . . . . . . . . . .
I.2.3 Low Substrate Resistivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
I.3 Dissertation Motivations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
I.4 Dissertation Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1
1
1
3
4
4
5
6
6
7
II
9
9
11
11
13
15
18
19
22
22
25
26
27
29
vi
IV.3
IV.4
IV.5
IV.6
V
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
vii
LIST OF FIGURES
II.1
II.2
II.3
II.4
II.5
III.1
III.2
III.3
III.4
III.5
III.6
III.7
III.8
III.9
III.10
III.11
III.12
III.13
III.14
III.15
12
32
viii
21
22
24
28
33
35
37
38
39
41
45
46
49
50
53
54
56
56
III.16 Plots of Id versus VGS for an ideal class-B operation, and a short-channel
device biased near the threshold voltage. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
III.17 Schematic and equivalent circuit of a high-pass, L-match network . . . . . . . .
III.18 Cascade of two lossy L-match networks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
III.19 Output matching networks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
III.20 Circuit and equivalent model of the interstage matching network. . . . . . . . .
III.21 Schematic and linear model of the two-stage CMOS class-AB power amplifier for illustrating ground connections. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
III.22 Two-stage CMOS class-AB PAs for illustrating the impact of ground connections on gain. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
III.23 Power gain of the two-stage CMOS class-AB power amplifiers versus total ground bondwire inductance for the two ground configurations shown
in Fig. III.22. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
III.24 Two-stage CMOS class-AB power amplifier for one-chip-ground and twochip-ground configurations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
III.25 Small-signal equivalent model of the two-stage CMOS class-AB power
amplifier for one-chip-ground and two-chip-ground configurations. . . . . . . .
III.26 Maximum stable ground bondwire inductance of the two-stage CMOS
class-AB PA for the ground configurations in Table III.2. . . . . . . . . . . . . . . .
III.27 Schematic of the fully matched two-stage CMOS class-AB power amplifiers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
III.28 Layout of a basic transistor cell. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
III.29 Schematic modelling and sideview of device layouts regarding the effect
of substrate coupling. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
III.30 Layout structure employing both large substrate guardrings and deep trench
blocks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
III.31 Final Layout of the fully integrated and compensated two-stage CMOS
PA (PA2). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
III.32 Die microphotograph of the fully integrated and compensated two-stage
CMOS PA (PA2). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
III.33 Photograph and cross section drawing of the MLF package. . . . . . . . . . . . . .
III.34 Output and input off-chip matching network for PA3. . . . . . . . . . . . . . . . . . .
III.35 ADS schematic and simulated impedance of off-chip output matching network for PA3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
III.36 Application schematic of PA3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
ix
58
61
63
64
67
71
73
74
76
77
80
81
84
87
88
89
90
91
93
95
97
IV.7
IV.8
IV.9
IV.10
IV.11
IV.12
IV.13
IV.14
Conceptual block diagram and actual implementation of the dynamic biasing technique. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
Schematic and equivalent large-signal model of the envelope detection
and gate-bias-control circuit. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
Source-drain current of Mp as a function of time. . . . . . . . . . . . . . . . . . . . . . . 111
SPECTRE simulated and MATLAB fitted PMOS source-drain current versus gate voltage. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
Schematic of the designed two-stage CMOS power amplifier. . . . . . . . . . . . 115
Approximate time-domain waveforms of the input gate voltage, sourcedrain current of Mp , and output voltage of the envelope detector for a
two-tone test signal. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
Circuit for the Volterra calculation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
IDS versus VGS for a long-channel class-A device and an ideal class-B device.126
Ratio of g2 and g3 to g1 for the four gate bias voltages (0.75 - 0.90 V) of
the implemented class-AB device. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
Comparison of the calculated and simulated IM3 of the load voltage at
21 2 versus peak-envelope output power. . . . . . . . . . . . . . . . . . . . . . . . . . 129
Contributions to the load-voltage IM3 from the g2 , g3 , and Ceff nonlinearities.130
Die microphotograph of the highly integrated and compensated two-stage
CMOS PA (PA3). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
Calculated, simulated, and measured Venv versus output power for a singletone input. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
Measured gain and power-added efficiency versus output power for PA3
with the dynamic biasing technique, and PA3 when the envelope detector
is disabled and the gate is biased at VGG0 = 0.85 V, respectively. . . . . . . . . . 133
IV.15 Measured power consumption improvement versus the PA output power. . . 134
IV.16 Measured IM3 , adjacent-channel leakage power, and alternate-channel
power versus peak-envelope output power for the three PAs. . . . . . . . . . . . . . 135
IV.17 Comparison between the calculated and measured load-voltage IM3 versus output power. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
xi
LIST OF TABLES
I.1
I.2
II.1
II.2
III.1
III.2
III.3
III.4
III.5
Characteristics of digital wireless systems relevant to power amplifier performance in mobile station [1]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Comparison of efficiency and linearity for different classes of power amplifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3
4
xii
ACKNOWLEDGEMENTS
Pursuing the doctoral degree begins with a great deal of excitement and expectations.
However, after years of endeavors, the initial love and devotions to the research topic gradually become frustrations by the overwhelming obstacle and adversity, and I start to realize
that this journey would never be fulfilled without the support of many people.
First and foremost, I would like to express my sincere gratitude and appreciation to
my advisor Professor Lawrence E. Larson. Without his continuous guidance and encouragement, my research work towards this thesis would never be possible. I would also like
to thank Professor Peter M. Asbeck for his assistance throughout the years, as his knowledge and experience were also valuable. Furthermore, I am indebted to Professor Walter
H. Ku, Professor Chung-Kuan Cheng, and Professor Bill Hodgkiss, for their patience and
efforts of being my dissertation committee.
This work has benefited from the contributions of many other individuals. Special
thanks to Mani Vaidyanathan and Liwei Sheng for their invaluable suggestions and ideas
to my research projects. I would also like to thank John Fairbanks, Matt Wetzel, Jonathan
Jensen, and Masaya Iwamoto for their assiduous assistance in solving the laboratory and
CAD problems I experienced. In addition, Xuejun Zhang, Junxiong Deng, Vincent Leung,
Don Kimball, Robert Wang and other brilliant colleagues deserve my sincere thanks for
their enthusiastic help and encouragement.
Finally, I am grateful to my family (my parents and my sisters Judy and Fang) and
my dear friends Changchun Shi, Yong Wang, Wei Lin, Jian Ma and others. Without their
xiii
continuous support, this Ph.D work would have been much more difficult.
This research was supported by the UCSD Center for Wireless Communications and
its member companies. Their supports are greatly acknowledged.
xiv
VITA
1992-1997
1997-1999
1999-2003
Ph.D., Electrical Engineering (Electronic Circuits & Systems), University of California, San Diego, United States
PUBLICATIONS
C. Wang and L.E. Larson, Analysis of a Microwave CMOS Class-E Power Amplifier
with Finite Switching On-resistance, 1999 IEEE Topical Workshop on Power Amplifier
for Wireless Communications, La Jolla, CA, Sept. 1999.
C. Wang and L.E. Larson, Highly Integrated Linear Class-AB CMOS Power Amplifier
with Nonlinear Capacitor Compensation, 1999 IEEE Topical Workshop on Power Amplifier for Wireless Communications, La Jolla, CA, Sept. 2000.
C. Wang, L.E. Larson, and P.M. Asbeck, A Nonlinear Capacitance Cancellation Technique
and its Application to a CMOS Class-AB Power Amplifier, presented at 2001 IEEE International Microwave Symposium (RFIC), Phoenix, AZ, May 2001.
C. Wang, L.E. Larson, and P.M. Asbeck, Improved Design Technique of a Microwave
Class-E Power Amplifier with Finite Switching On-Resistance, presented at 2002 IEEE
Radio and Wireless Conference, Boston, MA, Aug 2002.
C. Wang, M. Vaidyanathan and L.E. Larson, A Capacitance-Compensation Technique for
Improved Linearity in CMOS Class-AB Power Amplifiers, submitted to IEEE Journal of
Solid-State Circuits.
C. Wang, and L.E. Larson, A Dynamic Biasing Technique for Efficiency Improvement in
CMOS Class-AB Power Amplifiers , in preparation to IEEE Transactions on Microwave
Theory and Techniques.
FIELDS OF STUDY
Major Field: Electrical and Computer Engineering
Studies in Radio Frequency Integrated Circuit Design.
Professor Lawrence E. Larson
xv
by
Chengzhou Wang
Doctor of Philosophy in Electrical Engineering (Electronic Circuits & Systems)
University of California, San Diego, 2003
Professor Lawrence E. Larson, Chair
Linearity and efficiency are the two most important characteristics of power amplifiers (PAs) for wireless applications. In this dissertation, we investigate three topics on
CMOS power amplifiers: class-E, class-AB, and dynamic biasing technique.
Previous analytical efforts on class-E power amplifiers assumed either zero switch
resistance and/or infinite drain inductance, leading to less optimized design. In this dissertation, we developed an improved design technique by accounting for both finite drain
inductance and finite on resistance for a CMOS device. A design example based on the
developed algorithm achieves an output power of 0.25 W and a drain efficiency of 87% for
a 3.5 mm NMOS class-E device with VDD = 2 V and fc = 1.90 GHz.
The intrinsic linearity obtained in a CMOS class-AB operation is often insufficient
to meet the stringent linearity requirement imposed by modern wireless standards. In this
dissertation, we propose a capacitance compensation technique to improve PA linearity.
xvi
Experiments show that the compensation technique can improve both the two-tone, thirdorder intermodulation (IM3 ) and adjacent-channel leakage power (ACP) by approximately
8 dB. While meeting the 3GPP-WCDMA ACP requirements, the linearized two-stage amplifier is capable of delivering an output power of 24 dBm with a small-signal gain of nearly
24 dB and an overall power-added efficiency of 29 %.
The designed two-stage CMOS class-AB power amplifier suffers serious efficiency
degradation when operated at low output power levels. In this dissertation, it was demonstrated that a dynamic biasing technique can improve the average efficiency of a CMOS
class-AB power amplifier by controlling the gate bias voltage with the envelope of input
RF signal. However, the envelope signal introduced by the dynamic biasing technique can
significantly limit the overall linearity of the CMOS class-AB PA. Both analysis and experiments show that the dynamic biasing technique can significantly degrade the IM3 and
ACP performances of the designed two-stage CMOS class-AB power amplifier.
xvii
Chapter I
Introduction
I.1 Background
I.1.1 Power Amplifiers in Wireless Communication Systems
Recent years have witnessed a tremendous growth of wireless communication products. Consumer electronics, such as cellular phones, wireless local area networks, and
wireless computer peripherals, are just a few examples of the wireless devices that become part of our everyday lives. This constantly growing market drives an intense effort
to develop improved wireless standards and transceiver architectures, as well as reduce
implementation costs by using low-cost technologies and higher integration solutions.
Current implementation of wireless communication devices, such as cellular phones,
employ several chips implemented in different semiconductor technologies in order to realize high performance digital, analog, and RF circuit building blocks. Different technologies
are suited for different functions. For example, CMOS models an ideal switch very well,
thus is very suitable for digital functions and switch-capacitor circuits, but it is a poor technology for high-frequency, high performance analog functions for its low transconductance
and large parasitics; on the contrary, bipolar is well suited for high-frequency, high perfor-
2
mance analog functions, but not ideal for realizing digital functions and switch-capacitor
circuits due to its finite base current and other non-ideal switching characteristics.
The multi-chip solution limits the minimum cost and size of the final device. In
addition, the interface matching between different chips also adds cost, size, and time-tomarket to the final products. Thus, a single-chip solution is highly desirable.
With the advance of CMOS technology, many RF front-end functions, such as lownoise amplifier, mixer, and voltage-controlled oscillator can be implemented in a low-cost,
high-volume CMOS technology. However, a fully integration of CMOS power amplifier
(PA) still remains a design challenge because, as described later in this section, the limitations of the CMOS technology is especially severe for PA implementations.
Another reason for the reluctance of implementing PA in CMOS technology is the
high performance requirements imposed by modern wireless standards. With the growing
emphasis on channel capacity, more and more wireless communication systems employ
spectrally efficient modulation schemes, such as QPSK and QAM. These schemes results
in signals with highly time-varying envelopes, thus imposes a stringent linearity requirement of the power amplifiers to preserve modulation accuracy and limit spectral regrowth.
Meanwhile, to prolong battery life, a reasonable efficiency is also required for the power
amplifiers in such systems. Table 1 lists features pertinent to power amplifier design for
several digital wireless standards.
Table I.1: Characteristics of digital wireless systems relevant to power amplifier performance in mobile station [1].
CELLULAR
CORDLESS
Standard
GSM
NADC
IS-95
PDC
PHS
Uplink frequency
890-915
825-849
825-849
940-956
1895-1907
Channel BW (kHz)
200
32.81
1223
31.5
288
Multiple access
TDMA
TDMA
CDMA
TDMA
TDMA
Modulation
GMSK
/4-QPSK
/4-QPSK
Duplex mode
FDD
FDD
FDD
FDD
TDD
30
27.8
27.8
30.0
19.0
Long-term mean
21.0
23.0
10.0
N/A
10.0
12.5
33.3
Variable
33.3
33.3
PA voltage (V)
3.56.0
3.56.0
3.56.0
3.54.8
3.13.6
ACPR (dBc)
N/A
-26
-26
-48
-50
3.2
5.1
2.6
2.6
Typical PA quies-
20
180
200
150
100
> 50
> 40
> 30
> 50
> 50
band (MHz)
/4-QPSK O-QPSK
power (dBm)
4
two conflicting parameters in PA design. Table 2 shows the comparison of efficiency and
linearity among these power amplifiers.
Table I.2: Comparison of efficiency and linearity for different classes of power amplifiers
Classification
AB
Maximum Efficiency(%)
50
50-78
78
100
100
100
100
Typical Efficiency(%)
35
35-60
60
70
75
80
75
Linearity
Excellent
Good
Good
Bad
Bad
Bad
Bad
Vpeak(V)
2VDD
2VDD
2VDD
2VDD
2VDD
3.6VDD
2VDD
5
device is greatly limited.
In addition to gate oxide breakdown, the drain-substrate pn junction will conduct a
large current if the reverse bias applied to it exceeds a certain value [2]. This breakdown is
nondestructive, but limits the maximum PA voltage swing at the drain of the device.
(I.1)
For a bipolar device, this ratio is 1/VT , where the thermal voltage VT is 26 mV. In contrast,
the overdrive Vov for MOS transistors is typically chosen as several hundred mV. Thus,
the transconductance per given current is much lower for MOS devices than for bipolar
devices.
To accommodate this small transconductance, either the input signal amplitude or
the device size of the PA output stage have to be increased. However, either approach will
increase the loading for the driving stage, thus resulting in higher power consumption of
the driver stage. Increasing the input signal amplitude can also dramatically degrade the PA
linearity because the third-order nonlinearity of the device current is directly proportional
to the cube of the input voltage amplitude. Thus, higher nonlinearity will be expected for
MOS devices than for bipolar devices.
7
communications.
Although the designed two-stage CMOS class-AB PA exhibits good linearity and
maximum efficiency, it still suffers significant efficiency reductions when operated at low
power levels. Thus, we would like to explore the possibilities to improve the efficiency of
the CMOS class-AB PA at low output power levels.
8
also proves the feasibility of linear CMOS class-AB PAs for wireless communication systems.
Chapter IV describes a dynamic biasing technique to improve PA efficiency at low
output power levels. The transient response of the envelope detector and the average efficiency for a CMOS class-AB PA are analyzed. The impact of the technique on PA linearity
is verified using both Volterra analysis and SPECTRE simulations. Finally, the experimental results of a prototype amplifier is presented.
Chapter V concludes the whole dissertation.
Chapter II
Class-E Power Amplifiers
II.1 Introduction
The class-E amplifier was first introduced by Ewing [5] in his doctoral thesis, and
then was further elaborated by many other researchers [6]-[11]. As one of the switchingmode amplifiers, the class-E amplifier realizes very high efficiency (theoretically 100%) by
operating the device as a switch, i.e.,
1. The device sustains zero voltage when it carries current.
2. The device carries zero current when it sustains a finite voltage.
3. There is no transition time between the on and off states of the device.
This is also referred as the non-overlapping-current-and-voltage condition and underlies
all switching-mode amplifiers. One unique feature, which distinguishes the class-E amplifier from other switching-mode amplifiers, is that it requires zero slope of the drain (or
collector) voltage at the moment when the device turns on. This requirement substantially
lowers the sensitivity of the amplifiers efficiency as a function of the component variations
and other non-ideal effects in practical implementations [12][13].
10
Ewings original development of the class-E amplifier assumed an infinite collector
inductance, but a finite saturation resistance of the transistor. In 1975, Sokal [6] derived a method to analyze class-E amplifiers in optimum performance, where he assumed
both an infinite choke inductor and zero switching-on resistance. In Lis analysis [8], a
finite choke inductor was introduced, but with an ideal switching-on condition. Recently
published class-E literatures [9]-[11] relied on the previously reported analysis and concentrated mostly on the implementation details. In practice, however, both the switching-on
resistance of the active device and the choke inductance are finite. The latter is especially
true if MOS devices are used; as will be shown in this chapter, the large shunt parasitic
capacitance of MOS devices requires relatively small drain inductance to achieve the optimum class-E operation at gigahertz-range frequencies. Therefore, an improved analytical
method which takes into account both the finite drain inductance and finite switching-on
resistance is necessary. In this chapter, this optimum is established under the constraint of a
given Rswitch-on Cswitch-off product, which is a more realistic estimate of typical MOS devices.
This new technique expresses the circuit parameters in terms of the device width and the
design specifications, such as the output power and operating frequency. The agreement
obtained between the analytical and simulated results is outstanding, verifying the utility
of the technique.
This chapter begins with a brief class-E circuit description, followed by a detailed
presentation of the improved analytical method. Then, a design example based on the
developed algorithm is described and SPICE simulation results are compared with the the-
11
oretical calculations. Finally, key design issues, such as the validity of assumptions and the
choice of device width, are discussed and conclusions are summarized.
12
VDD
L1
L2
C2
Vo
M0
RL
Vs
(a)
VDD
L1
i L(t)
i d(t)
on
R on
off
vd (t)
v2 (t)
i o (t)
resonator at
jX
vo (t)
C1
RL
(b)
Figure II.1: CMOS class-E power amplifier. Part (a) shows the simplified schematic, and
part (b) shows the model accounting for both finite on resistance and finite drain inductance.
13
shaping the current and voltage waveforms for the optimum class-E operation.
(II.1)
where Io is the amplitude of the output current, and o is the phase shift constant.
The voltage at node 2 is also sinusoidal but with an extra phase shift by jX:
v2 (t) = V sin(c t + 1 )
(II.2)
where
s
V
1
X2
= Io RL
1+ 2
RL
X
= o + tan1
RL
(II.3a)
(II.3b)
14
Drain inductor current and drain voltage
To evaluate the current flowing through the finite drain inductor L1 , we apply KCL
at node 1:
iL (t) = id (t) + Io sin(c t + o ).
(II.4)
From the inductor characteristics, iL (t) is related to the drain voltage vd (t) by
VDD vd (t) = L1
diL (t)
.
dt
(II.5)
Since the device is switched between on and off states, the operation of the amplifier
can be divided into two parts:
Off state (nT t (n + 12 )T ): When the active device is off, vdoff (t) and idoff (t) are
governed by the characteristics of the capacitance C1 , i.e.,
idoff (t) = C1
dvdoff (t)
.
dt
(II.6)
Substituting (II.6) and (II.5) into (II.4) results in the following second-order differential equation:
L1 C1
d2 iLoff (t)
+ iLoff (t) = Io sin(c t + o ).
dt2
(II.7)
Io
sin(c t + o )
1 2
(II.8)
where
o =
1
L1 C1
= c /o
(II.9a)
(II.9b)
15
and the coefficients A and B are two constants to be determined.
On state ((n + 12 )T t (n + 1)T ): When the active device is on, it is modelled
as a small resistor, thus
vdon (t) = idon (t)Ron .
(II.10)
Substituting (II.10) and (II.4) into (II.5) gives a first-order differential equation:
VDD iLon (t)Ron + Io Ron sin(c t + o ) = L1
diLon (t)
.
dt
(II.11)
Io
VDD
+ Cet (II.12)
[ sin(c t + o ) c cos(c t + o )] +
2
+ c
Ron
Ron
.
L1
(II.13)
II.2.3 Conditions
To evaluate the constants A, B, C, Io and o , we need to apply the periodic, boundary,
and class-E conditions to the above circuit equations. Those conditions are:
Periodic conditions: According to the characteristics of the inductance and capacitance, the current of L1 and the drain voltage (also the voltage of C1 ) satisfy
(II.14a)
(II.14b)
16
Boundary condition: iLon (t) must be continuous, i.e.,
(II.15)
= 0
(II.16a)
dvdoff (t)
t=(n+1/2)T
dt
= 0.
(II.16b)
Substituting (II.6), (II.8), (II.10), and (II.12) into (II.14)-(II.16) gives the following equation array:
VDD
Io
Io sin o
+ CeT + 2
(II.17a)
(
sin
cos
)
=
A
+
o
c
o
Ron
( + c2 )
(1 2 )
Io
B Io cos o
CeT 2
(II.17b)
(
cos
sin
)
=
o
c
o
c
( + c2 )
(1 2 )
VDD
Io
+ CeT /2 2
+
B
sin
(
sin
cos
)
=
A
cos
o
c
o
Ron
( + c2 )
Io sin o
(II.17c)
(1 2 )
Io c
VDD
Ao sin Bo cos +
cos
(II.17d)
=
(1 2 )
Ron
Io c
VDD
CeT /2 + 2
(II.17e)
(
cos
sin
)
=
o
c
o
( + c2 )
Ron
Here, VDD , T , and c are fixed by the design specifications; and are functions of L1 and
C1 ; Ron and C1 are determined by the choice of device size.
17
DC Power Dissipation and Output Power
The total dc power Pdc is defined as the product of the power supply voltage VDD and
the dc current Idc drawn from the power supply:
Pdc = VDD Idc
(II.18)
where
1
Idc =
T
=
1
T
(n+1)T
iL (t) dt
nT
Z
(n+1/2)T
iLoff (t) dt +
nT
(n+1)T
iLon (t) dt .
(II.19)
(n+1/2)T
Meanwhile, Pdc is the sum of the power consumed by the load and the power dissipated in
the active device, i.e.,
Pdc = Pout + Pd .
(II.20)
During the switching-off period, only capacitive current flows into the device, implying no dc power dissipation in the device; during the switching-on period, Ron is the only
source of power consumption. Thus, the total power dissipation in the device, during one
period, is
1
Pd =
T
(n+1)T
(n+1/2)T
(II.21)
(n+1/2)T
iLoff (t) dt +
nT
(n+1)T
iLon (t) dt
(n+1/2)T
1
= Pout +
T
(n+1)T
(n+1/2)T
i2don (t)Ron dt
(II.22)
18
where iLoff (t) and iLon (t) are expressed in (II.8) and (II.12), respectively, and idon (t) is
related with iLon (t) by (II.4).
Io2
RL .
2
(II.23)
(n+1)T
(II.24)
19
Substituting (II.3a) into (II.24) results in
s
Io RL
X2
1+ 2
RL
2
=
T
(n+1)T
vd (t) sin(c t + 1 ) dt
(II.25)
nT
thus the excessive reactance X is evaluated. There is no specific requirements for the
loaded Q of the output matching network, as long as it is large enough to allow a sinusoidal
output only. In practice, a Q of 5 is enough. Once Q is chosen, L2 is evaluated by
Q=
c L2
.
RL
(II.26)
and C2 is solved by
jc L2 +
1
= jX.
jc C2
(II.27)
Based on (II.23)-(II.27), the design algorithm of a CMOS class-E amplifiers in optimum performance is straightforward. MATHEMATICA scripts were developed to perform
the calculations.
20
The device parameters are those of a 0.6 m digital CMOS technology, in which Ron is
approximately 3 and C1 is roughly 1 pF for a 1 mm device.
We first picked the NMOS width (WN ) as 3.5 mm, so the values of Ron and C1
were obtained. Following the component-evaluation procedure described in Section II.2.4,
we were able to compute L1 , L2 , C2 , and RL , as well as the expressions of id (t), iL (t),
io (t), and vd (t). Then, HSPICE netlists were constructed and simulated, and the results
were compared with the theoretical calculations. Table II.1 shows such comparison for the
amplifiers output power Pout and drain efficiency Peff ; also shown are the employed component values. As can be seen, less than 5% difference between the theoretical prediction
and simulation was achieved, verifying the utility of the technique.
The calculated and simulated current and voltage waveforms are compared in Fig. II.2.
Note that the pike in vd (t) comes from the sharp transition of the input square-wave signal.
Table II.1: Comparison of Pout and Peff between theoretical prediction and HSPICE simulation for the designed CMOS class-E power amplifier.
WN (mm)
RL ()
Pout (W)
Peff (%)
Theory
3.5
7.1
17
0.248
85
Simulation
3.5
7.1
17
0.25
87
Since we have the freedom in choosing the NMOS width WN , further calculations
and simulations were performed by sweeping WN from 2.5 mm to 5 mm. The resulting
Pout and Peff are shown in Fig. II.3.
21
0.5
0.4
0.3
CURRENT (A)
i (t)
L
0.2
0.1
i (t)
d
0.1
Calculation
Simulation
0.2
0.3
7.8
7.9
8.1
8.2
8.3
8.4
8.5
TIME (ns)
(a)
8
Calculation
Simulation
VOLTAGE (V)
6
vd(t)
4
vo(t)
2
0
2
4
7.8
7.9
8.1
8.2
8.3
8.4
8.5
TIME (ns)
(b)
Figure II.2: Comparison of the current and voltage waveforms between the calculation and
simulation. Part (a) shows the drain inductor current iL (t) and the drain current id (t); part
(b) shows the drain voltage vd (t) and the load voltage vo (t).
0.25
95
0.2
90
0.15
85
0.1
80
0.05
2.5
Calculation
Simulation
3
3.5
4.5
22
75
Figure II.3: Output power and the drain efficiency versus NMOS width.
II.4 Discussions
II.4.1 Validity of Assumptions
As described in Section II.2.1, the following assumptions were made for our analysis:
Ron , the switching-on resistance of the NMOS transistor, is constant and dominates
the total output impedance of the device during the on period.
C1 , the switching-off capacitance of the NMOS transistor, dominates the total output
impedance of the device and is independent of the switch voltage Vd (t) during the
off period.
The loaded quality factor (Q) of the output circuit is high enough to allow a sinusoidal
output only.
Since the third assumption can be easily met by a proper choice of Q, we will only investigate the first two assumptions.
23
When the NMOS transistor turns on, it is in the triode region. Since the VDS is small,
the simplified model shown in Fig. II.4(a) is often used [14]. Here, rds corresponds to Ron ,
and is given by
rds =
n Cox
W
L
(II.28)
(VGS VTn )
The gate-to-channel capacitance is evenly divided between the source and drain nodes,
Cgs = Cgd =
W LCox
.
2
(II.29)
The channel-to-substrate capacitance is divided in half and shared between the source and
drain junctions. At the drain node, this channel capacitor, together with the junction-tosubstrate capacitance and the junction-sidewall capacitance, consists of the drain-bulk capacitance:
Cdb = Cj0 (Ad +
Ach
) + Cj-sw0 Pd .
2
(II.30)
For typical CMOS processes, Cdb and Cgd are in the range of 1 pF/mm. The quantity rds
depends on the gate-source voltage VGS , but has typical values of 2-4 /mm. At 1.90 GHz,
the impedance of Cdb and Cgd are much higher than the on resistance, thus verifying our
utilization of the first assumption.
When the transistor turns off, the model changes dramatically. A reasonable model
is shown in Fig. II.4(b). Since the channel has disappeared, Cgs and Cgd are now due to
only overlap and fringing capacitances:
Cgs = Cgd =
W Lov Cox
.
2
(II.31)
24
Vg
Cgs
Vs
rds
Csb
Cgd
Vd
Cdb
(a)
Vg
Cgd
Cgs
Vs
Vd
Cgb
Csb
Cdb
(b)
Figure II.4: Simplified NMOS small-signal model (a) in triode region and (b) in cut-off
region.
25
The capacitor Cdb , which is also smaller when the channel is not present, is
Cj0 Ad
.
VDB
1+
0
Cdb = r
(II.32)
The total drain capacitance, if the input is treated as ac ground, is the sum of Cgd and Cdb .
Thus, we have
C1 =
Cj0 Ad
W Lov Cox
.
+r
2
VDB
1+
0
(II.33)
As shown in (II.33), C1 is a nonlinear function of its own voltage VDB , as opposed to the
constant switching-off capacitance of the second assumption. The simulations, however,
showed that this nonlinear capacitance does not introduce significant errors, as illustrated
in Fig II.2 and II.3.
26
(II.34)
Io0 = kIo
(II.35)
A0 = kA
(II.36)
B 0 = kB
(II.37)
C 0 = kC
(II.38)
and
all the equations will be reduced to their original forms. Therefore, all the component
values L1 , L2 , C2 , and RL are unchanged, and all the current and voltages id (t), iL (t),
io (t), and vd (t) are k times their original values. The output power becomes
0
Pout
= k 2 Pout .
(II.39)
This implies a perfect application of class-E power amplifiers in envelope elimination and
restoration (EER) systems, where the envelope variation of the modulated signal is imposed
to the switching power amplifier through the power supply.
It is important to mention, however, that our analysis assumed the constant switchingoff capacitance C1 . For the actual devices, as described in Section II.4.1, C1 is nonlinear
and varies with the drain-voltage swing; this may introduce some errors.
27
Some papers [15][16] claimed that the nonlinear parasitic capacitor does not influence the class-E performance. This conclusion, however, is made based on the ideal
switching-on condition (Ron =0) and infinite drain inductance. In addition, the resulted operation, due to the nonlinear capacitance, does not predict the linear relationship with VDD ,
as opposed to our above conclusion. The details can be seen in (4.1)-(4.5) of [15].
Sokal [75]
Li [94]
This work
Switching-on resistance
finite
zero
zero
finite
Drain inductance
infinite
infinite
finite
finite
To make the comparison fair, we employed the same devices and set the same design
specifications of Pout = 0.25 W and fc = 1.9 GHz. Since both Ewing and Sokal assumed
an infinite choke inductance, their designs have one degree less freedom than Lis and ours.
To achieve the design specifications and make the comparison possible, VDD was varied for
28
Ewings and Sokals designs and was fixed as 2 V for Lis and our designs.
Figure II.5 shows the simulated output power and drain efficiency versus the device
width by the four design approaches. As can be seen, Ewings approach has good output
power performance but poor drain efficiency, while both Sokals and Lis works achieve
good efficiency but predict poor output power. Our design technique, however, not only
achieves the designed output powers, but also obtains the optimized drain efficiency.
0.3
Design goal
OUTPUT POWER (W)
0.25
0.2
0.15
Ewing [64]
Sokal [75]
Li [94]
This work
0.1
0.05
2.5
3.5
4.5
(a)
100
80
60
Ewing [64]
Sokal [75]
Li [94]
This work
40
2.5
3.5
4.5
(b)
Figure II.5: Simulated (a) output power and (b) drain efficiency versus NMOS width for
the design approaches developed by Ewing, Sokal, Li, and this work.
29
II.5 Conclusions
An improved design technique is developed to derive the optimum performance of a
CMOS class-E power amplifier. Compared with other theoretical approaches, this design
approach models not only the finite drain inductance, but also the switching-on resistance
of the transistor, thus it leads to a more optimized design. With this design technique,
optimum circuit parameters, as well as the voltage and current waveforms, are derived and
numerically computed.
The design algorithm we developed is applicable not only for bulk MOS devices, but
also for other active devices, such as bipolar transistors, as long as they are operated as
switches.
The disadvantage of this technique is the analytical complexity rising from the inclusion of both the finite choke inductance and the finite switching-on resistance. Although
the analysis leads to more accurate and optimized designs, it does not provide intuitive and
straightforward expressions.
Chapter III
Linear CMOS Class-AB Power
Amplifiers
III.1 Introduction
As described in the first chapter, to meet the simultaneous requirements of high linearity and reasonable efficiency, power amplifiers in non-constant-envelope systems are
often operated in a class-AB mode. Although more linear than a class-B or higher amplifier, the intrinsic linearity obtained in class-AB operation is often still insufficient to meet
required specifications. This is especially true if a MOS device is employed because the
low transconductance associated with the MOS device requires a relatively large input voltage signal, and since the third-order nonlinearity (e.g., IM3) is directly proportional to the
cube of the input signal amplitude, this large signal amplitude will yield significant nonlinearity at the output. While many external linearization techniques are known [12], they
are complex and inconvenient for handset applications, and it is thus important that the
intrinsic amplifier linearity be made as high as possible. In this chapter, it is shown that the
gate-source capacitance of a MOS device is a major source of nonlinearity that can limit
the performance of a CMOS class-AB power amplifier. A simple technique to compensate
30
31
the nonlinearity is suggested, and simulations and experiments on a prototype amplifier are
used to demonstrate its effectiveness.
This chapter will begin with a description of distortion effects of the gate-source
capacitance. Then a capacitance compensation technique will be introduced, followed by
the verification of this technique using Volterra analysis. The detailed schematic and layout
designs will be presented, along with the implementation issues and experimental results
of the prototype power amplifiers. Finally, the conclusions will be summarized.
32
elements could be absorbed into I and O). These simplifications are justified, since the purpose of the model is merely to illustrate the main sources of nonlinearity under class-AB
operation. For accurate simulation results needed in final designs, however, it should be
noted that radio-frequency (RF) MOS models should include the omitted elements [17]
[21].
Cgdn
is
Cgbn
Cgsn
i dsn
RL
RL
s
(a)
Cgdn
is
Cgbn
Cgsn
i dsn
Cgdp
Cgbp
Cgsp
i dsp
(b)
Figure III.1: Simplified models of CMOS class-AB power amplifiers. Part (a) shows an
NMOS device working alone, and part (b) shows an NMOS device along with a PMOS
device used to provide a compensating input capacitance.
33
CAPACITANCE (pF)
16
12
0.5
1.5
Figure III.2: Plots of the simulated NMOS device capacitances as a function of gate-source
voltage, for a fixed drain-source voltage of 3.3 V. The device length and width are 0.5 m
and 3 mm, respectively, and the device threshold voltage is VTn = 0.66 V.
from IBMs SiGe5AM technology, and the plots were obtained using SPECTRE circuit
simulator and the associated commercial MOS model released by IBM; the model employs BSIM3v3.2 as an intrinsic subcircuit, along with extrinsic parasitics to account for
RF effects [22, p. 53].
Figure III.2 confirms that the total capacitance seen looking into the gate, as found
34
from an ac simulation at each gate-source voltage, Cggn Im {y11 }/, where y11 is the
short-circuit, common-source input admittance and = 2(2 GHz) is the radian frequency, is equal to the sum of the individual capacitance components mentioned earlier:
Cggn = Cgsn + Cgbn + Cgdn . This is to be expected when the devices parasitic resistances
are negligible [19, eq. (9)], and helps to validate the simplified model of Fig. III.1(a). More
importantly, Fig. III.2 shows that while Cgdn and Cgbn are relatively constant, Cgsn varies
substantially as the device transits from an off (below threshold) to an on (above threshold) state. While Cgsn as plotted includes both intrinsic and extrinsic parts, almost all of this
variation can be traced to a change in the intrinsic part [19, Fig. 3(a)]. This variation is particularly germane for class-AB operation, because the transition in the capacitance occurs
at the devices threshold voltage, close to where it is typically biased. As will be shown,
the change in capacitance leads to substantial distortion at the gate, and can subsequently
limit overall amplifier linearity.
35
V GG
is
V DD
Input
matching
network
Output
matching
network
RL
Output
matching
network
RL
(a)
V GG
is
V DD
Input
matching
network
V PP
(b)
Figure III.3: Simplified schematics of class-AB amplifiers used to illustrate the impact of
the gate-source capacitance on linearity. The basic amplifier is in (a), and the linearized version is in (b). The NMOS and PMOS devices are the same as those in Figs. III.2 and III.6,
respectively.
36
and output matching networks include short-circuit terminations at the harmonic frequencies, which we found helped overall linearity1 ; they also helped to boost the fundamental
output power [23, p. 384]. The input network includes the source admittance, chosen in
this case to represent the output admittance of a driving class-A stage. In fact, the circuits
in Fig. III.3 are simplified versions of actual two-stage, class-AB amplifiers that were built
and tested, and which will be described in Section III.6.
Figures III.4 and III.5 show SPECTRE simulations of the third-order, intermodulation distortion (IM3) at 21 2 for a two-tone input at frequencies 1 = 2(1.96 GHz)
and 2 = 2(1.94 GHz), at the gate and drain, respectively; note that the drain IM3 is
equivalent to the load IM3, since O and RL are linear and 21 2 1 .
As shown, the basic amplifier of Fig. III.3(a) incurs substantial distortion at both
the gate and drain; it will be proven in Section III.3 B that most of this distortion is due
to the change in gate-source capacitance as the device turns on and off during class-AB
operation. On the other hand, Figs. III.4 and III.5 show that much better performance can
be obtained by employing the scheme illustrated in Fig. III.3(b), where a compensating
nonlinear capacitance is added at the input.
The details are described in the out-of-band termination part of Section III.4.
37
VGG = 0.75V
VGG = 0.80V
20
basic
40
60
linearized
SPECTRE (basic)
SPECTRE (linearized)
Volterra (basic)
Volterra (linearized)
80
0
10
20
20
basic
40
linearized
60
80
30
linearized
80
10
20
30
VGG = 0.90V
40
30
20
basic
60
20
VGG = 0.85V
20
10
basic
40
60
linearized
80
0
10
20
30
38
VGG = 0.75V
VGG = 0.80V
20
basic
40
linearized
60
SPECTRE (basic)
SPECTRE (linearized)
Volterra (basic)
Volterra (linearized)
80
0
10
20
20
basic
40
linearized
60
80
30
linearized
80
10
20
30
VGG = 0.90V
40
30
20
basic
60
20
VGG = 0.85V
20
10
basic
40
linearized
60
80
0
10
20
30
39
CAPACITANCE (pF)
12
Cgsp+Cgbp+Cgdp
C
gsp
Cgbp
Cgdp
0.5
0.5
Figure III.6: Plots of the device capacitances of a PMOS transistor as a function of its gatesource voltage, with its drain-source voltage held at zero. The device length and width are
0.5 m and 2 mm, respectively, and the device threshold voltage is VTp = 0.49 V.
As can be seen, while Cgbp is relatively constant, Cgdp and Cgsp change2 from a high
to a low value as the device transits from an on to an off state. This behavior is exactly
complementary to that of Cgsn in Fig. III.2. Therefore, it should be possible to linearize
or compensate Cgsn with the aid of a PMOS device. The basic idea is simply to place a
PMOS device alongside the NMOS device as illustrated in Fig. III.3(b); the model for the
situation is shown in Fig. III.1(b). When the PMOS device is properly biased and sized,
2
Since the drain-source voltage is zero, Cgdp should equal Cgsp ; the small discrepancy occurs due to an
implementation limit in BSIM3v3 [24, Ch. 4].
40
the total capacitance Cggn + Cggp seen at the NMOS gate will be a constant, which reduces
the distortion generated at the gate, and subsequently at the drain.
Since the change in the NMOS and PMOS capacitances occurs at their respective
threshold voltages, it is clear that the PMOS bias voltage VPP in Fig. III.3(b) should be
VPP = VTn VTp .
(III.1)
Neglecting Cgbn and Cgbp and extrinsic contributions to the capacitances, an appropriate
figure for the sizing of the PMOS device can be obtained by noting that the NMOS device
switches between weak and strong inversion, and the PMOS device works in the triode
region. Therefore [2, Sec. 8.3.2], the changes in NMOS and PMOS capacitances are approximately
2
Cggn Cgsn Wn Ln Cox n
3
(III.2)
and
Cggp
Wp Lp Cox p
(Cgsp + Cgdp ) 2
= Wp Lp Cox p
2
(III.3)
where Wn and Ln , and Wp and Lp , are the widths and lengths of the NMOS and PMOS devices, and Cox n and Cox p are their oxide capacitances, respectively. Assuming the changes
in the capacitances are abrupt, we then require
Cggn
2 Wn Ln Cox n
1
Cggp
3 Wp Lp Cox p
which can be used as a guide to size the PMOS device.
(III.4)
41
Cggn+Cggp
C
ggn
Cggp
CAPACITANCE (pF)
20
16
12
0.5
1.5
Figure III.7: Plots of simulated Cggn , Cggp , and the sum Cggn + Cggp for the NMOS and
PMOS devices of Figs. III.2 and III.6.
Figure III.7 shows plots of Cggn and Cggp , found from Im {y11 }/, and of the sum
Cggn + Cggp , for the NMOS and PMOS devices of Figs. III.2 and III.6. As shown, while
both Cggn and Cggp vary with the NMOS gate-source voltage, the sum Cggn + Cggp remains
roughly constant. The small ripple that occurs in the sum at the transition point arises
because the capacitances do not change abruptly; the slope of the Cggn curve is not exactly
equal (in magnitude) to that of the Cggp curve. The ripple can be minimized by adjusting
the bias and size of the PMOS device from the nominal values given by (III.1) and (III.4).
The impact of linearizing or compensating the input capacitance can be understood
with the aid of Volterra analysis.
42
only on the circuits bias point. Such analysis cannot be used to describe a highly nonlinear
circuit, such as a class-AB power amplifier. However, we will attempt to alleviate this problem by employing power-series expansions of order greater than three, and by allowing the
series coefficients to depend on both the bias point and the RF signal power.
(III.5)
(III.6)
and
At each bias point, the RF signal power determines the range of excursion of the NMOS
gate-source voltage; for simplicity, this range can be approximated to be the peak-to-peak
excursion of the two-tone envelope (i.e., the envelope arising from the fundamental signal
components at 1 and 2 , and neglecting the much smaller harmonic and intermodulation
components). With knowledge from SPECTRE of the behavior of the individual components of Ceff versus this voltage, Ceff can then be modelled as a power series. We found that
a fifth-order power series would work well for all bias points and for all RF signal powers
43
considered, i.e., Ceff could always be written as follows:
2
3
4
Ceff = c1 + c2 vgs + c3 vgs
+ c4 vgs
+ c5 vgs
.
(III.7)
It is important to emphasize that when the bias point or RF signal power changes, the
coefficients c1 through c5 also change, such that the expansion in (III.7) always traces out
the appropriate Ceff versus vgs curve.
The behavior of the large-signal, quasi-static, drain-source current iDSN (vGS , vDS ) for
the NMOS transistor as a function of vGS and vDS can be simulated with SPECTRE, and
the results can be used to expand the corresponding signal current idsn in Figs. III.1(a)
and III.1(b) as a power series. In performing the expansion, for simplicity, the dependence
on the drain-source voltage is first eliminated. Referring to Figs. III.3(a) and III.3(b), this
is done by approximating vDS to be a superposition of the dc bias and the purely linear part
of the output signal:
vDS VDD gm vgs RO
(III.8)
44
of points traced out by iDSN (VGG + vgs , VDD gm vgs RO ) can be used to find a power series
for idsn in terms of vgs . In this case, we found a series of order three sufficed, i.e., idsn could
be written as follows:
2
3
idsn = g1 vgs + g2 vgs
+ g3 vgs
(III.9)
where, as before, the coefficients g1 through g3 change with both the bias point and the RF
signal power, such that (III.9) always traces out the appropriate idsn versus vgs curve.
Figure III.8 shows the SPECTRE simulated and MATLAB fitted curves for Ceff and
idsn as functions of the NMOS gate-source voltage. The fitted curves shown are for the gate
bias of VGG = 0.8 V, and the input voltage amplitude vgs of 0.2 and 0.6 V, respectively. As
can be seen, the third-order current and fifth-order capacitance polynomials can fit idsn and
the compensated Ceff very well at all signal levels we are interested; for the uncompensated
Ceff , however, the fifth-order polynomial can fit well only at low power levels. This is not
surprising considering the strong nonlinear relationship between the uncompensated Ceff
and the NMOS gate-source voltage. It can be shown that a higher order polynomial will
yield a better fit but a much more complicated analysis. Thus, the choice of a fifth-order
polynomial fit for the capacitance is a compromise between accuracy and complexity.
45
20
SPECTRE data
Curvefit (vgs=0.2 V)
Curvefit (vgs=0.6 V)
16
12
0
0
0.2
0.4
0.6
0.8
1.2
1.4
(a)
0.6
SPECTRE data
Curvefit (vgs=0.2 V)
Curvefit (vgs=0.6 V)
0.4
0.2
0
0
0.2
0.4
0.6
0.8
1.2
1.4
1.6
(b)
Figure III.8: SPECTRE simulated and MATLAB fitted curves for (a) Ceff and (b) idsn as
functions of the NMOS gate-source voltage. The fitted curves shown are for the gate bias
of VGG = 0.8 V, and the input voltage amplitude vgs of 0.2 and 0.6 V, respectively.
46
iC
is
Ys
vC
definition of capacitance,
C=
dQ
dQ dt
dt
=
= iC
dvC
dt dvC
dvC
(III.10)
Substituting (III.7) into (III.10) and rearranging gives the current in the capacitor as
iC = c1
dvC c2 dvC2
c3 dvC3
+
+
.
dt
2 dt
3 dt
(III.11)
Here, for illustration purposes, only the first three terms in (III.7) were used. Let vC be
vC = H1 (ja ) is + H2 (ja , jb ) i2s + H3 (ja , jb , jc ) i3s .
(III.12)
(III.13)
Substituting (III.11) and (III.12) into (III.13) and equating the first-order terms gives
is = Ys (ja )H1 (ja ) is + ja c1 H1 (ja ) is .
(III.14)
47
Rearranging (III.14) yields
H1 (ja ) =
1
.
ja c1 + Ys (ja )
(III.15)
The same procedure can be applied for the second and third order terms, and we obtain
H2 (ja , jb ) =
(III.16)
and
H3 (ja , jb , jc ) =
j(a + b + c )
3[j(a + b + c )c1 + Ys (ja + jb + jc )]
[c3 H1 (ja )H1 (jb )H1 (jc ) + 3c2 H1 (ja )H2 (jb , jc )]
(III.17)
where
1
H1 (ja )H2 (jb , jc ) = [H1 (ja )H2 (jb , jc ) + H1 (jb )H2 (ja , jc )
3
+ H1 (jc )H2 (ja , jb )].
(III.18)
The IM3 of vC at 22 1 is
IM3 =
(III.19)
Since the tone spacing 2 1 is generally much smaller than 1 and 2 , let = 2 1
and 2 1 . Then (III.17) reduces to
H3 (j2 ,j2 , j1 ) =
j
3[jc1 + Ys (j)]
[c3 H12 (j)H1 (j) + c2 (2H1 (j)H2 (j) + H1 (j)H2 (j2))] (III.20)
48
where
jc2 |H1 (j)|2
2[jc1 + Ys (j)]
j2c2 H1 (j)2
H2 (j2) =
.
2[j2c1 + Ys (j2)]
H2 (j) =
(III.21a)
(III.21b)
Special attention should be paid to the terms in the second bracket of (III.20). The
first term comes from the intrinsic third-order nonlinearity c3 ; the second term comes
from the second-order nonlinearity c2 , which yields third-order products by first generating second-order products and then mixing them with the fundamental signals. (III.21)
shows that H2 (j) and H2 (j2) are greatly influenced by the source conductance Ys at
the sub- and second harmonic frequencies. For example, H2 (j) and H2 (j2) can be
set to zero by letting Ys (j) and Ys (j2) be infinity, which is equivalent to shorting the
impedance at and 2.
The same conclusions can be drawn for the simplified nonlinear model of the PA
output stage shown in Fig. III.10. Note that after compensation, Ceff can be approximated
as a linear capacitor, which has no second-order nonlinearity. However, the sub- and second
harmonics can still appear at vgs through the Cgdn feedback.
The impact of out-of-band (in particular, the sub- and second-harmonic) impedances
on circuits linearity are also described in [25][27]. For weakly nonlinear circuits like
LNA, Volterra analysis can be applied to derive the output linearity as a function of the
out-of-band impedances, and it was shown that if the out-of-band terminations are properly
chosen, the circuits linearity can be improved dramatically [27]. However, this optimization technique is not applicable for strong nonlinear circuits such as class-AB PAs, because
49
Cgdn
i ds1
ZI
v gs
Ceff
i dsn
Cdsn
ZL
the strong nonlinearity sources associated with the class-AB operation do not generally
have constant second and third order coefficients for the entire signal range, as exemplified in Fig. III.5. Optimizing the linearity at one signal level could worsen the linearity at
other signal levels. In addition, varying the input second harmonic of a strong nonlinear
source can also influence the generation of its intrinsic third-order term, thus making the
optimization untractable.
However, leaving the out-of-band impedances unattended is not a good strategy either. Our calculations and simulations for the PA output stage show that while the subharmonic impedances of the initially designed input and output matching networks have
slight effects on the load linearity, the second-harmonic impedance of the output matching network can deteriorate the load IM3 4 5 dB for a wide signal range. Thus, on-chip
second-harmonic short circuits are included in the final amplifier.
50
IM3 Calculation
With the power series in (III.7) and (III.9) established and the out-of-band short circuitry applied, the circuit for the Volterra calculation, based on the method of nonlinear
currents [23, pp. 190-207], is shown in Fig. III.11. Here, ZI represents the impedance
Cgdn
+
ZI
~
v gs, 21 2
c1
Ceff, 21 2
~
g1 v
gs, 21 2
dsn, 21 2
ZO
seen looking into the input matching network from the NMOS gate, and ZO represents the
impedance seen looking into the output matching network from the NMOS drain. Since ZI
presents a short circuit at even-order frequencies (see Section III.2), the distortion currents
generated by idsn and Ceff have the following phasor amplitudes:
3 2
dsn,21 2 = g3 vgs,
v
1 gs,2
4
(III.22)
and
Ceff ,21 2
1 2
1 3
2
2
= j(21 2 ) c3 vgs,1 vgs,2 + c5 2
vgs,1 vgs,1 vgs,2 + 3
vgs,1 vgs,2 vgs,2
4
8
(III.23)
where vgs,1 and vgs,2 are the phasor amplitudes of the gate-source voltage at the fundamental frequencies, and denotes complex conjugation. The distortion voltages that
51
result at the gate and drain can then be computed using the circuit of Fig. III.11:
vgs,21 2 =
vds,21 2 =
ZO {dsn,21 2 [1 + j(21 2 )Cgdn ZI0 ] Ceff ,21 2 [g1 j(21 2 )Cgdn ]ZI0 }
1 + j(21 2 )Cgdn (ZI0 + ZO + g1 ZI0 ZO )
(III.25)
where ZI0 ZI k c1 , and the impedances ZI0 and ZO should be evaluated at the intermodulation frequency 21 2 . The drain voltage at the fundamental frequency is also easily
found to be
vds,1 =
g1 ZO + j1 Cgdn ZO
vgs,1
1 + j1 Cgdn ZO
(III.26)
where, in this case, ZO should be evaluated at the fundamental frequency 1 . The IM3 at
the gate and drain are then simply
vgs,21 2
IM3G = 20 log
vgs,1
(III.27)
vds,21 2
.
IM3D = 20 log
vds,1
(III.28)
and
52
expressions are able to predict the main trends in IM3 as a function of both bias and power
level. Of course, since the power-series coefficients in (III.7) and (III.9), and the values of
vgs,1 vgs,2 , were all found from SPECTRE, this agreement may not be too surprising.
However, the real utility of the Volterra expressions lies in their ability to isolate the impact
of the individual nonlinearities.
Figure III.12 shows the contributions to the drain IM3 arising from the Ceff and idsn
nonlinearities, as computed from (III.25), (III.26), and (III.28). The contribution from Ceff
is found by setting dsn,21 2 0 in the expressions, and the contribution from idsn is
found by setting Ceff ,21 2 0. The Ceff contributions are shown for both the basic and
linearized amplifiers; the idsn contributions do not change, so only one curve is shown.
As illustrated, in the basic amplifier, the Ceff nonlinearity limits the drain IM3 over
most power levels; only at very high power levels does the idsn nonlinearity become important, which is simply a result of increased clipping in class-AB mode. On the other
hand, in the linearized amplifier, the impact of the Ceff nonlinearity is greatly reduced,
and correspondingly, except at high power levels where the idsn nonlinearity dominates, the
compensation scheme leads to the improved performance originally seen in Fig. III.5. Similar analysis could be undertaken and comments made for the gate IM3 in Fig. III.4. (Again,
there is no improvement at very high power levels due to the idsn nonlinearity, which can
impact the gate IM3 by way of feedback through Cgdn .)
53
VGG = 0.75V
VGG = 0.80V
40
20
C eff
(basic)
i dsn
60
C eff
(linearized)
80
0
10
20
20
C eff
(basic)
40
i dsn
60
C eff
(linearized)
80
30
20
30
VGG = 0.85V
VGG = 0.90V
20
C eff
(basic)
40
i dsn
60
C eff
(linearized)
80
0
10
20
30
20
10
C eff
(basic)
40
i dsn
60
C eff
(linearized)
80
0
10
20
30
Figure III.12: Calculated contributions to the drain IM3 from the Ceff and idsn nonlinearities
for both the basic and linearized amplifiers in Figs. III.3(a) and III.3(b), respectively. The
values are computed from the Volterra expressions (III.22)(III.28), as described in the text.
54
VDD
V GG0
VDD
VGG1
RF
choke
RF
choke
Rs
Vs
Input
matching
network
M1
Interstage
matching
network
M0
Output
matching
network
R50
V PP
Mp
Driver stage
Output stage
Figure III.13: Simplified block diagram of designed two-stage CMOS class-AB power
amplifiers.
This section will begin with the design of the output and driver stages. Then the
influence of out-of-band impedances on the amplifier linearity will be discussed, followed
by the study of the impact of ground connections on the amplifier gain and stability. Finally, the schematic of a fully matched two-stage CMOS class-AB power amplifier will be
55
presented.
2
vds0
2Pout
(III.29)
56
VDD
VGG0
v out
v d0
v g0
M0
ZL
Output
matching
network
R50
V PP
Mp
(a)
Output
matching network
v g0
v d0
v out
gm0eqv vgs0
Cg0tot
Rds0
RL
R50
Cd0tot
(b)
Figure III.14: Output stage. (a) Schematic. (b) Simplified linear model for first-order
analysis. The Miller effects of Cgd0 are included in Cg0tot and Cd0tot .
ID
Saturation Region
Vg0max
VGG
Vd0min
VDD
VDS
57
where Pout is the output power delivered to the load and vds0 is the amplitude of the drainsource voltage signal. vds0 is related with the supply voltage VDD by
vds0 = VDD Vds0min .
(III.30)
Here, Vds0min is the minimum drain voltage. Since Vds0min is on the boundary between the
triode and saturation region, as shown in Fig. III.15, we have
Vds0min = Vgs0max VTn
= VGG0 VTn + vgs0
(III.31)
where VGG0 is the gate bias voltage and vgs0 is the amplitude of the gate voltage signal.
Substituting (III.31) and (III.30) into (III.29) gives
RL =
(III.32)
In the actual design, Vds0min is chosen to be slightly larger than (III.31) to include a
small margin voltage Vm . In other words,
Vds0min = VGG0 VTn + vgs0 + Vm .
(III.33)
(III.34)
V = VGG0 VTn + Vm .
(III.35)
where
58
Choice of Device Width
It can be shown as follows that the choice of device width, W0 , is equivalent to the
choice of vgs0 . As illustrated in Fig. III.14 (b), the equivalent transconductance of the output
stage is defined as
gm0eqv =
vds0
.
vgs0 RL
(III.36)
Here, Rds0 is much larger than RL , thus ignored. For a general class-AB operation, gm0eqv is
a complicated function of the gate bias, VGG0 , and the signal amplitude, vgs0 ; but for an ideal
class-B operation, as illustrated in Fig. III.16 (a), gm0eqv is one half of the transconductance
because the signal conducts exactly a half period. Thus, we have
gm0eqv =
Id
gm0
.
2
(III.37)
Id
VGG
VT
0
VGS
vgs
(a)
VGS
vgs
(b)
Figure III.16: Plots of Id versus VGS for (a) an ideal class-B operation, and (b) a shortchannel device biased near the threshold voltage.
59
For hand calculation purposes, the short-channel device that is biased near the threshold voltage can be approximately treated as the ideal class-B case, as shown in Fig. III.16 (b).
When the gate voltage signal is large enough, the carrier velocity is saturated and the
transconductance of a short-channel MOS device is
gm0 = W0 Cox vscl
(III.38)
where W0 is the device width, Cox is the oxide capacitance per unit area, and vscl is a
constant called the scattering-limited velocity [3]. Substituting (III.36) and (III.38) into
(III.37) and rearranging gives
W0 =
2vds0
.
Cox vscl vgs0 RL
(III.39)
4Pout
Cox vscl vgs0 (VDD vgs0 V )
(III.40)
Here, Cox and vscl are constants; Pout and VDD are fixed by the design specifications; V is
a function of VGG0 and defined in (III.35). Thus, (III.40) shows that if the gate bias VGG0 is
fixed, W0 is only a function of vgs0 , proving our claim that the choice of W0 is equivalent
to the choice of vgs0 .
It is worth mentioning that (III.40) is highly simplified and only for hand calculations.
The actual design should take all non-ideal effects into account and use simulations for final
verifications.
The choice of vgs0 (or W0 ) involves a variety of tradeoffs. With respect to linearity,
small vgs0 is preferred because the third-order nonlinearity is proportional to the cube of
60
input signal amplitude. But as shown in (III.40), small vgs0 corresponds to a large device
width, which causes the increase of all the device parasitics, thus making the design of the
matching networks more challenging. Large vgs0 can alleviate the parasitic problems, but
will deteriorate the linearity. In our design, we found that the choice of a vgs0 of 0.6 V and
device width of 6 mm is a good compromise among all these tradeoffs.
Rp kRLp
1
=
1
1
Rs
+
Rs
Rp QL L
1
1
1
+
2
Qt0 QL Qt
(III.41)
where Qt0 is the total Q of the network when the inductor is lossless, i.e.,
r
Qt0 = Qt |QL =
Rp
.
Rs
(III.42)
61
L
Rp
Rs
RLs
(a)
C
Rs
RLp
Rp
(b)
Figure III.17: High-pass, L-match network. (a) Schematic. (b) Equivalent circuit.
62
Rearranging and solving (III.41) gives
Qt =
Q2t0 +
p
Q4t0 + 4Q2L Q2t0
.
2QL
Qt0
Qt Qt0 1
.
2QL
(III.43)
(III.44)
Define loss as the ratio of the power dissipated in the on-chip inductor to the power
delivered to Rp , i.e.,
loss =
Rp
Rp
Rp
Q2t0
=
=
.
=
RLp
QL L
QL Qt Rs
QL Qt
(III.45)
Qt0
t0
=
Qt0
QL
QL 1
2QL
Qt0
1+
.
2QL
(III.46)
Again, we assume Qt0 2QL . For a matching network of Qt0 = 3 and QL of 10, (III.46)
gives 0.35, which implies that the power loss in the on-chip inductor is approximately 35 %
of the power dissipated at the load. The same conclusion can be obtained for a low-pass,
L-match network as well.
Note that if two matching networks are cascaded together, as shown in Fig. III.18,
the total power loss ratio is
loss, cascade =
PL1 + PL2
PL1 PL2
+
= loss, 1 + loss, 2
PRp
PRi
PRp
(III.47)
Here, we assume that PRi PRp , and Q2t0,1 and Q2t0,1 are much larger than one. The total
Qt0 for a cascade structure is
r
Qt0 =
Rp
=
Rs
Rp
Ri
Ri
= Qt0,1 Qt0,2 .
Rs
(III.48)
63
Lossy
L-match
network
1
Rs
Lossy
L-match
network
2
Rp
1
=
QL
Qt0
1
Q2t0
2
Qt0,1 +
+
Qt0,1 + 2
.
Qt0,1
2Q2L
Qt0,1
(III.49)
Qt0
(III.50)
and the resulting minimum total power loss ratio of a cascaded structure is
Qt0
2 Qt0
min(loss, cascade ) =
1+
.
QL
2QL
(III.51)
The same approach can be applied to a cascade structure of more than two matching networks, and the same conclusion can be drawn, as long as the Q2t0 for each stage is much
larger than one. Comparing (III.51) with (III.46), we conclude that a cascade structure of
two on-chip matching networks can reduce the total inductor loss when the total Qt0 is
larger than four.
Let us return to the design of output matching network. The output impedance of
M0 is approximately Rds0 in parallel with Cd0tot , as shown in Fig. III.14 (b). The output
64
RL
Cd0tot
RL
R50
Cd0tot
(a)
R50
(b)
Cb0
RL
Lo1
RLo1s
L0
Cd0tot
Co2
RL0s
R50
C0
(c)
Figure III.19: Output matching networks. (a) High-pass L-match. (b) Low-pass L-match.
(c) Actual implementation.
65
matching network is required to match 50 to an impedance of RL in parallel with an
inductive impedance (to cancel Cd0tot ). In this case, RL = 8 and Cd0tot = 9.6 pF, so the
load impedance to be matched is RL k( sC1d0tot ), which results 4.3 + 4j.
Since the total Qt0 of the output matching network is less than 4, there is no benefits
to cascade more than one L-match networks. If the on-chip inductors have a constant QL of
10, the matching topologies in Fig. III.19 (a) and (b) yield power loss ratios (with respect
to the load power) of 38 % and 56 %, respectively, which are close to the 38 % predicted
by (III.46). The final output matching network is chosen as Fig. III.19 (c), where Lo1 and
Co1 match 50 to 8 , and L0 and C0 cancels Cd0tot . The power and efficiency loss ratio
associated with this matching network is approximately 36 %, slightly better than the 38 %
of the matching network in Fig. III.19 (a).
66
this overdrive can enforce M1 into nonlinear regions and degrade the linearity of the PA.
The SiGe5AM design guide recommends a minimum distance of 80 m between any on-chip inductors
and adjacent conductors, thus a large amount of chip area is consumed.
67
Cb1
Rg0tot
L1
g m1 vgs1
Rds1
Cd1tot
Cg0tot
RL1s
Lossy interstage
matching network
Driver stage
Output stage
(a)
g m1v gs1
Rds1
R L1p
L1
C tot
QC R g0tot
(b)
Figure III.20: Interstage matching network. (a) Circuit implementation. (b) Equivalent
model. Here, Cd1tot is ignored, and Ctot represents the total capacitance of Cb1 in series with
Cg0tot .
68
Our goal is to find the optimum values of L1 and Cb1 to achieve the maximum power
transfer to Rg0tot under the constraint of a finite QL of L1 . Let Ctot represent the total
capacitance of Cb1 in series with Cg0tot ,
Ctot =
1
1
+
Cb1 Cg0tot
1
(III.52)
QC =
Rg0tot Ctot
(III.53)
Assuming that QL and QC are much larger than one and Cd1tot is much smaller than Cg0tot ,
thus ignored, Fig. III.20 (a) can then be simplified to (b), where
RL1p = QL L1 .
(III.54)
Rds1 + QL L1 + Q2 Rg0tot
C
2
2
gm1
vgs1
+j
. (III.55)
2
1
1 2
Q Rg0tot
QC Rg0tot L1 C
QC Rg0tot
.
(III.56)
QC
1
1
(III.57)
69
Then let the derivative of the denominator in (III.55) with respect to QC be zero, the optimum QC is calculated as
s
QC =
Rds1
.
Rg0tot
(III.58)
Note that (III.58) gives us the same conclusion as in the lossless matching case. Cb1 and
L1 can then be calculated from (III.53) and (III.56), respectively.
If Cd1tot is not ignored, it can be shown that (III.53) and (III.56) will be changed to
QC Rg0tot
(1 + QC Rg0tot Cd1tot )
v
u
1
u
QC = u
1
Cd1tot
t
Rg0tot
+
Rds1
QL
L1
(III.59)
(III.60)
The estimated model parameters of the driver and output stages are shown in Table III.1, where M1 is biased at VGG1 = 0.9 V and has a width of 3 mm. If QL is 15,
(III.59) and (III.60) gives L1 = 0.26 nH and Cb1 = 300 pF. Due to the parasitic inductance
of large on-chip capacitors, at 1.95 GHz, the maximum allowable on-chip capacitor (50 pF
calculated by size) exhibits the same impedance as an ideal 200 pF, thus used for Cb1 . L1
is implemented using a microstrip line, which does provide a Q of 15.
Driver stage
gm1 ( ) Rds1 () Cd1tot (pF)
0.28
76
4.2
1
Output stage
Rg0tot () Cg0tot (pF)
0.25
22.6
70
Impact on Gain
It is illustrative to examine the impact of the ground bondwire inductor on the gain of
output stage in Fig. III.21. When biased at VGG0 = 0.85 V, M0 has the estimated device parameters of gm0 = 0.43, Rg0tot = 0.25, Cg0tot = 22.6 pF, and Cd0tot = 9.6 pF. At 1.95 GHz,
the transconductance and input and output impedances of M0 are
gm0 = 0.43
zg0 = Rg0 +
zd0 =
(III.61a)
1
= 0.25 3.7 j
jCg0
1
= 8.5 j.
jCd0tot
(III.61b)
(III.61c)
As implied in [3], the bondwire inductor Ls0 at the source node of M0 behaves as a seriesseries feedback. Assuming Ls0 is 0.1 nH and ignoring the effect of Cgd0 , the Ls0 feedback
71
VDD
VDD
RF
choke
RF
choke
Cb0
L o1
Vout
L1
C f1
Cb2
Co1
M0
R f1
L i1
Vin
M1
C i1
C0
Rb0
C1
On-chip
2f termination
s0
Rb1
VGG0
VGG1
C
Cdc
V PP
s1
On-chip
2f termination
L0
Cb1
L s1
L s0
Mp
Compensation
circuitry
(a)
C f1
R f1
Cb2
Vin
Rs
C i1
L i1
C gd1
Cgs1
Cb1
L1
Cds1
s1
A
On-chip
2f termination
Cgs0
C1
L s1
Cb0 L o1
C gd0
Cds0
L0
Co1
RL
C0
s0
Vout
L s0
On-chip
2f termination
(b)
Figure III.21: Two-stage CMOS class-AB power amplifier for illustrating ground connections. Part (a) shows the schematic, and part (b) shows the simplified linear model.
Here, the bias resistance and channel-length-modulation resistances of the transistors are
not shown.
72
transforms the transconductance and input and output impedances to
0
gm0
gm0
= 0.34 0.18 j
1 + gm0 jLs0
(III.62a)
0
zg0
zg0 (1 + gm0 jLs0 ) = 2.2 3.5 j
(III.62b)
0
zd0
zd0 (1 + gm0 jLs0 ) = 4.5 8.4 j.
(III.62c)
73
VDD
VDD
RF
choke
RF
choke
Cb0
L o1
Vout
L1
C f1
Cb2
L i1
L0
Cb1
Vin
M1
C1
Rb1
C i1
Co1
M0
R f1
C0
Rb0
On-chip
2f termination
VGG0
VGG1
Cdc
s0
On-chip
2f termination
V PP
Mp
L s0
Compensation
circuitry
(a)
VDD
VDD
RF
choke
RF
choke
Cb0
L o1
Vout
L1
C f1
Cb2
L i1
L0
Cb1
R f1
Vin
M1
Rb1
C i1
Co1
M0
C1
C0
Rb0
s0
On-chip
2f termination
VGG0
Cdc
VGG1
On-chip
2f termination
V PP
s1
L s1
L s0
Mp
Compensation
circuitry
(b)
Figure III.22: Two-stage CMOS class-AB PAs for illustrating the impact of ground connections on gain. (a) Ground nodes are connected internally together. (b) Ground nodes
are connected separately.
74
30
25
20
15
10
5
0
0.05
0.1
0.15
0.2
Figure III.23: Power gain of the two-stage CMOS class-AB power amplifiers versus total
ground bondwire inductance for the two ground configurations shown in Fig. III.22 (a)
and (b), respectively.
Fig. III.21 and each bonding wire is 0.5-1 nH, if mutual inductances among the bondwire
inductors are ignored, both Ls0 and Ls1 will be 0.05-0.1 nH, which is consistent with our
0.1 nH estimation.
Impact on Stability
In addition to the impact on gain, ground connections also play an important role in
power amplifier stability. Two techniques gain their popularity in stability analysis. The
first is the root-locus technique, which involves calculation of the poles and zeros of the
amplifier and of their movement in the s plane as the low-frequency, loop-gain magnitude
is changed. This technique is widely used in analog circuit designs in solving feedbackinduced stability problems. At microwave frequencies, however, it is often difficult to
identify the feedback loops that cause the circuit to become unstable. Under this circum-
75
stance, the second technique Stern stability factor K is usually employed. K is defined
as
K=
(III.63)
where = S11 S22 S12 S21 . If K > 1 and < 1, the circuit is unconditionally stable, i.e.,
it does not oscillate with any combination of source and load impedances as long as their
real parts are positive. However, a disadvantage of using K factor is that the S parameters
of the circuit must be calculated (or measured) for a wide frequency range to ensure that
K remains greater than unity at all frequencies. Thus, a great deal of effort is involved,
and most importantly, little insight can be obtained. It is also worth noting that K is a
pessimistic measure of stability since it allows arbitrary source and load impedances.
In our investigation of PA stability, the following criterion [28] is examined. If the
determinant of a linear network contains any zeros in the right half plane (RHP), the
network will be unstable, otherwise the network is stable. This criterion is equivalent
to the pole analysis of a linear network [28].
The ground configurations are divided into two categories: one-chip-ground and twochip-ground, as shown in Fig. III.24. The first is defined as the configurations where s1 and
s0 are joined together before connecting to the external ground; the latter is defined as those
where s1 and s0 are connected independently to the external ground. Table III.2 lists all the
possible ground configurations.
Since oscillations start from noise, the small-signal equivalent models in Fig. III.25
were used for the one-chip-ground and two-chip-ground configurations, respectively. Nodal
76
VDD
VDD
RF
choke
RF
choke
Cb0
L o1
Vout
L1
C f1
Cb2
L i1
R f1
M1
C0
Rb0
C1
On-chip
2f termination
Rb1
C i1
Co1
M0
Vin
VGG0
VGG1
C
Cdc
s0
On-chip
2f termination
L0
Cb1
V PP
Mp
L s0
Compensation
circuitry
(a)
VDD
VDD
RF
choke
RF
choke
Cb0
L o1
Vout
L1
C f1
Cb2
L i1
M1
C1
Rb1
C i1
C0
Rb0
s0
On-chip
2f termination
VGG0
VGG1
On-chip
2f termination
Co1
M0
R f1
Vin
L0
Cb1
C
Cdc
V PP
s1
L s1
L s0
Mp
Compensation
circuitry
(b)
Figure III.24: Two-stage CMOS class-AB power amplifier for (a) one-chip-ground and (b)
two-chip-ground configurations.
77
C f1
R f1
Cb2
Vin
Rs
L i1
C gd1
Cb1
Cgs1
C i1
L1
Cds1
C1
On-chip
2f termination
Cb0 L o1
C gd0
Cgs0
Cds0
L0
Co1
RL
C0
s0
Vout
On-chip
2f termination
L s0
(a)
C f1
R f1
Cb2
Vin
Rs
C i1
L i1
C gd1
Cgs1
Cb1
L1
Cds1
s1
A
On-chip
2f termination
Cgs0
C1
L s1
Cb0 L o1
C gd0
Cds0
L0
Co1
RL
C0
s0
Vout
L s0
On-chip
2f termination
(b)
Figure III.25: Small-signal equivalent model of the two-stage CMOS class-AB power amplifier for (a) one-chip-ground and (b) two-chip-ground configurations.
78
Ground
One-chip-ground Two-chip-ground
configuration
configurations
configurations
index
A B C D
A B C D
0
0 0 0
0
0 0 0
0
1
0 0 0 s0
0 0 0 s0
2
0 0 s0 0
0 0 s0 0
3
0 0 s0 s0
0 0 s0 s0
4
0 s0 0
0
0 s1 0
0
5
0 s0 0 s0
0 s1 0 s0
6
0 s0 s0 0
0 s1 s0 0
7
0 s0 s0 s0
0 s1 s0 s0
8
s0 0 0
0
s1 0 0
0
9
s0 0 0 s0 s1 0 0 s0
10
s0 0 s0 0
s1 0 s0 0
11
s0 0 s0 s0 s1 0 s0 s0
12
s0 s0 0
0
s1 s1 0
0
13
s0 s0 0 s0 s1 s1 0 s0
14
s0 s0 s0 0
s1 s1 s0 0
15
s0 s0 s0 s0 s1 s1 s0 s0
0 represents the external ground.
Table III.2: Ground configurations for the two-stage CMOS class-AB PAs in Fig. III.24.
79
analysis [28] was employed to calculate the determinant of each ground configuration in
Table III.2, and the resulting determinant is a high-order polynomial with coefficients expressed by the circuit parameters, including the total ground bondwire inductance Lstot . It
is apparent that Lstot is equal to Ls0 for one-chip-ground configurations and 12 Ls0 for twochip-ground configurations if we let Ls1 = Ls0 . Then Lstot was swept from zero to 4 nH4
to find its maximum that is capable of keeping all the roots of the determinant polynomial
in the left half s plane. This value, as defined by the stability criterion, sets the upper
limit of the ground bondwire inductance to avoid oscillation. Figure III.26 shows the calculated maximum stable ground bondwire inductance of the two-stage CMOS class-AB
PA for the ground configurations in Table III.2. As can be seen, the stability of the twostage power amplifier is strongly dependent on how the ground nodes were connected: one
unappropriate connection could make a stable PA oscillate. It is also shown that, for our
case, most of one-chip-ground configurations have better stability performance than their
two-chip-ground counterparts.
To verify our analysis, transient simulations based on the schematic in Fig. III.21 (a)
were carried out using SPECTRE. Again, for each ground configuration, the total ground
bondwire inductance was swept to find the value that began to make the transient waveforms unstable. Then it was recorded and compared with the value predicted by the calculation. Less than 10 % difference between the calculated and simulated Lstot were obtained
for all ground configurations, proving the validity of our analysis.
4
The value of 4 nH is arbitrarily chosen. In fact, any value can be chosen as long as it is much larger than
the typical ground bondwire inductance.
80
4
One chip ground
Two chip grounds
3
10
15
Figure III.26: Maximum stable ground bondwire inductance of the two-stage CMOS classAB PA for the ground configurations in Table III.2. The plot does not show the data points
exceeding 4 nH.
81
V DD
V DD
RF
Choke
Onchip
interstage
matching
Cb2
L i1
Cb0
L o1
V out
L1
Cf1
Onchip
input matching
Onchip
output matching
(PA1 and PA2 only)
RF
Choke
L0
Cb1
Co1
M0
Rf1
C0
V in
M1
C1
Rb0
Rb1
V GG0
V GG1
s0
Ci1
Onchip
2f termination
Onchip
2f termination
Cdc
V PP
Mp
L s0
Compensation
circuitry
(PA2 and PA3 only)
Equivalent
bondwire inductance
Figure III.27: Schematic of the fully matched two-stage CMOS class-AB power amplifiers.
82
83
sheet resistances at 25 C.
current limits at 100 C.
84
Drain
Gate
Substrate
Diffusion
Polysilicon
M1 & M2
AM
Contact
Source
Figure III.28: Layout (not scaled) of a basic transistor cell. The fingers are 20 m long.
drain nodes. Figure III.28 shows our layout of a basic transistor cell. First, the two ends
of gate fingers were connected together to reduce the gate resistance by a factor of four.
Second, M1 and M2 were combined to connect both drain and source not only to reduce
their parasitic resistances but also to increase their current handling capabilities. Third, the
substrate were connected to the source for each transistor cell, thus keeping the substrate
voltage equally distributed in the whole transistor. To reduce the parasitic resistance and
capacitance, both the gate and drain were routed through the top metal layer AM.
85
may result in a 10-15 % degradation of the inductors quality factor Q. The disadvantage
associated with this design rule is an enormous waste of chip area. Therefore, it is advised
that inductors should be used as little as possible during the initial design phase.
Since an on-chip inductor has a maximum width of 25 m, it should not be used for
passing through a large amount of current. The initial design should be carried out with
this in mind.
86
Table III.4: Comparison between maximum allowable layout currents and corresponding
maximum designed currents of all critical components in PA2.
Critical
Layout
Design
components Metal ID Metal width Idc
Irms
Idc
Irms
(m)
(A) (A)
(A) (A)
M1&M2
160
0.32 2.22
D
0.26 0.45
AM
75
0.46 0.67
M0
G
AM
40
N/A 0.37 N/A 0.09
S M1&M2
160
0.32 2.22 0.26 0.45
MT
40
N/A 0.22
L0
N/A 0.19
AM
20
N/A 0.21
MT
40
N/A 0.22
Lo1
N/A 0.18
AM
20
N/A 0.21
MT
50
N/A 0.27
Co1
N/A 0.20
AM
20
N/A 0.21
87
VDD
L d0
M0
VDD
Cdb0
M2
PA
Distributed
Substrate
Model
M1
Sensitive
Circuits
Lb
(a)
L d0
PA
p+
n+
M0
M1
n+
n+
Cdb0
Sensitive
Circuits
n+
Cdb1
p - substrate
(b)
Figure III.29: Effect of substrate coupling. (a) Schematic modelling. (b) Sideview of
device layouts.
88
Large Substrate
Guardrings
L s0
PA
p+
Small
Bondwires
M0
n+
M1
n+
p+
Cdb0
p+
n+
Sensitive
Circuits
n+
Cdb1
p - substrate
Deep Trench
Blocks
Figure III.30: Layout structure employing both large substrate guardrings and deep trench
blocks.
in our design. First, large areas of substrate guardrings were used to encompass all the
power transistors since they are the primary sources of substrate noise. Second, multiple
deep trench blocks were placed at the boundary of the power amplifier to further increase
the isolation between the PA and other circuit blocks. The layout structure employing these
two methods is illustrated in Fig. III.30.
89
Figure III.31: Final Layout of the fully integrated and compensated two-stage CMOS PA
(PA2). The components are labelled using the same names as in Fig. III.27.
90
Figure III.32: Die microphotograph of the fully integrated and compensated two-stage
CMOS PA (PA2).
91
Package Choice
The Amkor MicroLeadFrame (MLF) package was chosen primarily for its enhanced
thermal and electrical characteristics. It is a plastic encapsulated and leadless package
where electrical contact to the PCB is made by soldering the lands on the bottom surface of
the package to the PCB. The enhanced thermal and electrical properties of the MLF package is achieved by incorporating an exposed die paddle on the bottom, which efficiently
conducts heat to the PCB and provides a stable ground through down bonds and electrical
connections through conductive die attach material. Figure III.33 shows the photograph
and cross section drawing of the MLF package.
(a)
(b)
Figure III.33: MLF package (a) photograph and (b) cross section drawing.
There is a variety of options in choosing the size and lead numbers of the MLF
package. In order to relax the handling and soldering issues, the final package was chosen
to have a large profile of 6 6 mm2 and a total of 20 leads (5 on each side).
92
Printed Circuit Board Choice
The PA evaluation board utilizes a two-layer RO4350 with a dielectric constant of
3.48 and a dielectric thickness of 20 mil. In addition to good dimensional stability and low
processing and assembly costs, RO4350 provides excellent high-frequency performance
due to its low dielectric loss and stable electrical properties over frequency. The low thermal
coefficient of the dielectric constant of RO4350 also makes it suitable for PA applications.
93
L bw
TL 1
C1
TL 2
TL 3
ZL
L1
R50
(a)
TL 1
TL 2
C1
TL 3
L bw
Zs
Z in
C2
(b)
Figure III.34: Hybrid off-chip matching network for PA3. (a) Output. (b) Input.
94
topology, as shown in Fig. III.34 (a). Here, Lbw models the output bondwire inductance,
TL1, TL2, and TL3 model the transmission line effects of the connection traces. Since
ZL is very small (approximately 4 + 4j), any slight imperfections in the matching network
could influence the value of ZL and consequently degrade the gain and efficiency of the
output stage. It can be shown that, for the output matching network in Fig. III.34, the imaginary part of ZL is most sensitive to the variations of the matching components. Thus, it
is very desirable to design the matching network to be capable of continuously tuning the
imaginary part of ZL . If TL1 and TL2 is short enough, adjusting the length of TL1 or TL2
can achieve the continuous tuning of the imaginary part of ZL , while keeping the real part
of ZL approximately unchanged. Since changing the length of TL1 involves physically
cutting the TL1 trace, tuning the length of TL2 is preferable, and this is accomplished by
designing the impedances of TL2 and TL3 as 50 and sliding L1 along the trace of TL2
and TL3.
The values of the matching components were first calculated in MATHEMATICA and
further verified and tuned using Agilent ADS. Figure III.35 shows the ADS schematic and
simulated ZL of the output matching network for PA3. Due to the large size of the package, the output bonding wire has a length of more than 1.5 mm and exhibits approximately
1.8 nH.
Considering the uncertainty of the bondwire inductance and the variations of the actual values of chip capacitors and inductors, it is necessary to tune the output matching
networks in conjunction with the exhibited testing phenomena. Since the output matching
95
(a)
(b)
Figure III.35: Off-chip output matching network for PA3 in ADS. (a) Schematic. (b) Simulated ZL .
96
is the load-line matching instead of maximum power matching, optimizing the gain (|s21 |)
does not necessarily imply optimized load matching. To still achieve the optimum load
matching, the gain and dc current for various output power levels were observed and compared with the SPECTRE simulations, and corresponding adjustments were made in the
output matching network until good agreement was obtained. Since we only need to tune
the value and position of L1 , few iterations were needed before the optimum output match
was achieved.
After the output matching network was implemented, the input matching network
can be designed by first measuring the input impedance and then matching it with any
matching structure. Figure III.34 (b) shows our input matching network for PA3. Again,
Lbw models the input bondwire inductance, TL1, TL2, and TL3 model the transmission
line effects of the connection traces.
The final application schematic is shown in Fig. III.36, where the 47 F capacitors at
VDD1 and VDD0 are for bypassing the ac signals to ground. This is very important not only
for stability considerations, but also for linearity concerns. As illustrated in Section III.3.2,
the out-of-band (the sub-harmonic frequencies, in this case) impedance can dramatically
impact the PA linearity. As expected, the measurements showed that the inclusion of these
two capacitors can significantly improve both linearity and stability of the PA. Any value
can be chosen for these two capacitors as long as they provide good ac short at the subharmonic frequencies. The photograph of the PCB implementation of PA3 is shown in
Fig. III.37.
97
VDD1
47uF
VDD0
47uF
10nH
10nH
MLF
5.6nH
Die
RF in
2.0pF
1.3pF
RF out
1.3pF
Bias
circuit
1.5nH
98
Power supplies
Agilent 6612C
Spectrum analyzer
Agilent 6612C
Agilent E4440A
Signal generator
Agilent E4438C
VDD1
VDD0
PA
HP E3610A
...
...
HP E3610A
Power supplies
Figure III.38: Test setup for evaluating the PAs. The ground connections of the test equipments and the PA are not shown.
99
35
PA1
PA2
PA3
30
25
25
20
20
15
15
10
10
0
5
10
15
20
25
PAE (%)
GAIN (dB)
30
0
30
Figure III.39: Measured gain and power-added efficiency versus output power of the three
PAs. The output stages of the PAs are all biased at 0.8 V.
100
The measured results were compared with those from SPECTRE simulations and
good agreement was obtained. Figure III.40 shows the simulate and measured gain and
power-added efficiency (PAE) for for the three PAs.
Linearity
To verify their linearity performances, the PAs were tested under various bias and
power levels using both two-tone and WCDMA signals. Figures III.41 show the measured third-order intermodulation, adjacent-channel leakage power (ACP1), and alternatechannel power (ACP2) for the three PAs. As can be seen, the compensated PAs (PA2 and
PA3) have much better linearity than the uncompensated PA (PA1) for various gate biases
and a wide range of output power; in addition, the IM3 measurements show similar trends
as those shown in Fig. III.5 of Section III.2 C.
PA3 achieves an ACP1 of -35 dBc and ACP2 of -55 dBc at a carrier output power of
24 dBm, which is compliant with the 3GPP-WCDMA ACP requirements of -33 dBc and
-43 dBc [29], respectively. Due to the loss of on-chip output matching, PA1 and PA2 can
only meet the WCDMA ACP requirements at output powers of 22 and 23 dBm, respectively. Figure III.42 shows the measured WCDMA spectra of PA1 and PA2 at a carrier
output power of nearly 20 dBm.
It is worth mentioning that all the bias voltages utilized in our measurements are
almost exactly the designed values; in addition, no oscillation was observed during the
entire measurement procedure, even when both the source and load were disconnected.
101
40
40
Simulation
Measurement
35
30
30
25
25
20
20
15
15
10
10
0
5
10
15
20
25
PAE (%)
GAIN (dB)
35
0
30
(a)
40
40
Simulation
Measurement
35
30
30
25
25
20
20
15
15
10
10
0
5
10
15
20
25
PAE (%)
GAIN (dB)
35
0
30
(b)
40
40
Simulation
Measurement
35
30
30
25
25
20
20
15
15
10
10
0
5
10
15
20
25
PAE (%)
GAIN (dB)
35
0
30
(c)
Figure III.40: Simulated and measured gain and power-added efficiency versus output
power for (a) PA1, (b) PA2, and (c) PA3. The output stages of the PAs are biased at 0.8 V.
102
PA1
PA2
PA3
20
30
40
50
5
10
15
20
25
30
25
30
25
30
(a)
20
PA1
PA2
PA3
30
40
50
10
15
20
(b)
40
PA1
PA2
PA3
50
60
70
5
10
15
20
(c)
Figure III.41: Measured (a) IM3 , (b) adjacent-channel leakage power, and (c) alternatechannel power versus peak-envelope output power for the three PAs. The output stages of
the PAs are all biased at 0.8 V.
103
Figure III.42: Measured WCDMA spectra of PA1 and PA2 at a carrier output power of
nearly 20 dBm. The output stages of the PAs are both biased at 0.8 V.
Table III.5 compares the performance of recently reported linear power amplifiers
for handset applications. As can be seen, although a CMOS PAs peak efficiency is generally lower than its GaAs HBT (FET) counterpart, if properly linearized, it can effectively
be used as a low-cost alternative, especially for low-supply voltage and medium-power
applications.
III.7 Summary
The nonlinear gate-source capacitance is a dominant source of distortion that may
limit the linearity of CMOS class-AB power amplifiers. Improved performance can be
obtained by using a compensating nonlinearity, provided by the gate-source capacitance of
an appropriately biased and sized PMOS device placed alongside the NMOS device that
provides the class-AB amplification. Simulations and experiments show that the method
can improve both the two-tone, third-order intermodulation and adjacent-channel leakage
104
Table III.5: Performance comparison of recently reported linear power amplifiers for handset applications.
Ref.
Technology
Pout
PAE
(dBm)
Su 98
CMOS
[30]
0.8 m
Giry 00
CMOS
[31]
0.35 m
Yen 03
CMOS
[32]
0.25 m
This work
CMOS
(PA3)
0.5 m
Vintola 01
AlGaAs/GaAs
[33]
HBT
Jager 02
InGaP/GaAs
[34]
HBT
Srirattana 03
GaAs
[35]
FET
28
33 %
Gain
[Signal]
VDD
Freq.
Operating
(dB)
ACPR @ Pout
(V)
(MHz)
class
[NADC]
836
N/A
35 %
24.6
20
28 %
11.2
[PDC]
AB
(linearized)
2.5
1910
AB
2.5
2450
AB
29 %
23.9
[WCDMA]
(linearized)
3.3
1750
>27 %
>30
[WCDMA]
AB
(linearized)
3.5
1950
AB
N/A
1950
AB
N/A
1950
Doherty
38 %
22.6
29.7
46 %
8.5
[WCDMA]
-37 dBc @ 27 dBm
[WCDMA]
-38 dBc @ 28.6 dBm
3-stage
power by approximately 8 dB. While meeting the 3GPP-WCDMA ACP requirements, the
linearized two-stage amplifier is capable of delivering an output power of 24 dBm with a
small-signal gain of nearly 24 dB and a power-added efficiency of 29 %.
Chapter IV
Dynamic Biasing Technique
IV.1 Introduction
Efficient power amplifiers are highly desirable in mobile wireless communication
systems to prolong battery life. Meanwhile, spectrally efficient modulation schemes in
many wireless standards result in signals with highly time-varying envelopes, thus imposing a stringent linearity requirement on the employed power amplifiers to preserve modulation accuracy and limit spectral regrowth.
To achieve the linearity requirement, PAs are generally operated in class-A or classAB modes. Although class-A and AB power amplifiers have reasonable maximum efficiencies (theoretical 50% for A and 50-78.5% for AB), they suffer significant efficiency
degradation if operated at low power levels. For example, the efficiency of a class-A amplifier is in proportion to the output power, Pout , and this results in a maximum of only
0.5 % when Pout is backed off 20 dB. This efficiency degeneration at low power levels
deserves special attentions, if taking into account the statistical nature of power usage in
wireless communication systems. As exemplified in [36], despite the maximum of 0.5 W,
the output power in a IS-95-CDMA system has the most probable value of only 1 mW,
105
106
yielding an extremely low PA efficiency.
Various techniques [36]-[37] were developed to improve PA efficiency at low power
levels. The dynamic biasing technique described in [37] is most appropriate for IC implementation. This technique uses the envelope of the input signal to dynamically control the
gate dc bias voltage of the power amplifier, thus reducing the current consumption of the
amplifier at low power levels. Since the previously designed two-stage CMOS class-AB
power amplifier exhibits good linearity and maximum efficiency, we would like to explore
the utility of the dynamic biasing technique on our class-AB PA.
This chapter will begin with a brief description of the dynamic biasing technique.
Then it will be followed by the detailed analysis of the envelope detector circuit. The
average efficiency improvement and distortion impact of the technique will be discussed
successively. Finally, the experimental results of a prototype amplifier will be presented
and conclusions will be drawn.
107
schematic in Figure IV.1 (b) is used. Here, the envelope detector is directly connected to
the with the input of the output stage. Since the input impedance of the output stage of the
PA is much larger than that of the envelope detector (ED), the inclusion of the ED does not
influence the RF signal performance.
VDD
RF in
Directional
coupler
FET
Envelope
detector
RF out
Gate bias
control
(a)
VDD
RF in
VDD
Driver
Gate bias
control
Output
Envelope
detector
RF out
Gate bias
control
(b)
Figure IV.1: (a) Conceptual block diagram and (b) actual implementation of the dynamic
biasing technique.
108
(IV.1)
109
V B0
VDD
C bp
Vgp(t)
R1
Mp
R3
Venv (t)
Rb1
Vin(t)
C1
VGGP
R2
Envelope
detector
PA input
Z PA
Gate bias
control
PA input
(a)
C bp
Vin(t)
PA input
R3
Vgp(t)
Cgp
Isdp (t)
C1
R1
Low-pass filter
R2
Venv (t)
C PA
PA input
(b)
Figure IV.2: Envelope detection and gate-bias-control circuit. (a) Schematic. (b) Equivalent large-signal model.
110
where d and c represent the angle frequencies of the envelope and carrier, i.e.,
d = 1 2
(IV.2)
1 + 2
2
(IV.3)
c =
and A is the tone amplitude.
To obtain the transient response of the envelope detector, Fourier transform of Isdp (t)
should be first calculated. However, Isdp as a function of Vgp depends on the current characteristics of Mp . In the following calculations, we will derive the Fourier transform of Isdp (t)
for a long-channel device and an ideal linear device, respectively. Then we will show that
the implemented device exhibits approximately long-channel current characteristics for the
gate-voltage range we are interested. In the analysis, all the devices are assumed at ideal
class-B biases.
kp (VSG + VTp )2
VSG > VTp
ISDP =
(IV.4)
0
VSG VTp
where
p Cox p
kp =
2
W
L
(IV.5)
p
111
and the channel-length modulation effect is ignored because only a small envelope
voltage amplitude will appear at the drain of the PMOS device.
To simplify our analysis, the envelope period (Td ) is chosen as a multiple of the
carrier period (Tc ), i.e., Td = (2M + 1)Tc , where M is an integer and much larger
than one. This is shown in Fig. IV.3.
Isdp(t)
T
- __d
2
...
t -m-1
t -m
...
Tc
__
2
...
tm
tm+1
...
Tc
__
2
time
Td
__
2
(IV.6)
Substituting (IV.1) and (IV.6) to (IV.4) and since c is much larger than d , we have
the drain current for one carrier period as
4kp A2 cos2
tm cos2 c t
2
Isdp (t)
(tm
Tc
)
4
t<
(tm +
Tc
)
4
t < (tm+1
(tm +
Tc
)
4
Tc
)
4
(IV.7)
112
Since Isdp (t) is conveniently chosen as an even function with the period of Td , it can
be expanded to the following Fourier series:
a0 X
Isdp (t) =
+
an cos(nwd t)
2
n=1
(IV.8)
where
Z
2
an =
Td
Td
2
Td
2
(IV.9)
Among all the frequency components, we are only interested in those near the envelope frequency (including some of the envelope harmonics) because all the RF
frequency components will be removed by the low-pass filter. The calculation of a1
is shown here:
2
a1 =
Td
Td
2
Td
2
Z
!
M
tm + T4c
2 X
kp A2
.
2
(IV.10)
a0 can be calculated as
a0 = kp A2 .
(IV.11)
Td
2
Td
2
cos2 (
d
t) cos(nd t) dt = 0
2
(IV.12)
113
where n = 2, 3, ..., and
nd c .
(IV.13)
It can be shown from (IV.10) to (IV.12) that the envelope component of the longchannel PMOS current for a two-tone input signal is
isdp (t) =
kp A2
(1 + cos d t).
2
(IV.14)
0
VSG VTp
(IV.15)
where kp is a constant. The same approach can be applied, and it can be shown that
the current envelope component of the ideal linear PMOS device is directly proportional to the input envelope signal amplitude, i.e.,
0
2kp A
d
cos t.
idsp (t) =
(IV.16)
c) Actual Device
The gate length of the PMOS device for the envelope detector is chosen as 0.5 m,
thus it is possible that the device current exhibits short-channel characteristics. At
low gate-bias voltages, such as in the class-B case, a short-channel device still exhibits long-channel characteristics. To verify this claim, we fitted the SPECTRE
114
simulated PMOS current data using the square function in (IV.4) and the linear function in (IV.15). The fitting was carried out for the gate voltage between 2.0 and 2.9
V. Figure IV.4 shows the SPECTRE simulated, the square-function fitted, and the
linear-function fitted curves, respectively. As can be seen, the square function can
fit the current very well in the voltage range we are interested. Thus, the employed
PMOS can be approximately modelled as a long-channel device.
10
SPECTRE DATA
Curvefit (square)
Curvefit (linear)
ISD (mA)
6
4
2
0
2
1.9
2.1
2.3
2.5
2.7
2.9
3.1
3.3
Figure IV.4: SPECTRE simulated and MATLAB fitted PMOS source-drain current versus
gate voltage. The two fitting functions are those in (IV.4) and (IV.15), respectively. The
fitting was carried out for the gate voltage between 2.0 and 2.9 V.
115
VDD
VDD
RF
choke
RF
choke
VGG0
L1
C f1
Cb2
L i1
Vout
C1
Cb1
PA
input
M0
R f1
Vin
M1
Rb1
C i1
On-chip
2f termination
Cdc
VGG1
On-chip
2f termination
V PP
Mp
Compensation
circuitry
(IV.17)
Venv (s)
R1,2
=
Isdp (s)
1 + sC1 R1,2 + sCPA (R3 + R1,2 + sC1 R3 R1,2 )
(IV.18)
where R1,2 represents the total resistance of R1 in parallel with R2 . Note that the low-pass
filter will introduce both magnitude distortion and phase delay. To minimize these effects,
C1 , R1 , R2 , and R3 should be chosen as small as possible to maximize the frequency
of the dominant pole in (IV.18). However, to isolate the influence of the bias circuitry
on the PA RF path and minimize the current consumption, R1 , R2 , and R3 should be
116
maximized. In addition, to remove the RF frequency components at the envelope detector
output, C1 should be maximized. Therefore, tradeoffs in these regards have to be made. In
our implementation, C1 is chosen as 30 pF, R1 and R2 are chosen as 200 , R3 is 100 ;
the input of the PA is approximately 76 pF. This yields the two poles of the low-pass filter
as 9.4 and 117.6 MHz, respectively.
R2
a0
VB0 + R1,2 + a1 |G(jd )| cos(d t + G(jd ))
R1 + R2
2
+ a2 |G(j2d )| cos(2d t + G(j2d )) +
(IV.19)
where G(j) represents the phase delay introduced by the low-pass filter. Since the
PMOS device we used exhibits long-channel current characteristics, we can substitute
(IV.10)-(IV.12) to (IV.19). This gives
Venv (t) =
R2
kp A2
VB0 +
[R1,2 + |G(jd )| cos(d t + G(jd ))].
R1 + R2
2
(IV.20)
For envelope frequencies much less than 9.4 MHz (the dominant pole of the low-pass filter),
(IV.20) reduces to
Venv (t)
kp A2
R2
VB0 +
R1,2 (1 + cos d t).
R1 + R2
2
(IV.21)
Here, the first term is the initial bias voltage set by the gate-bias-control circuit, as shown
in Fig. IV.2 (b); the second term is the output envelope signal. As can be seen, the output
117
envelope signal is proportional to the square of the input envelope signal amplitude. This
is due to the square relationship between the source-drain current and the gate voltage of
the long-channel PMOS device.
For illustration purposes, Fig. IV.6 shows the approximate time-domain waveforms
of Vgp (t), Isdp (t), and Venv (t). The waveforms are not scaled.
k(VGS VTn )2
(1 + VDS )
1 + (VGS VTn )
IDS =
The I V curve of the employed device can be obtained from the SPECTRE simulator
for the VGS and VDS ranges where the device will be operated. Using the MATLAB least-
118
Vgp(t)
V GG0
time
(a)
Isdp(t)
time
(b)
Venv(t)
Isdp(t)
envelope
Output
envelope
time
(c)
Figure IV.6: Approximate time-domain waveforms of (a) input gate voltage, (b) sourcedrain current of Mp , and (c) output voltage of the envelope detector for a two-tone test
signal. The dashed lines in (a) and (b) are the corresponding signal envelopes. The dashed
line in (c) is the envelope of Isdp (t) for illustrating the delay of the ED output.
119
square curvefit function, we can fit (IV.22) to the device I V data with appropriate
coefficients of k, , and .
The single-tone input voltage signals for the fixed and dynamic biasing schemes are
Vgs (t) = VGG + A cos(c t + )
(IV.23)
(IV.24)
and
respectively, where A stands for the RF signal amplitude. The corresponding drain current
is
Ids =
k(Vgs VTn )2
(1 + Vds )
1 + (Vgs VTn )
Vgs VTn
(IV.25)
To simplify our calculation, the Ids dependence on Vds was eliminated by approximating
Vds as a superposition of the dc bias and the purely linear part of the output signal:
Vds = VDD gv vgs = VDD gv A cos(c t + )
(IV.26)
where gv , the voltage gain of the amplifier at the fundamental frequency, can be estimated
from first-order simulations. Substituting (IV.23), (IV.24), and (IV.26) to (IV.25), Ids (t) can
be derived.
The dc and fundamental components of the drain current can be obtained by ap-
120
plying Fourier-series transforms to (IV.25)
IDD
1
=
T1
2
io =
T1
T1
2
T1
2
T1
2
Ids (t) dt
(IV.27)
(IV.28)
T1
2
Peff
1 2
io RO
Pout
=
= 2
.
Pdc
VDD IDD
(IV.29)
Substituting (IV.22)-(IV.28) to (IV.29), we are able to calculate the drain efficiency of the
amplifier.
avg =
(IV.30)
Here, hPdc i is the average dc power consumed by the amplifier, which directly corresponds with battery energy consumption.
121
(IV.31)
2
3
idsn = g1 vgs + g2 vgs
+ g3 vgs
.
(IV.32)
and
122
It is important to reemphasize that when the bias point or RF signal power changes, the
coefficients (c1 through c5 and g1 through g3 ) also change, such that the expansions always
trace out the appropriate Ceff and idsn versus vgs curve.
The voltage signal at the gate of M0 , when the dynamic biasing technique is applied,
contains both RF and envelope components. Assuming the spacing of the two tones is
much less than the dominant pole (9.4 MHz) of the low-pass filter, from (IV.21), the gate
voltage signal of the output device, M0 , is
Vg0 (t) = A(cos 1 t + cos 2 t) +
=(
R2
kp A2
VGG0 +
R1,2 (1 + cos d t)
R1 + R2
2
R2
VGG0 + A2 ) + [A(cos 1 t + cos 2 t) + A2 cos d t]
R1 + R2
(IV.33)
kp
R1,2
2
(IV.34)
where
=
d = 1 2 .
(IV.35)
The first term in (IV.33) is the dc bias voltage, and the second term is the ac signal.
Note that this dc bias voltage varies with the input envelope amplitude, thus will have
impact on PA linearity. This will be discussed later in this section.
Substituting the ac signal in (IV.33) to (IV.32) gives
9
3
3
idsn = (g1 A + g3 A3 + g3 2 A5 ) cos(1 t) + g3 A3 cos((22 1 )t)
4
2
4
3
+ g2 A3 cos((22 1 )t) + g3 2 A5 cos((22 1 )t) +
4
3
g1 A cos(1 t) + ( g3 + g2 )A3 cos((22 1 )t) + .
4
(IV.36)
123
Cgdn
+
ZI
~
v gs, 21 2
c1
Ceff, 21 2
~
g1 v
gs, 21 2
dsn, 21 2
ZO
Ceff ,21 2
1 3
5 5
= j(21 2 ) c3 vgs,
+ c5 vgs,
1
1
4
8
(IV.37)
(IV.38)
where vgs,1 is the phasor amplitude of the gate-source voltage at the fundamental fre-
124
quency. The distortion voltages that result at the gate and drain can then be computed using
the circuit of Fig. IV.7:
vds,21 2 =
ZO {dsn,21 2 [1 + j(21 2 )Cgdn ZI0 ] Ceff ,21 2 [g1 j(21 2 )Cgdn ]ZI0 }
1 + j(21 2 )Cgdn (ZI0 + ZO + g1 ZI0 ZO )
(IV.39)
where ZI0 ZI k c1 , and the impedances ZI0 and ZO should be evaluated at the intermodulation frequency 21 2 . The drain voltage at the fundamental frequency is also easily
found to be
vds,1 =
g1 ZO + j1 Cgdn ZO
vgs,1
1 + j1 Cgdn ZO
(IV.40)
where, in this case, ZO should be evaluated at the fundamental frequency 1 . The IM3 at
the drain are then simply
vds,21 2
.
IM3D = 20 log
vds,1
(IV.41)
125
Estimation of g2
a) a long-channel class-A device
For a long-channel class-A device, g2 is in the same order of g1 , as illustrated in the
long-channel current expression:
Cox
ids (vgs ) =
2
Cox
=
2
W
L
W
L
(vgs + VGG VT )2
2
[(VGG VT )2 + 2(VGG VT )vgs + vgs
].
(IV.42)
The IDS versus VGS curve for such a device is illustrated in Fig. IV.8 (a). Comparing (IV.42) with (IV.32), we have
g2
1
=
.
g1
2(VGG VT )
(IV.43)
Typical values of (VGG VT ) for a CMOS class-A PA is in the range of 0.25 0.5 V,
which implies that g2 is approximately 1-2 times of g1 .
b) an ideal class-B device
For an ideal class-B device that has current characteristics shown in Fig. IV.8 (b),
g2 is also in the same order of g1 . Assuming the current can be expanded for three
terms, i.e.,
3
2
.
+ g3 vgs
idsn = g1 vgs + g2 vgs
(IV.44)
g2
cos 2t + .
2
(IV.45)
126
IDS
IDS
VT
VGG
VT
VGS
VGS
vgs
vgs
(a)
(b)
Figure IV.8: IDS versus VGS for (a) a long-channel class-A device, and (b) an ideal class-B
device.
In the meantime, the output current of an ideal class-B device is also
cos t
T4 < t T4
idsn (t) =
T
0
< t 3T
4
4
(IV.46)
Here, to simplify our analysis, we let the slope of the IDS versus VGS curve as unity.
Fourier expansion on (IV.46) gives
idsn (t) =
1
2
cos t +
cos 2t + .
2
3
(IV.47)
(IV.48)
Thus, for an ideal class-B device, g2 is approximately equal to g1 . Note that the above
derivations are only for the first-order estimation.
127
c) the implemented class-AB device
For a general class-AB device, g2 varies with the device characteristics, bias voltages,
and signal amplitude. In such cases, numerical calculations described in the previous
chapter is necessary. The ratio of g2 to g1 for the four gate bias voltages (0.75 0.90 V) of the implemented class-AB device is shown in Fig. IV.9 (a). As can be
seen, this ratio for most bias voltages and power levels are larger than one.
Estimation of g3
The estimation of g3 for a CMOS class-AB device is not straightforward. Again,
numerical calculations are employed, and it was found that g3 varies dramatically with
both the gate bias voltage, VGG0 , and the output power, Pout , as shown in Fig. IV.9 (b).
(IV.49)
Thus, can be calculated. The output IM3 can then be calculated from (IV.37)(IV.41).
Figure IV.10 shows the comparison of the calculated and simulated load-voltage IM3 of
the designed CMOS class-AB PA with the dynamic biasing technique. As illustrated in
128
5
VGG0=0.75 V
VGG0=0.80 V
V
=0.85 V
GG0
VGG0=0.90 V
g2/g1
0
10
10
20
30
(a)
g3/g1
VGG0=0.75 V
VGG0=0.80 V
V
=0.85 V
GG0
VGG0=0.90 V
1
10
10
20
30
(b)
Figure IV.9: Ratio of (a) g2 and (b) g3 to g1 for the four gate bias voltages (0.75 - 0.90 V)
of the implemented class-AB device.
129
(IV.33), the dc gate bias voltage of M0 varies with the tone amplitude, A. Thus, each
swept A corresponds to a different set of power-series coefficients of Ceff and idsn , which is
obtained by performing the interpolations among the four sets of coefficients at VGG0 from
0.75 to 0.90 V.
20
40
60
Calculation (ED)
Simulation (ED)
80
10
20
30
Figure IV.10: Comparison of the calculated and simulated IM3 of the load voltage at 21
2 versus peak-envelope output power. The gate bias is designed to vary from 0.75 V to
0.85 V.
Figure IV.11 shows the contributions to the load IM3 arising from g2 , g3 and Ceff
nonlinearities, as computed from (IV.37)(IV.41). The contribution from one nonlinearity
source is found by setting the other two to zero. As shown, the g2 nonlinearity limits the
load IM3 over a wide range of power levels.
130
20
40
60
g3 contribution
g contribution
2
Ceff contribution
80
10
20
30
Figure IV.11: Contributions to the load-voltage IM3 from the g2 , g3 , and Ceff nonlinearities.
The values are computed from the Volterra expressions (IV.37)(IV.41), as described in the
text.
131
Figure IV.12: Die microphotograph of the highly integrated and compensated two-stage
CMOS PA (PA3). The ED block is the envelope detector circuit.
132
Venv (V)
0.83
Calculation
Simulation
Measurement
0.81
0.79
0.77
0.75
2
10
14
18
22
26
Figure IV.13: Calculated, simulated, and measured Venv versus output power for a singletone input.
133
dynamic biasing technique improves the PAs 1-dB compress point due to the increased
gate bias. However, this does not yield better linearity, as shown later in the linearity measurements.
35
35
VGG0=0.85 V
VGG0=dynamic
30
25
25
20
20
15
15
10
10
0
5
10
15
20
25
PAE (%)
GAIN (dB)
30
0
30
Figure IV.14: Measured gain and power-added efficiency versus output power for PA3 with
the dynamic biasing technique, and PA3 when the envelope detector is disabled and the gate
is biased at VGG0 = 0.85 V, respectively.
(IV.50)
where Pdc, 0.85 V and Pdc, dynamic represent the dc power consumption of PA3 when the
dynamic biasing technique is applied, and PA3 when the dynamic biasing technique is
disabled and the gate is biased at VGG0 = 0.85 V, respectively. Figure IV.15 shows the
measured power consumption improvement versus the PA output power. As can be seen,
the dynamic biasing technique can improve the power consumption by nearly 50 % for the
output stage and 30 % for the total two-stage.
134
60
Twostage
Output stage
50
40
30
20
10
0
5
10
15
20
25
30
Figure IV.15: Measured power consumption improvement versus the PA output power.
Linearity
To verify its linearity performance, PA3 was tested using both two-tone and WCDMA
signals. Again, the testings were carried out for both biasing schemes. Figures IV.16 show
the measured IM3 , adjacent-channel leakage power (ACP1), and alternate-channel power
(ACP2). Figures IV.17 shows the comparison between the calculated and measured loadvoltage IM3 versus output power. As can be seen, a good agreement is obtained between
the calculations and the measurements, verifying our distortion analysis. The linearity
measurements also validate our claim that the dynamic biasing technique can introduce
significant nonlinearity into the CMOS class-AB PA.
Although more nonlinear, PA3 with the dynamic biasing technique can marginally
meet the 3GPP-WCDMA ACP requirements of -33 dBc and -43 dBc at the output power
of 24 dBm.
135
VGG0=0.85 V
V
=dynamic
20
MEASURED IM3 (dBc)
GG0
30
40
50
5
10
15
20
25
30
25
30
25
30
(a)
20
VGG0=0.85 V
V
=dynamic
GG0
30
40
50
10
15
20
(b)
40
VGG0=0.85 V
V
=dynamic
GG0
50
60
70
5
10
15
20
(c)
Figure IV.16: Measured (a) IM3 , (b) adjacent-channel leakage power, and (c) alternatechannel power versus peak-envelope output power of PA3 for both biasing schemes.
136
Calculation
Measurement
20
30
40
50
5
10
15
20
25
30
Figure IV.17: Comparison between the calculated and measured load-voltage IM3 versus
output power.
IV.6 Summary
The dynamic biasing technique can improve the efficiency of a CMOS class-AB
PA at low output power levels, as demonstrated by both calculations and experiments.
However, the envelope signal introduced by the dynamic biasing technique can significantly
limit the overall linearity of the CMOS class-AB PA, as verified by both Volterra analysis
and experimental results. Thus, further linearization methods are necessary to reduce this
nonlinearity.
Chapter V
Conclusions
Linearity and efficiency are the two most important characteristics of power amplifiers for wireless applications. In this dissertation, we investigate three topics on CMOS
power amplifiers: class-E, class-AB, and dynamic biasing technique.
Class-E power amplifier is a promising candidate for realizing high efficiency. Previous analytical efforts on class-E power amplifiers assumed either zero switch resistance
and/or infinite drain inductance, leading to less optimized design. In this dissertation, we
developed an improved design technique by accounting for both finite drain inductance
and finite on resistance for a CMOS device. This design technique expresses the circuit
parameters in terms of the device width and the design specifications, such as the output
power and operating frequency fc . A design example based on the developed algorithm
achieves an output power of 0.25 W and a drain efficiency of 87% for a 3.5 mm NMOS
class-E device with VDD = 2 V and fc = 1.90 GHz.
The intrinsic linearity obtained in a CMOS class-AB operation is often insufficient to
meet the stringent linearity requirement imposed by modern wireless standards. In this dissertation, we found that the nonlinear gate-source capacitance is a dominant source of distortion that limits the linearity of CMOS class-AB power amplifiers. A simple technique is
137
138
proposed to cancel this nonlinearity by using a compensating nonlinearity, provided by the
gate-source capacitance of an appropriately biased and sized PMOS device placed alongside the NMOS device that provides the class-AB amplification. Volterra analysis and
two-tone SPECTRE simulations were used to verify the technique. Prototype two-stage
CMOS class-AB power amplifiers were implemented. Experiments show that the amplifiers employing the compensation technique can improve both the two-tone, third-order
intermodulation and adjacent-channel leakage power by approximately 8 dB. When operated at VDD = 3.3 V, the final linearized power amplifier is capable of delivering an output
power of 24 dBm with a small-signal gain of nearly 24 dB and an overall power-added
efficiency of 29 %. At the designed output power of 24 dBm, the adjacent-channel leakage
power of the linearized amplifier is -35 dBc, meeting the 3GPP-WCDMA requirements
of -32 dBc. The experimental results also prove the feasibility of linear CMOS class-AB
power amplifiers for wireless communication systems.
Although the designed two-stage CMOS class-AB power amplifier exhibits good linearity and maximum efficiency, it still suffers serious efficiency degradation when operated
at low output power levels. This deserves special attentions considering the statistical nature of power usage in wireless communication systems: as exemplified in [36], the most
probable output power of a IS-95-CDMA system is only 1 mW, despite the maximum of
0.5 W. In this dissertation, it was demonstrated that a dynamic biasing technique can improve the efficiency of a CMOS class-AB power amplifier by controlling the gate bias
voltage with the envelope of input RF signal. However, the envelope signal introduced by
139
the dynamic biasing technique can significantly limit the overall linearity of the CMOS
class-AB PA, as verified by both Volterra analysis and experimental results. The prototype
power amplifier employing the dynamic biasing technique exhibited more than 6 dB worse
in IM3 and ACP performances than the one without the technique applied. Thus, further
linearization or compensation methods are necessary to reduce the nonlinearity introduced
by the dynamic biasing technique.
Bibliography
[1] X. Zhang, An Improved Outphasing Power Amplifier System of Wireless Communications. PhD thesis, University of California, San Diego, 2001.
[2] Y. Tsividis, Operation and Modeling of The MOS Transistor. Boston, MA: McGrawHill, second ed., 1999.
[3] P. R. Gray, P. J. Hurst, S. H. Lewis, and R. G. Meyer, Analysis and Design of Analog
Integrated Circuits. New York, NY: Wiley, 2001.
[4] R. S. Narayanaswami, RF CMOS Class C Power Amplifiers for Wireless Communications. PhD thesis, University of California, Berkeley, Fall 2001.
[5] G. D. Ewing, High-Efficiency Radio-frequency Power Amplifier. PhD thesis, Oregon
State University, June 1964.
[6] N. Sokal and A. Sokal, Class E-a new class of high-efficiency tuned single-ended
switching power amplifiers, IEEE Journal of Solid-State Circuits, vol. 10, pp. 168
76, June 1975.
[7] F. Raab, Idealized operation of the class E tuned power amplifier, IEEE Transactions on Circuits and Systems, vol. 24, pp. 72535, December 1977.
[8] C.-H. Li and Y.-O. Yam, Maximum frequency and optimum performance of class
E power amplifiers, in IEE Proceedings-Circuits, Devices and Systems, pp. 17484,
1994.
[9] K.-C. Tsai and P. Gray, A 1.9-GHz, 1-W CMOS Class-E power amplfier for wireless
communications, IEEE Journal of Solid-State Circuits, vol. 34, pp. 96270, July
1999.
[10] C. Yoo and Q. Huang, A common-gate switched 0.9-W class-E power amplifier with
41 % PAE in 0.25-m CMOS, IEEE Journal of Solid-State Circuits, vol. 36, pp. 823
30, May 2001.
[11] K. Mertens and M. Steyaert, A 700-MHz 1-W fully differential CMOS class-E power
amplifier, IEEE Journal of Solid-State Circuits, vol. 37, pp. 13741, February 2002.
2
[12] S. C. Cripps, RF Power Amplifiers for Wireless Communications. Boston, MA: Artech
House, 1999.
[13] B. Razavi, RF Microelectronics. Upper Saddle River, NJ: Prentice Hall, 1998.
[14] D. A. Johns and K. Martin, Analog Integrated Circuit Design. New York, NY: Wiley,
1997.
[15] M. J. Chudobiak, The use of parasitic nonlinear capacitor in class-E amplifier, IEEE
Transactions on Circuits and Systems-I: Fundamental Theories and Applications,
vol. 41, pp. 941944, December 1994.
[16] C. K. T. Chan and C. Toumazou, Physically based design of a class e power amplifier with non-linear transistor output capacitance, in IEE Analog Signal Processing
Seminar, pp. 5/15/4, 2000.
[17] S. H.-M. Jen, C. C. Enz, D. R. Pehlke, M. Schroter, and B. J. Sheu, Accurate modeling and parameter extraction for MOS transistors valid up to 10 GHz, IEEE Transactions on Electron Devices, vol. 46, pp. 22172227, November 1999.
[18] C. C. Enz, MOS transistor modeling for RF integrated circuit design, in Proceedings
of the IEEE Custom Integrated Circuits Conference, pp. 189196, 2000.
[19] C. C. Enz and Y. Cheng, MOS transistor modeling for RF IC design, IEEE Transactions on Electron Devices, vol. 35, pp. 201231, February 2000.
[20] Y. Cheng, C.-H. Chen, C. Enz, M. Matloubian, and M. J. Deen, MOS modeling for
RF circuit design, in Proceedings of the Third IEEE International Caracas Conference on Devices, Circuits and Systems, pp. D23/18, 2000.
[21] Y. Cheng, C.-H. Chen, M. Matloubian, and M. J. Deen, High-frequency small signal AC and noise modeling of MOSFETs for RF IC design, IEEE Transactions on
Electron Devices, vol. 49, pp. 400408, March 2002.
[22] IBM Corporation, SiGe5AM Model Reference Guide, September 2002.
[23] S. A. Maas, Nonlinear Microwave Circuits. Piscataway, NJ: IEEE Press, 1997.
[24] W. Liu, MOSFET Models for SPICE Simulation Including BSIM3v3 and BSIM4. New
York, NY: Wiley, 2001.
[25] S. Watanabe, S. Takatuka, K. Takagi, H. Kuroda, and Y. Oda, Simulation and experimental results of source harmonic tuning on linearity of power GaAs FET under class
AB operation, in IEEE MTT Symposium, pp. 17711774, 1996.
[26] R. Nagy, J. Bartolic, and B. Modlic, GaAs MESFET small signal amplifier intermodulation distortion improvement by the second harmonic termination, in Electrotechnical Conference, pp. 348361, 1998.
3
[27] V. Aparin and C. Persico, Effect of out-of-band terminations on intermodulation
distortion in common-emitter circuits, in IEEE MTT Symposium, pp. 977980, 1995.
[28] H. Bode, Network Analysis and Feedback Amplifier Design. New York, NY: Nostrand,
1945.
[29] Technical Specification 3GPP TS 25.101 V6.1.0, June 2003.
[30] D. Su and W. McFarland, An IC for linearizing RF power amplifiers using envelope
elimination and restoration, IEEE Journal of Solid-State Circuits, vol. 33, pp. 2252
2258, December 1998.
[31] A. Giry, J.-M. Fourier, and M. Pons, A 1.9 GHz low voltage CMOS power amplifier
for medium power RF applications, in IEEE RFIC Symposium, pp. 121124, 2000.
[32] C.-C. Yen and H.-R. Chuang, A 0.25-m 20-dBm 2.4-GHz CMOS power amplifier with an integrated diode linearizer, IEEE Microwave and Wireless Components
Letters, vol. 13, pp. 4547, February 2003.
[33] V. Vintola, M. Matilainen, S. Kalajo, and E. Jarvinen, Variable-gain power amplifier
for mobile WCDMA applications, IEEE Transactions on Microwave Theory and
Techniques, vol. 49, pp. 24642471, December 2001.
[34] H. Jager, A. Grebennikov, E. Heaney, and R. Weigel, Broadband high-efficiency
monolithlic InGaP/GaAs HBT power amplifiers for 3G handset applications, in IEEE
MTT-S digest, pp. 10351038, 2002.
[35] N. Srirattana, A. Raghavan, D. Heo, P. Allen, and J. Laskar, A high-efficiency multistage doherty power amplifier for WCDMA, in IEEE MTT Symposium, pp. 397400,
2003.
[36] G. Hanington, P.-F. Chen, P. M. Asbeck, and L. E. Larson, High-efficiency power
amplifier using dynamic power-supply voltage for CDMA applications, IEEE Transactions on Microwave Theory and Techniques, vol. 47, pp. 14711476, August 1999.
[37] A. A. M. Saleh and D. C. Cox, Improving the power-added efficiency of FET amplifiers operating with varying-envelope signals, IEEE Transactions on Microwave
Theory and Techniques, vol. 86, pp. 5156, January 1983.
[38] J. F. Sevic, Statistical characterization of RF power amplifier efficiency for CDMA
wireless communication systems, in Wireless Communications Conference, pp. 110
113, 1997.