A Dissertation
by
CHUNYU XIN
DOCTOR OF PHILOSOPHY
August 2004
A Dissertation
by
CHUNYU XIN
DOCTOR OF PHILOSOPHY
August 2004
ABSTRACT
The beginning of the 21st century sees great development and demands on wireless
Low operational power, low cost, small form factor, and function diversity are the cru-
cial requirements for a successful wireless product. The receiver’s front-end circuits
play an important role in faithfully recovering the information transmitted through
receiver architecture was proposed and designed using a pure CMOS process. The
front-end of the receiver consists of a low noise amplifier (LNA) and mixer. The
intermediate frequency was chosen to be 2MHz to save battery power and alleviate
the low frequency noise problem. A conventional LNA architecture was used for
reliability. The mixer is a modified Gilbert-cell using the current bleeding technique
to further reduce the low frequency noise. The front-end draws 10 mA current from a
3 V power supply, has a 8.5 dB noise figure, and a voltage gain of 25 dB and -9 dBm
IIP3.
a multi-standard application. The two standards are IEEE 802.11b and Bluetooth.
They work together making the wireless experience more exciting. The front-end is
iv
Besides the system level contributions, intensive studies were carried out on the
development of quality LNA circuits. Based on the multi-gated LNA structure, a
This LNA configuration can achieve comparable linearity to its CMOS multi-gated
counterpart and work at a higher frequency with less power consumption. A LNA us-
ing an on-chip transformer source degeneration is proposed to realize input impedance
media applications.
v
ACKNOWLEDGMENTS
Upon finishing my graduate study at Texas A&M University, I would like first to
has been the sole support for my research and study in the past several years. I
have benefited from his academic foresight and technical insight in research. His
particularly grateful to him for his advice in the academic realm and also for my
professional career.
I would like to thank Dr. Jóse Silva-Martı́nez for his advice on my course of study
and project questions. He was always available for questions and comments. I am
also grateful to Dr. Sherif Embabi. I was his teaching assistant for one semester and
really learned much from helping him with the class. His attitude toward work and
study gave me great encouragement. I also give thanks to Dr. Aydin Ilker Karsilayan.
He not only gave advice on my studies, but also supported me and the whole analog
and mixed signal center group by maintaining a great UNIX computing environment.
Some of the simple yet effective scripts he developed really made life a lot easier.
Most of my research time was spent on two major research projects, the Bluetooth
and Chameleon projects. The contribution and cooperation of the team members
made the learning and research study a great experience. Here I would like to thank
all of the team members, including Wenjun Sheng, Ahmed Emira, Bo Xia, Sung Tae
Moon, Ari Yakov Valero-Lopez, Alberto Valdes-Garcia, Ahmed Mohieldin and David
Hernandez-Garduño for the great team environment we created together. In these
two big projects, we learned from each other, tackled problems together and cheered
on each other. I felt the power of team work. I think this valuable experience will
vii
I would like to thank the secretary of our group, Ella Gallagher. Her enthusiasm,
warm sense of humor and positive attitude is a valuable asset for the group. I am
also indebted to all the other professors and students in the analog and mixed signal
Special thanks go to my parents for their love. Finally, I want to thank my dear
wife, Jingding Dou. I would not have been able to achieve this goal without her love
and support.
viii
TABLE OF CONTENTS
CHAPTER Page
I INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . 1
A. Research Motivation . . . . . . . . . . . . . . . . . . . . . 3
B. Dissertation Overview . . . . . . . . . . . . . . . . . . . . 3
CHAPTER Page
5. Harmonic Mixer . . . . . . . . . . . . . . . . . . . . . 66
D. Mixer Design for a Low-IF Bluetooth Receiver . . . . . . . 67
1. Low-IF Bluetooth Receiver Architecture . . . . . . . . 69
2. Implementation of the Down-Conversion Mixer . . . . 69
3. Layout Considerations and Simulation Results . . . . 74
4. Experimental Results of the Mixer Within the Receiver 77
CHAPTER Page
REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
APPENDIX A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
APPENDIX B . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
VITA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
xi
LIST OF TABLES
TABLE Page
TABLE Page
LIST OF FIGURES
FIGURE Page
1 A receiver RF front-end . . . . . . . . . . . . . . . . . . . . . . . . . 1
6 MOS transistor thermal noise model (a) gate and drain noise
sources (b) input-referred noise sources . . . . . . . . . . . . . . . . 21
12 Triple-order beats . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
FIGURE Page
20 Simulation plots of the designed LNA (a) Voltage gain, S11 and
noise figure (b) IIP3 and P1dB . . . . . . . . . . . . . . . . . . . . . . 52
23 Mixing process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
27 LO-to-RF leakage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
28 Single-diode mixer . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
33 Sub-sampling mixer . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
34 Harmonic mixer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
FIGURE Page
42 Bluetooth/Wi-Fi receiver . . . . . . . . . . . . . . . . . . . . . . . . 83
46 Noise optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
47 LNA layout . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
51 Mixer layout . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
FIGURE Page
60 IIP3 plot for 2-tone test at 12MHz and 25MHz offset . . . . . . . . . 107
61 IIP2 plot for 2-tone test at 12.2MHz and 12.8MHz offset . . . . . . . 108
FIGURE Page
FIGURE Page
104 Simulated small signal performance of the dual-band LNA (a) S11
and S21 (b) noise figure and minimum noise figure . . . . . . . . . . 173
105 Simulated IIP3 of the dual-band LNA (a) GSM-900 (b) DCS-1800 . . 174
111 Coplanar stripline operation mode (a) Even mode (b) Odd mode . . 185
FIGURE Page
CHAPTER I
INTRODUCTION
It is generally known that a typical receiver consists of two major parts: i) the RF
front-end, which performs small signal amplification and frequency down conversion,
and ii) a base-band portion, which performs demodulation and generates the required
digital control to the system. Although there is no specific definition for the RF front-
end, it usually includes a low noise amplifier (LNA), image rejection filter (IRF), down
Mixer
ADC DSP
The LNA is a very important block of the whole receiver. It interfaces directly
with the antenna (usually there is a passive RF band-pass filter between the antenna
and the LNA, the function of this filter is to provide band selection). The LNA
must be able to provide enough power amplification and introduce small additional
amounts of noise to the system. It also should be linear enough to tolerate high power
For a heterodyne receiver, the image rejection filter follows the LNA to remove
the image frequency component from the band. The image of the signal resides on the
other side of the local oscillator frequency, if not removed or attenuated somehow,
it will fold into the IF band with the signal, and degrade the signal to noise and
distortion ratio. For a zero-IF receiver architecture, there is no image problem, so
the image rejection filter is not needed. For a low-IF receiver, which usually has a
relatively low intermediate frequency (several MHz), the image is rejected after the
important block in the receiver. It converts the signal from a high frequency to a low
frequency to make the signal process easier and more effective. Because the mixer
operates among three frequencies (RF, LO and IF), it must be able to provide enough
isolation between them, deliver enough conversion gain, and keep reasonable dynamic
range.
The local oscillator and frequency synthesizer provide a stable and clean pro-
grammable LO signal to the mixer for channel selection. The IF filters and amplifiers
are used to further filter out the unwanted signal, and provide the proper signal level
to the base-band analog-to-digital converter (ADC). The IF amplifier is usually a
variable gain amplifier (VGA), adjusting its gain according to the received signal
strength such that the ADC always sees the optimum input signal level.
3
A. Research Motivation
The rapid development in the wireless technology introduces new design issues and
challenges, such as the low power consumption, high speed, low cost, small form factor
circuits: LNA and mixer for both narrow band and broad band applications. The
LNA and mixer are the two blocks in the signal path operating at the highest signal
frequency. They see all the interferences and noise coming from the wireless channel.
The quality of the front-end circuits directly affect the performance of the whole
system. The challenge of LNA design is to implement it by meeting the voltage gain,
noise, linearity, silicon area and power consumption at the same time. The mixer
design in low-IF and direction conversion architecture requires mainly low noise and
high linearity. Also the non-linear switching behavior of mixer’s current commutating
pair make the design of mixer more difficult. The emphasis will be put on the system
integration of the LNA and mixer in Bluetooth and IEEE 802.11b receivers, LNA
linearization technique using bipolar transistor, LNA input matching network design
using on-chip transformer and possible wide-band LNA implementations. The next
B. Dissertation Overview
The whole dissertation is organized as follows. Chapter I discusses the research mo-
tivations, and the outline skeleton of the work. In Chapter II and Chapter III, an
overview of the LNA and mixer design issues and techniques will be given. Based on
the current bleeding/injection technique, the later sections of Chapter III introduces
a mixer implementation for a low-IF Bluetooth receiver. Chapter IV is dedicated to
receiver. The LNA and mixer design trade-offs and considerations are elaborated
in this chapter. Chapter V and Chapter VI give novel LNA design techniques for
dissertation.
5
CHAPTER II
Low noise amplifier (LNA) is the first gain stage encountered in a receiver environ-
ment either wired or wireless. It must meet several specifications at the same time,
which make its design really challenging. Signal coming from the receiver antenna
is very small, usually from -100 dBm (3.2 µV ) to -70 dBm (0.1 mV )1 , therefore sig-
nal amplification is needed for the following stage (mixer) to handle. This sets the
requirement of a certain gain to the LNA. The received signal should have a certain
of the LNA. Large signal or blocker can occur at the input of LNA. The circuits should
be sufficiently linear in order to have a reasonable signal reception. For portable and
A. Basics on S-parameters
parameter set that relates to the traveling waves that are scattered or reflected when
an n-port network is inserted into a transmission line. S-parameter analysis is basi-
1
The unit dBm is a power unit referred to 1 mW in dB scale. If a signal’s power
is P Watt, then in dBm, it is 10 log (P/1mV) dBm. When related to a voltage, there
is a reference impedance R involved. This impedance is usually 50 Ω. When we say
0.1 mV is corresponding to -70 dBm, what we mean is that the power dissipated into
a 50 Ω resistor by a sinusoidal voltage signal which has a peak value of 0.1 mV is 10−7
mW, and corresponds to -70 dBm.
6
cally a modeling method to characterize an n-port linear network. There are other
parameters. All of them fall into the same modeling category for a network together
with S-parameters. They are behavioral modeling methods. The network or device is
treated as a black box, only the interaction between the ports and outer environment
is modeled.
For low frequencies, H-, Y-, or Z-parameters are more widely used. They use port
voltage and current as variables. But when the frequency moves higher and higher,
short and/or open circuit condition at each port. This can be hard to do, especially at
RF frequencies where lead inductance and capacitance make short and open circuits
difficult to obtain. Active devices, such as transistors and tunnel diode, very often
can not be connected in stable short or open circuit conditions. S-parameters, on the
other hand, are usually measured with the device imbedded between a 50 Ω load and
source, and there is very little chance for oscillations to occur. Another important
advantage of S-parameters stems from the fact that traveling waves, unlike terminal
voltages and currents, do not vary in magnitude at points along a lossless transmission
line. This means that scattering parameters can be measured on a device located at
some distance from the measurement transducers, provided that the measuring device
ZS I1 I2
+ Linear +
VS V1 Two-Port V2 ZL
- Network -
a1 a2
b1 b2
Fig. 2. Two-port network showing incident waves (a1 , a2 ) and reflected waves (b1 , b2 )
used in S-parameter definiations
impedances, but the same reference impedance Z0 will the used for all the ports here.
The independent variables a1 and a2 can be related to port voltages (V1 , V2 ) and
V 1 + I 1 Z0 Vi1
a1 = √ =√ (2.7)
2 Z0 Z0
V 2 + I 2 Z0 Vi2
a2 = √ =√ (2.8)
2 Z0 Z0
V1 +I1 Z0 V2 +I2 Z0
where Vi1 = 2
and Vi2 = 2
are the voltage waves incident on port 1 and
port 2 respectively. Therefore, |a1 |2 is the power incident on the input of the network,
8
and is also the power available from a source impedance Z0 . |a2 |2 is the power incident
on the output of the network, and is also the power reflected from the load.
Similarly, the dependent variables b1 and b2 can be related to port voltages and
currents as follows:
V 1 − I 1 Z0 Vr1
b1 = √ =√ (2.9)
2 Z0 Z0
V 2 − I 2 Z0 Vr2
b2 = √ =√ (2.10)
2 Z0 Z0
V1 −I1 Z0 V2 −I2 Z0
where Vr1 = 2
and Vr2 = 2
are the voltage waves reflected from port 1
and port 2 respectively. Therefore, |b1 |2 is the power reflected from the input port
of the network, or the power available from a Z0 source minus the power delivered
to the input of the network. |b2 |2 is the power reflected from the output port of the
network, or the power incident on the load, which is also the power that would be
delivered to a Z0 load.
From the above explanation of a1 , a2 and b1 , b2 , the four S-parameters are simply
|s12 |2 = Reverse transducer power gain with Z0 load and source (2.14)
It needs to be noticed that all of these two-port parameters are equivalent, be-
cause they describe the same network property. However, for all the practical pur-
poses, S-parameters are easier for the LNA design and characterization, while under
certain scenarios, Y-, H-, or Z-parameters maybe more straightforward. Observe that
9
not all the networks (e.g. ideal transformer, a serials impedance or a shunt admit-
Gain performance for a RF amplifier is determined by the RF transistor itself and the
input/output matching network. Fig. 3 shows a single-stage amplifier block diagram.
and load impedance Zs and ZL . s11 and s22 are the input and output reflection coeffi-
cients with Z0 source and load termination. For an arbitrary impedance termination,
the input and output reflection coefficient Γin and Γout for a two-port network [1] can
be found to be
b1 s12 s21 ΓL
Γin = = s11 + (2.15)
a1 1 − s22 ΓL
b2 s12 s21 Γs
Γout = = s22 + (2.16)
a2 1 − s11 Γs
Zs −Z0 ZL −Z0
where Γs = Zs +Z0
and ΓL = ZL +Z0
are the source and load reflection coefficient
respectively.
a1 a2 PL
b1 b2 Vo
ZC
Input + + Output
Match Match Load or
V1 RF Amplifier V2
Vs
Zin
and
Filtering
_ [S] _ and
Filtering
2nd
Stage
Zout
Network Network
Γs Γin Γout Γ L
If the input and output are simultaneously complex conjugate matched, i.e. Γin =
10
Γ∗s and Γout = Γ∗L , the amplifier has maximum power transfer. Usually it is not easy
for an unilateral device, s12 is practically zero, then Γin = s11 and Γout = s22 . The
input and output are decoupled from each other, matching can be done at the input
Several gain definitions exist for an amplifier. Power gain (G) characterizes the
actual power amplification of an amplifier. It is defined as the power delivered to the
load divided by the power input to the network. Available power gain (GA ) shows
LNA input is interfaced off-chip and usually matched to specific impedance (50 Ω or
75 Ω). Its output is not necessarily matched if directly driving on-chip blocks such as
the voltage at the input port of the amplifier. Relating to the s-parameters of the
amplifier
V2 s21 (1 + ΓL )
AV = = (2.17)
V1 (1 − s22 ΓL ) (1 + Γin )
The transducer power (GT ) [1] is defined as the power delivered to the load
divided by the power available from the source:
PL
GT = (2.18)
PAV S
where PL equals the power incident on load minus the power reflected from load:
PL = |b2 |2 1 − |ΓL |2
¡ ¢
(2.19)
11
and PAV S is
|bs |2
PAV S = (2.20)
1 − |Γs |2
√
Vs Z0
where bs = ZS +Z0
. Therefore
¯ ¯2
¯ b2 ¯ ¡
GT = ¯¯ ¯¯ 1 − |Γs |2 1 − |ΓL |2
¢¡ ¢
(2.21)
bs
b2
Using the signal flow chart, the ratio bs
can be found to be
b2 s21
= (2.22)
bs (1 − s11 Γs ) (1 − s22 ΓL ) − s12 s21 Γs ΓL
lateral, i.e., s12 is small and effectively zero, (2.23) will be reduced to
1−|Γs |2 1−|ΓL |2
GT u = GT |s12 =0 = |1−s11 Γs |2
|s21 |2 |1−s22 ΓL |2
(2.24)
= GS |s21 |2 GL
where GS and GL are the source and load mismatch factor respectively:
1 − |Γs |2
GS = (2.25)
|1 − s11 Γs |2
1 − |ΓL |2
GL = (2.26)
|1 − s22 ΓL |2
Once the device and its bias condition is established, s21 remains unchanged. GS is
only related to the input parameters s11 and Γs . GS shows the degree of mismatch
between the source impedance and the input impedance. Similarly, ML is only related
to the output parameters s22 and ΓL and shows the matching condition at the output.
When Γs equals s11 ’s complex conjugate s∗11 , GS reaches its maximum value of
12
1
GS, max = 1−|s11 |2
. This can be verified by writing s11 = a + jb and Γs = x + jy,
∂GS ∂GS
substituting them into (2.25) and solve equations ∂x
= 0 and ∂y
= 0 for x and y.
The same result can be obtain for ΓL : when ΓL = s∗22 , GL reaches its maximum value
1
of GL, max = 1−|s22 |2
.
It is also easy to see that when |Γs | = 1, GS has a minimum value of zero and
GL have different circles. These circles are usually referred to as constant gain circles.
They have their centers located on the line drawn from the center of the Smith Chart
to the point of s∗11 or s∗22 . Specifically, the center of the circles resides at
gi s∗ii
ci = (2.27)
1 − |sii |2 (1 − gi )
where gi = Gi 1 − |sii |2 = Gi
¡ ¢
Gi, max
is the normalized gain value for the gain circle
Gi . The subscript i represents S for input and L for output, ii represents 11 for
input and 22 for output. The detailed proof of the constant gain circle is given in
Appendix A.
The stability of RF amplifiers is also a very important metric. Only when the
amplifier is stable, the other metrics such as gain, noise figure are meaningful. Suppose
the input impedance at the input port of the amplifier is Zi = Ri + jXi , then the
From (2.29) it is easy to see that if the real part of the input impedance is negative,
13
i.e. if Ri < 0, then |Γin | > 1. If this negative impedance compensates the loss
coming from the input termination network, oscillation can occur. The amplifier is
potentially unstable. The same argument holds for the output. If for all the passive
terminations at the input and output, the following conditions hold, then the amplifier
A meaning for (2.30) is that the real parts of input and output impedance of the
amplifier are resistive, or, in Smith Chart, Γin and Γout will never go out of the unity
circle.
It can be shown that the conditions for unconditionally stable [1] in term of
s-parameters are
|s11 | < 1
|s22 | < 1 (2.31)
K > 1
where K is the stability factor given by
s11 = 0.38∠ − 115◦ , s12 = 0.06∠14◦ , s21 = 6.0∠104◦ , s22 = 0.50∠ − 52◦ (2.33)
It can be calculated from the above data that |s11 | = 0.38 < 1, |s22 | = 0.5 < 1, and
K = 1.02 > 1. So the amplifier is unconditionally stable. While at 750 MHz, the
same amplifier has the following S-parameter:
s11 = 0.56∠ − 78◦ , s12 = 0.05∠33◦ , s21 = 8.6∠122◦ , s22 = 0.66∠ − 42◦ (2.34)
14
In this case, |s11 | = 0.56 < 1, |s22 | = 0.66 < 1, but K = 0.60 < 1, So it is potentially
unstable.
In order to stabilize an active device, one simple way is to add a series resistance
or a shunt conductance to the unstable port. For example, the input port of the
amplifier at 750 MHz mentioned above can be stabilized by series a 16.5 Ω resistor or
shunt a 17.8 Ω resistor. Its output port can be stabilized with a 40 Ω series resistor or
a 161 Ω shunt resistor. In practise, due to the coupling between the input and output
ports of the amplifier, it is usually sufficient to just stabilize one of the ports, and
one should avoid resistive loading of the input port, because it will cause additional
noise to be amplified. Therefore, one will generally try to stabilize the output port.
The prices paid for stabilizing using resistive loading are poor impedance matching,
C. Noise Performance
to the amplifier, it is defined as the SNR at the input of the network divided by the
SN Rin
F = (2.35)
SN Rout
where SN Rin and SN Rout are the SNR at the input and output of the amplifier
respectively. The represent the signal’s quality in terms of noise before and after the
network. Noise figure is just the logarithm form of noise factor, its unit is dB:
It will be pretty involved if one tries to use a transistor’s equivalent noise circuit to
analysis the whole amplifier or network. Using a two-port network noise model can
simplify the calculation of its noise factor and also gain insight on minimize noise
figure [2].
An effective way to analyze noise in a given circuit is to assume that the circuit
is noiseless and model its internal noise by external noise sources at the input and
output ports. These noise sources must have the same noise power appearing at the
shown in Fig. 4. In the following discussions, it is assumed that port 1 is the input
i1 Vn1 Vn2 i2
Noiseless
V1 Network V2
(a)
i1 i2
Noiseless
V1 in1 Network
in2 V2
(b)
i1 Vn i2
Noiseless
V1 in Network V2
(c)
Using the Z-parameters in Fig. 4(a), the voltage-current relationship among ports
can be written as
The equivalent noise source vn1 and vn2 can be measured from the open-circuited
(o.c) measurements as
Using the Y-parameters in Fig. 4(b), the voltage-current relationship among ports
can be written as
i1 = y11 v1 + y12 v2 + in1 (2.41)
The equivalent noise source in1 and in2 can be obtained from the short-circuited (s.c.)
measurements as
in1 = i1 |v1 =v2 =0 (2.43)
Referring the noise sources to the input port is convenient for noise analysis.
This leads to the ABCD-parameter representation in Fig. 4(c),
The noise vn and in cannot be measured using the o.c. or s.c. measurement technique,
but they can be found as function of vn1 and vn2 or as function of in1 and in2 through
17
source connected to its input port as shown in Fig. 5. It is assumed that the noise of
the source is and the noise of the network, in and vn , are uncorrelated but vn and in
may be correlated.
vn i2
Noiseless +
is Ys in
Network v2
-
The total output noise power is proportional to the mean square value of the
short-circuited current (i2s.c. ) at the input port of the noiseless network. The noise
due to the source is only proportional to the mean square value of the current i2s . The
i2s.c. (in + vn Ys )2
F = =1+ (2.51)
i2s i2s
Typically there is correlation between vn and in . The gate and drain noise in
where inu is the part of in which is uncorrelated with vn , and inc is the part of in which
inc = Yc vn (2.53)
Note that Yc is not a physical admittance, it is called admittance only because it has
The noise factor can be rewritten using the above established equations as:
resistance Rn ,
vn2 = 4kT Rn ∆f (2.56)
The uncorrelated noise current inu can be expressed in terms of an equivalent (ficti-
tious) noise conductance Gu ,
Now the noise factor can be written in terms of noise parameters Rn , Gs and Gu ,
G u Rn £
(Gs + Gc )2 + (Bs + Bc )2
¤
F =1+ + (2.58)
Gs Gs
19
Note that Gs and Bs can be changed independently. So the noise factor can be
Now, the second square term in the square brackets of (2.58) becomes zero, and F is
reduced to
G u Rn
F =1+ + (Gs + Gc )2 (2.60)
Gs Gs
It can be further minimized by using a proper value for Gs , which can be found by
The above Gs and Bs describe the optimum source admittance which would
minimize the noise factor, and will be referred to as Gopt and Bopt respectively.
Rn
F = Fmin + |Ys − Yopt |2 (2.63)
Gs
By using normalized impedance and admittance, the noise factor can be ex-
pressed as
rn
F = Fmin + |ys − yopt |2 (2.64)
gs
20
Rn
where rn = Zo
, gs = Gs Zo , ys = Ys Zo and yopt = Yopt Zo .
1 − Γs
ys =
1 + Γs
1 − Γopt
yopt =
1 + Γopt
and
1 1 − |Γs |2 1 − |Γs |2
gs = (ys + ys∗ ) = =
2 1 + |Γs |2 + (Γs + Γ∗s ) |1 + Γs |2
So the noise factor can be expressed in terms of reflection coefficients Γs and Γopt ,
Now the concept of the constant noise circle will be introduced. For this purpose,
rewrite (2.65) as
|Γs − Γopt |2
=N (2.66)
1 − |Γs |2
where N = F −Fmin
4rn
|1 + Γopt |2 . Since F ≥ Fmin , then N ≥ 0. For constant F , N is
also constant. It can be shown that the source reflection coefficient Γs for constant
noise factor F is a circle in the Smith Chart Γs -plane (see Appendix A). Its radius
and
Γopt
cF = (2.68)
1+N
respectively.
moving the center cF toward the origin of the Γs -plane, and enlarge the radius of the
circle. It can be concluded that the outer circles have a larger noise factor than the
inner ones.
The most popular CMOS LNA usually consists of just one stage utilizing one MOS
transistor. Design insight can be gained by studying the two-port noise parameters
the noise factor can be calculated by knowing these four noise parameters.
For the first order approximation, the drain and induced gate current noise dom-
inate the noise performance of a MOS transistor. Fig. 6(a) shows the physical origin
i
Fig. 6. MOS transistor thermal noise model (a) gate and drain noise sources (b) in-
put-referred noise sources
where gdo is the drain source conductance at zero drain-source biasing. γ is about
2/3 for long channel devices in the saturation region. For short channel devices, γ is
typically 2-3 or even larger [3]. This increased value of γ is due to carrier heating by
large electric fields developed across drain and source. Thus keeping the drain-source
where δ is the gate noise coefficient and is about twice the value of γ for long channel
ing · i∗nd
c= q (2.72)
i2ng · i2nd
For long channel devices its value is j0.395. The pure imaginary value implies that
the correlation is due to capacitive coupling from channel to gate. So the gate noise
current can be expressed as the sum of component ingc which is fully correlated with
drain current noise and component ingu which is completely uncorrelated with drain
current noise.
Using the method described in the previous section, refer all the noise sources
in Fig. 6(a) to the input port (gate-source) of the MOS device as in Fig. 6(b). The
noise voltage is
4kT γgdo ∆f
vn2 = 2
(2.73)
gm
23
where inc1 = jωCgs vn , Cgs is the gate-source capacitance. inc1 is completely correlated
and
Whereby if the gate noise is ignored, noise voltage and noise current will be fully
correlated.
The four noise parameters [3] can be found to be
γgdo γ 1
Rn = 2
= (2.77)
gm α gm
às !
δ
Yc = Gc + jBc = jωCgs 1 + α |c| (2.78)
5γ
δω 2 Cgs
2
1 − |c|2
¡ ¢
Gu = (2.79)
5gdo
gm
where α = gdo
. It is seen that Gc , the real part of Yc , is essentially zero. Thus the
and às !
δ
Bopt = −Bc = −ωCgs 1 + α |c| (2.81)
5γ
respectively. The minimum noise factor is
2 ω
q ¡
γδ 1 − |c|2
¢
Fmin = 1 + 2Rn Gu = 1 + √ (2.82)
5 ωT
24
gm
where ωT = Cgs
.
In order to achieve as small of a noise factor as possible, one can design an input
and biasing condition of transistors. By making Fmin small will also reduce the
overall noise figure. From (2.82), one can conclude that increasing ωT and/or the
correlation coefficient c will reduce the minimum noise factor. For a chosen process,
the correlation coefficient is fixed by the technology. The choice left for the circuit
ωT
designer to reduce Fmin is to have large ω
. For example, suppose that γ = 2 and
ωT ωT
δ = 4, and c is assumed to be 0.395. If ω
= 5, then Fmin = 1.7 dB, while if ω
= 15,
Fmin will be reduce to 0.6 dB.
The importance of gain and noise factor specifications on LNA can be discussed
further from the receiver’s system sensitivity aspect. The sensitivity P s represents
the smallest input signal power that can be reliably detected by the system [4]
The first two terms in (2.83) are usually referred to as noise floor. BW is the system
2 Eb
BER is directly related with N 0
of the received signal in an additive white Gaus-
sian noise (AWGN) channel. Eb is the energy per bit. N0 is the thermal noise density.
If the date rate is R bits per second and the signal bandwidth is B, then the signal’s
Eb R Eb
SNR can be related to N 0
as: SN R = B N0
. So for a certain BER specification, lower-
ing data rate or increasing signal bandwidth can reduce required SNR, thus improve
sensitivity.
25
receiver, simulation shows that a 12.3 dB SNR is needed for a BER lower than 10−3 ,
and for an 802.11b receiver, an 11.4 dB SNR is required to achieve better than 10−5
BER. Ftot is the system total noise factor and is directly affected by the LNA’s gain
and noise factor. Ftot can be calculated by
Faf terLN A − 1
Ftot = FLN A + (2.84)
GLN A
The above equation shows that the LNA’s noise factor FLN A appears directly in
the system’s noise factor. For high sensitivity, low system noise factor is required,
therefore FLN A should be made as small as possible. The second term of (2.84)
shows that noise coming from the stages following the LNA will be suppressed by the
LNA’s gain, hence a high gain LNA is desirable for high sensitivity. For example, if
the LNA’s noise figure is 1.5 dB, its available power gain GLN A is 10 dB, and the
overall noise figure of the circuits following the LNA is 18 dB, then the system’s total
noise figure can be found using (2.84) to be 8.8 dB. If the LNA’s gain is increased
to 15 dB, the total noise figure will be reduce to 5.3 dB. System sensitivity will be
improved by 3.5 dB. If 8.8 dB noise figure is a fixed system specification, the noise
figure requirement on the circuits after the LNA can be relaxed to 22 dB.
On the other hand, a high gain of the first stage, which is the LNA in this case,
will put a more stringent linearity requirement on the following stages. Therefore
a trade-off must be made between gain, noise, and linearity. Table I summaries
important results obtained in the previous sections.
The RF amplifier is a non-linear system in nature. If the input signal is small enough,
the circuit can be modeled using a linear model around its operating point. But if the
26
s12 s21 ΓL
Input and output Γin = s11 + 1−s22 ΓL
s12 s21 Γs
reflection coefficient Γout = s22 + 1−s11 Γs
|s11 | < 1
Unconditional stable
|s22 | < 1
for passive source and
K > 1
load termination
1−|s11 |2 −|s22 |2 +|s11 s22 −s12 s21 |2
K= 2|s12 s21 |
F = Fmin + Rn
Gs
|Ys − Yopt |2
Two-port network
Fmin = 1 + 2Rn (Gopt + Gc )
noise factor q
Yopt = Gopt + jBopt = G2c + G u
Rn
+ j (−Bc )
γ 1
Rn = α gm
MOS transistor Gc ≈ 0
³ q ´
δ
four noise parameters Bc = ωCgs 1 + α |c| 5γ
(
2 1−|c|2
δω 2 Cgs )
Gu = 5gdo
MOS transistor q ¡
√2 ω γδ 1 − |c|2
¢
Fmin = 1 + 5 ωT
minimum noise factor
27
signal level is relatively high, due to non-linearity, the amplifier’s dynamic operation
point will be changed and become a function of the signal level. The LNA’s proper
operation must be checked by using a large signal input. On the other hand, although
the signal itself is small, large interferers may come together with the signal. This
situation is shown in Fig. 7(a). The interferers can be coming from the adjacent chan-
the two-tone test (f1 , f2 ) as depicted in Fig. 7(b). Usually distortion term 2f1 − f2
and 2f2 − f1 fall in-band, they are characterized by the 3rd order non-linearity. For
example, the desired signal channel in Fig. 7(a) has a bandwidth of 1 MHz and is
centered at 1000 MHz. Two large blockers are located 1 MHz away from the center of
the channel and separated by 1 MHz, i.e. f1 = 1001 MHz and f2 = 1002 MHz. Thus
the lower side IM3 component will be at 2f1 − f2 = 1000 MHz, which is right upon
the center of the channel, degrading the signal’s SNR. Detailed non-linear distortion
analysis can be found in Chapter V. A large in-band blocker tends to desensitize the
circuit, it is measured by the 1-dB compression point. Dynamic range measures the
signal handling capacity of a circuit, which is bounded by the IIP3 and system noise
floor.
The 1-dB compression point (P1dB ) is the point (input or output) where the funda-
mental gain reduced by 1 dB from the ideal small signal gain at a certain frequency
(see Fig. 8). Assume a non-linear system can be approximated by Taylor series
Unwanted
in-band blocker
Desired signal
-23dBm
-102dBm
IM3 fa<2f1-f2<fb
f=2f1-f2
010101010101010100
010101010101010100
010101010101010100
010101010101010100
010101010101010100
fa fb f1 f2
(a)
f1 f2
f1 f2 f1+f2
f2-f1 2f1-f2 2f2-f1
2f1 2f2
(b)
Fig. 7. (a) In-band blockers generate in-channel intermodulation term (b) The two–
tone test spectrum
where α1 and α3 are the 1st-order and 3rd-order coefficients of the Taylor series
2. Intercept Point
The two tone test is usually used to mimic the real-world scenario in which both
a desired signal and a potential interferer feed the input of the amplifier. Due to
29
2SR7TVUOL
HMIENWP
2+? @5AB8DCE=+>F<
=+> GH*I
2+J K$L
HMIONQP
2
3547678:9 ;<
=+>
the non-linearity of the circuit, the 2nd and 3rd order inter-modulation products will
appear at the output and they may lie within the pass band thus, degrading the
signal level. The 2nd order intercept point (IP2) is the extrapolated intersection of
the fundamental curve and the IM2 curve. The 3rd order intercept point (IP3) is the
extrapolated intersection of the fundamental and IM3 curve (Fig. 9).
For the system described by (2.85), the input-referred IP3 (IIP3) [4] is given by
s ¯ ¯
4 ¯¯ α1 ¯¯
IIP 3 = (2.87)
3 ¯ α3 ¯
From (2.86) and (2.87), it is easy to show that IIP3 is about 10 dB higher than P1dB
for the same system if the third-order non-linearity dominates the linearity behavior.
3. Dynamic Range
Dynamic range (DR) is generally defined as the ratio of the maximum input level that
the circuit can tolerate without appreciable distortion to the minimum input level at
30
m7n7o\p+q1rstvu
k c1jg
k c1je
_E` [ a ^ b
XZY\[Z]5^ cFdfe cFdhg
m*w x q1rsOtyu
cic1jle cic1jg
which the circuit provides a reasonable signal quality. Two different definitions are
Fig. 10(a) shows the definition of SFDR. The upper bound of SFDR is based on
test for which the third order intermodulation (IM3) products do not exceed the
noise floor. From Fig. 10(b), it can be shown that the input level for which the IM3
2 × IIP 3 + Nf loor
Pin,max = (2.88)
3
where Nf loor = −174 dBm + N F + 10 log BW , is the noise floor. All the quantities
required signal quality, the minimum SNR is SN Rmin , then the minimum detectable
31
Pout (dBm)
Testing tones
OIP3
SFDR
f | ~
SFDR
Minimum zS{F|F}B~
detectable
signal level IM3L IM3H SNRmin
SNRmin
Noise floor
IM5L IM5H
Pin (dBm)
3f1-2f2 f1 f2 3f2-2f1
2f1-f2 2f2-f1 N floor P in,min P in,max IIP3
(a) (b)
Fig. 10. (a) SFDR defination (b) Relationship between SFDR and IIP3
This is also shown in Fig. 10(b). The SFDR is then calculated by the difference
2
SF DR = (IIP 3 − N F − 10 log B + 174 dBm) − SN Rmin (2.90)
3
input-referred 1-dB compression point and the noise floor as in Fig. 11:
Pout (dBm)
1dB
SNRmin 11.4 dB
BW 6.0 MHz
NF 12.9 dB
The noise floor Nf loor calculated from the above data is -93 dBm. From (2.88) and
(2.89), Pin,max and Pin,min are -40 dBm and -82 dBm, respectively. Thus SFDR is
42 dB. The 1dB compression point can be estimated from the IIP3 which is usually
10 dB larger than P1dB , therefore CFDR is about 90 dB.
4. Wide-Band Non-Linearity
The non-linearities of the wide-band circuits are also measured by their third-order
and second-order linearity behavior. But different from narrow-band system, wide-
band signal occupies a large amount of bandwidth and the two-tone test for narrow-
band system is not sufficient. For example, the cable TV (CATV) system has many
equally spaced carries, the linearities are measured by the composite triple-order beat
33
(x1 + x2 + x3 )3 = H3 + IM 3 + T B (2.92)
where
H3 = x31 + x32 + x33 (2.93)
T B = 6x1 x2 x3 (2.95)
The third order harmonic products in H3 are at three times of frequency. There are
N of them for N tones or carriers. These products are 15.6 dB weaker than the triple-
beat term in (2.95) and do not fall near a carrier, thus they usually are ignored. The
third order intermodulation products in (2.94) are half of the magnitude of the triple-
beats and they are fewer in number. In a system with 20 channels, the contribution
of the IM3 terms is less than 0.1 dB and their contribution decreases with increasing
number of channels.
The triple-beat (TB) term creates spurious frequencies at f1 ± f2 ± f3 (f1 < f2 <
f3 ) and those spurs are 6 dB stronger than the third order intermodulation products
in (2.94). Fig. 12 illustrates the triple-order beats generation. The beats are drawn
offset from the carriers for showing purpose, they are actually reside on the carriers.
In the case of N tones/carriers, the total number of composite distortion products can
N2
be approximated by 4
at the edge of the band and 83 N 2 at the middle of the band. So
triple-beats dominate the third-order distortion performance in a wide-band system.
The triple-beat power in dBm can be related to the intercept point power PIP 3
34
x1 x2 x3
f0 f1 f2 f3 f4
f1-f3+f2 f1+f3-f2 f3+f2-f1
The CTB is usually measured by how many dB the triple-beat down from the
signal power. Because the mid-band has the maximum number of beats, so
µ ¶
3 2
CT B = 2 (PIP 3 − Ps ) − 10 log N −6 (2.97)
8
Composite second order distortion (CSO) is a result of one or two carries expe-
tem having fL as its lowest channel, fH as its highest channel and d is the frequency
offset from a multiple of 6 MHz (1.25 MHz), then the number of second-order beats
f − 2fL + d
NB = (N − 1) (2.98)
2 (fH − fL )
Each of the above second-order beats is an IP2 tone, so its power can be expressed
in dBm as
The CSO is specified as how many dBs the beats power down from the signal power
Note that all the quantities are measured at the output of the system. A typical value
of CTB and CSO for a CATV system would be -97 dBc and -89 dBc.
Non-linear system
y (t) = α1 x (t) + α2 x2 (t) + α3 x3 (t) + · · ·
characteristics
Non-linear circuit
r ¯ ¯
P1dB = 0.145 ¯ αα13 ¯
¯ ¯
1dB compression point
Non-linear circuit
r ¯ ¯
IIP 3 = 43 ¯ αα13 ¯
¯ ¯
3rd order intercept point
2
SF DR = 3
(IIP 3 − N F − 10 log B + 174dBm)
System dynamic range
− SN Rmin
Wide-band system ¡3 2
¢
CT B = 2 (PIP 3 − Ps ) − 10 log 8
N −6
triple-beat distortion
Wide-band system
CSO = PIP 2 − Ps − 10 log NB
second order distortion
36
Before continuing the discussion of LNA topologies, we will review the LNA speci-
fications in some wireless standards. For an LNA used in a GSM mobile receiver,
it will typically require about 20 dB gain, less than 2 dB noise figure, better than
-10 dBm IIP3. Its power consumption should be minimized and return loss should
no less than -8 dB. For the LNA used in a Bluetooth low-IF receiver, it will need to
have a voltage gain of 18 dB, noise figure of 3.5 dB and IIP3 of -3.5 dBm. The power
consumption also needs to be minimized and input return loss should be better than
10 dB. The challenge of LNA design is to fulfill all the specifications with limited
and there is not much degree of freedom to form different architectures. Still, for a
fixed technology, different circuit structures will produce a different performance and
matching is a must requirement for all the LNA’s listed in Figs. 13. The LNA struc-
tures are distinguished from each other by how the input impedance matching is
achieved.
4γ 1
F ≥2+ (2.102)
α gm R s
37
VDD VDD
ZL RL
OUT
VBB RF
M2 OUT
VDD
IN IN
M1 VBB M1
RL
Rs IN R1
M1 OUT
VDD
ZL VDD VDD
OUT
M2 RL
VBB
M2 IN OUT LF OUT
Lg Cgd
IN M1 CF
M1
Cgs M1
+ IN
M3 _ VBB
Ls Ls
Fig. 13. Popular single-ended CMOS LNA topologies (a) Resistive termination (b)
Common gate (c) Shunt-series feedback (d) Inductive source-degeneration (e)
Current-reuse (f) Cgd neutralization
38
So the noise figure is readily larger than 6dB. The primary contribution of noise comes
from the termination resistor and transistor drain noise. Only in very rare cases, is
this LNA structure employed and usually not referred to as an LNA anymore.
The difficulty of input impedance match comes from the high input impedance if
a common source configuration is used. Common gate configuration avoids the high
gate input impedance as in Fig. 13(b). For the first order approximation, the real part
1
of input impedance is just gm
. By carefully choosing the size of the transistor and
γ
F ≥1+ (2.103)
α
This bound is about 2.2 dB and 4.8 dB for a long and short channel device respectively.
The induced gate noise will make the noise factor larger, but the drain noise is still
the dominant factor.
Fig. 13(c) utilizes negative shunt feedback to modify the input impedance of a
RF
Zin = (2.104)
1+A
where A is the voltage gain from input to output and is approximately in the order of
RL
R1
, assuming M1 ’s gm is large enough. The noise figure of this structure is far better
than that of 13(a) but it is still too high to use in some applications. A noise factor
expression for this LNA without the source resistor R1 is given as [7]
µ ¶2 µ ¶2
Gs + G F Gs + g m
F = 1 + Rs δgg + Rs (GL + γgdo ) + Rs G F (2.105)
gm − G F gm − G F
where Gs is the conductance of the signal generator, GF = RF−1 and GL = RL−1 . For
lower power consumption, drain noise is the major noise contribution, but for high
39
power consumption, gate noise and the noise due to RF also become significant.
All of the three architectures introduced above have another similarity. Their
input impedance matching can cover a wide frequency band. Proper input impedance
is required by the RF filter preceding the LNA and the maximum power transfer. The
variations of Fig. 13(a), (b) and (c), are also widely used in wide-band systems [8] [9].
In the above equation, the second term represents the contribution of the drain noise,
the third term is the gate noise contribution, and the last term shows the noise con-
tribution due to the correlation between the gate and drain noise. When considering
R Lg
the series resistance of the inductors used in the LNA, an additional term Rs
should
be added to (2.106). The noise from series resistance of the degeneration inductor
is negligible. Among all the noise contributors, gate noise is the largest one. This
is because the gate noise current sees a high impedance due to the resonance of the
input matching network. Therefore, in order to reduce the noise factor, the Q value
The basic problem with using CMOS transistor for LNAs is its inherently low
transconductance and hence low gain. However, [6] uses a current-reuse technique
40
Fig. 13(e) shows a simplified schematic of this design. The key point is that given
the same bias current, the effective transconductance is gm1 + gm2 , while it is simply
gm1 in the case of no M2 presented. A major drawback of this design is its high input
and output impedances, thus requiring external impedance matching networks. This
prevents the use of this LNA in fully integrated applications Due to the high gain
property, strong Miller effect reduces the reverse isolation of this LNA. In the actual
design, two identical stages are cascaded to improve the reverse isolation.
In order to improve the reverse isolation of the LNA, Cgd neutralization technique
can be used as shown in Fig. 13(f). The LNA’s reverse isolation is limited by the
drain-source parasitic capacitor Cgd . An inductor LF is added in parallel with this
capacitor to provide a different feedback polarity to cancel the effect of Cgd . Care
must be taken to ensure that the inductive feedback does not incur any potential
stability issues.
Table III summarizes the performance of the five LNA circuits discussed in this
section.
In this section, an inductive source degenerated LNA will be designed using a 0.18 µm
CMOS technology to show the LNA design procedure and trade-offs. The simplified
LNA schematic is redrawn in Fig 14. We want to have an LNA working at 2.4 GHz
ISM band with less than 1.6 dB noise figure (noise factor: 1.45), -8 dBm IIP3, 20 dB
(10 V/V) voltage gain and drawing no more than 10 mA current from a 1.8 V power
supply.
The step by step design details from hand calculations to simulation verifications
41
are provided below. A flow chart is given in Fig. 15 to further visualize the design
flow.
Because low-noise is the most important requirement for a LNA, the design consider-
ation will start from the LNA’s noise factor equation. The major noise contribution
ω T Ls = R s (2.107)
42
VDD
LD CD
RL
Vo
VB
M2
Lg
M1
Rs Cgs
VIN Ls
and
1
p = ωo (2.108)
Cgs (Lg + Ls )
where Cgs and ωT are M1 ’s gate-source capacitance and cut-off frequency respectively.
ωo is the operation frequency. The noise factor of the LNA can be shown to be
µ ¶
ωo
F = 1 + κnf (2.109)
ωT
where
γ 1 £
1 − 2 |c| χd + 4 Q2 + 1 χ2d
¡ ¢ ¤
κnf = (2.110)
α 2Q
1
Q= (2.111)
2Rs ωo Cgs
s
δ
χd = α (2.112)
5γ
κnf is called noise factor scaling coefficient. Rs is the source impedance and usually
is 50 Ω or 75 Ω. Q is the input quality factor and also controls the voltage gain from
Rs
Ls = ≈ 0.2nH
ωT
Calculate Lg, LS and
Required Load 1
Lg = 2 − Ls ≈ 26nH
ωo C gs
ω
RL = o AV Rs ≈ 30Ω
ωT
Simulation
Verification
F ↓→ ↑, Q ↑, I den ↓
ωT
ωo NO Meet the
IIP3 ↑→ V gs − Vth ↑ Specs ?
with the above equations. The data can be collected through measurement or simu-
to gdo and changes with biasing condition. Fig. 16 plots the value of α and Vgs − Vth
versus drain current density Ids /W . gm and gdo are also shown on the graph in the
E
¬MV®
«
E
¾Æ
Ä
E Í
Ä ¸Å
ÉÎ Ï
ÂÃ
º §
ÂÍ
¦£§
À ¹ ¸Á ¤B¥
ÉÊ
¼ ÇË Ì
»º¼ ¾
¿ M
¾½ ©$ª
³¶ µ ³ ¯°
±1²´ ¸Å
¨ M
º½¼ ÉÊ
ÇÈ
¹ º» M
·¸
M
M
'E +
EO
O E E
* OE
V ES "VlS 7'¡£¢
Fig. 16. α, Vgs − Vth , gm /W , and gdo /W versus drain current density
Here minimum length of transistor is used. When bias current is low, MOS
transistor can be modeled using long channel approximation, thus α is near unity.
With the increasing of bias current, short channel effect becomes significant rapidly.
α deviates from unity. It is also observed that after current density is greater than
constant and assumed to be about 3. δ/γ is also assumed to be constant and equal
45
to 2.
The next two parameters are the gate-source capacitance normalized to gate
width, i.e. Cgs /W , and device cut off frequency fT when transistor is biased in
saturation region. Fig. 17 shows the curves of Cgs /W and fT versus gate over drive
ÓOÑ ÐO× Ó
ÖMÓ
ÐO× Ö ü
üû
ÐO× Õ
ÖMÑ ñø
øù ú
ÐO× Ô
ÕOÓ ÿî
óõö÷ ÐO× Ð ÷
ÕOÑ óõ
í ÐO× Ñ ó
óô
ñò Ô*Ó
Ñ× Û õö
ññ ð
Ô*Ñ ÿî ý
î\ï Ñ× Ù
ýö
ìí ÐÒÓ Ñ× Ú ì ýþ
ÐÒÑ Ñ× Ø
ÑM× Ñ Ñ× Ô Ñ× Ö ÑM× Ø Ñ× Ù Ð× Ñ Ð× Ô
ÜhÝSÞ1ßDà+áÒßOâVãâ1ä á+ßåáÒàOæ Þ\Ý+ç
ßéèiêë
Fig. 17. Cgs /W and fT versus gate over drive voltage Vgs − Vth
Some insights can be gained from these plots. fT increases with Vgs − Vth when
Vgs −Vth is small. When Vgs −Vth becomes larger than 0.3 V, short channel effects such
as mobility reduction and velocity saturation make gm increasing slowly with Vgs −Vth
and finally reaching a saturation value. Therefore the speed of fT increment with Vgs −
Vth is also reduce. When gate overdrive is small, transistor is biased at weak inversion
region, the charge controlled by the gate is sparse thus gate-source capacitance is
small. Channel charge increases rapidly with gate overdrive and so does the gate-
source capacitance. After Vgs − Vth becomes larger than 0.2 V, Cgs /W increases with
gate overdrive very slowly. At high gate overdrive, gm is almost constant and Cgs
46
increases slowly, so fT begins to degrade gradually after Vgs − Vth is larger than 0.8 V.
Now the noise factor scaling coefficient can be plotted against Q and drain current
density Ids /W as illustrated in Fig. 18, where (a) is the 3-D plot for visual inspection
and (b) is the 2-D plot for design lookup purpose. Generally, the larger the current
density, the smaller the scaling coefficient; and the larger the Q, the larger the larger
the scaling coefficient. For a really large current density there is an optimal value of
Q which minimize the scaling coefficient. But the current density maybe too large
for practical applications. It is also worth to notice that for a fixed current density,
The last plot is not related directly to noise factor but to the linearity perfor-
mance. Fig. 19 is the IIP3 versus gate overdrive at 2 GHz for different size of devices.
Note that IIP3 is more dependent on gate over-drive voltage than device size.
Because we have a small current budget, the gate over drive can not be very large.
It will probably some where between 0.2 V to 0.4 V. Here we do not intend to fix
the gate overdrive voltage but rather have a rough idea of what value of fT we can
use. From Fig. 17, fT is estimated to be about 40 GHz. For 1.6 dB noise figure, from
(2.109) we need κnf to be no larger than 7.5:
fT
κnf = (F − 1) = 7.5 (2.113)
fo
For 0.2 V to 0.4 V gate overdrive, from plots in Fig. 16, the current density is some-
where between 60 µA/µm and 140 µA/µm. From Fig. 19, the device raw IIP3 is from
7 dBm to 14 dBm, so if Q is chosen to be 4, IIP3 will have good change meet the
47
18
16
Noise factor scaling coefficient
14
12
10
600 500 6 8
400 300 200 4
100 2
Current density (µA/µm) Quality factor
(a)
HI.:J: KL4NMKL1O3 4+5
! "
C
=F
@@ = ?B
<? ## ! "
B
E = FGA $&%' ! "
>B
C <D $(#*) +"
@ AB
= >? %,, "
;<
-/.102 3 4+576+018*4
9:
(b)
Fig. 18. Noise factor scaling coefficient versus quality and current density for 0.18µm
NMOS device (a) 3-D plot (b) 2-D plot
48
specification. From Fig. 18(b), for Q of 4, the current density is chosen to be about
1
Cgs = ≈ 166 f F (2.114)
2QRs ωo
From Fig. 17, the gate-source capacitance density is about 1.3 f F/µm, so the required
device width will be
166 f F
W = ≈ 128 µm (2.115)
1.3 f F/µm
The device current can be estimated by
Under the above conditions, the transistor’s gate overdrive voltage is about 0.23 V
gm
ωT = ≈ 2π × 48 GHz (2.117)
Cgs
Considering the effect of gate-drain capacitance Cgd , the ωT is pretty near the value
Rs
Ls = ≈ 0.2 nH (2.118)
ωT
1
Lg = − Ls ≈ 26 nH (2.119)
ωo2 Cgs
The cascoded transistor M2 will use the same size as the main driving transistor and
its size can be optimized through simulation if required. The gate bias voltage of M 2
is tied to power supply in this example, its value can also be optimized. Generally,
it is chosen just large enough to bias M1 in saturation region for all possible input
levels. The use of M2 reduces the Miller effect and improves reverse isolation of the
LNA.
In order to obtain a desired voltage gain, the load impedance has to be set
correctly. Assuming the load is LC tuned and its equivalent impedance around ωo is
dominated by load resistor RL . The voltage gain of the LNA can be shown to be
µ ¶
ωT RL
Av = jQgm RL = j (2.120)
ωo Rs
50
Now the design calculation is finished. But most of the design parameters need to be
fine tuned through circuit simulation. The above procedure shows the design trade-
offs between all the design parameters and can be used to guide circuit adjustment in
simulation. The initially calculated and simulated final device parameters are listed in
Table IV. Table V compares the targeted specifications and simulation results. The
hand calculations match the simulation results well except that the gate inductance
is much more over estimated by hand calculations. Fig. 20 shows the simulation plots
of various parameters of this LNA. Note that it is assumed that high quality on-chip
inductors are available in our design, their Q values can be as large as 20 at the
working frequency.
A differential LNA can be designed using the same procedure provide above. A
MOS differential pair should be investigated in order to obtain the required design
plots. Detailed discussion of differential LNA design issues will be given in Chapter IV.
51
W 128 µm 127.5 µm
Ids 8.9 mA 8 mA
Ls 0.2 nH 0.2 nH
Lg 26 nH 16 nH
RL 30 Ω 40 Ω
Voltage gain 20 dB 21 dB
S11 – -17 dB
S12 – -25 dB
25.0 0.88
20.0 AV
15.0 0.86
10.0
0.84
5.0
0.0
0.82
-5.0 NF
-10.0
S11 0.80
-15.0
-20.0 0.78
2.20 2.25 2.30 2.35 2.40 2.45 2.50 2.55 2.60
20
P1dB=-20 dBm
0
Output Voltage (dBV)
-20
IIP3=-6.4 dBm
-40
-60
-80
-100
-40 -35 -30 -25 -20 -15 -10 -5 0
Input Power (dBm)
(b)
Fig. 20. Simulation plots of the designed LNA (a) Voltage gain, S11 and noise figure
(b) IIP3 and P1dB
53
CHAPTER III
BLUETOOTH RECEIVER
In the receiving path (Rx) of a communication system, one of the goals is to pick
up the useful signal from the presence of noise and interferer, i.e. select interested
channel band. While it is hard to design a filter with a 1 MHz band centered at
2.4 GHz (Q∼2400), it is relatively easier to select a 1 MHz band centered at 4 MHz
(Q∼4). The down-conversion mixer is the key block that transfers the spectrum from
the transmission of the baseband signal efficient and practical, the low frequency band
usually needs to be transferred to a much higher frequency band. Thus, the signal
can be radiated out by a compact enough antenna. Up-conversion mixer does this
job. It translates the low frequency baseband spectrum to the RF spectrum, then the
power amplifier amplifies this high frequency signal to a sufficient amount of power
and radiates out by the antenna. Fig. 21 is the simplified block diagram of a low-
IF/zero-IF transceiver. It shows the position and roles of the up- and down-conversion
mixers.
Most mixer implementations use some kind of multiplication of two signals, the signal
to be up- or down-converted (IF or RF) and the signal whose frequency determines
Down-Conversion
Mixer
LNA Rx Baseband
LO
Rx/Tx
Switch
PA Tx Baseband
Up-Conversion
Mixer
Frequency
RF Portion Conversion IF Portion
v RF Mixer Output
v LO
Assume the RF input signal is vRF (t) = ARF cos (ωRF t) and the LO signal is
vLO = ALO cos (ωLO t), then the output of an ideal multiplier is
ARF ALO
vIF = vRF (t) × vLO (t) = [cos (ωRF − ωLO ) + cos (ωRF + ωLO )]
2
For an up-conversion mixer, component ωRF − ωLO is filtered out. For a down-
conversion mixer, component ωRF + ωLO is filtered out. If the input signal occupies a
frequency band, that band of frequency will be moved to a lower frequency for down-
considering this mixing model in frequency domain. Usually the ideal LO signal is a
single tone, vLO = cos (ωLO t). Here the amplitude of the LO is normalized to unity.
The RF signal resides within a certain band around its center ωRF . The ampli-
fication in time domain will become convolution in frequency domain, i.e., the RF
signal will convolve with two δ-functions in frequency domain. Fig. 23 demonstrates
this process. It can be seen that the RF frequency band is translated into side-band,
the lower side-band centered around ωRF − ωLO and the upper side-band centered
around ωRF + ωLO .
VLO(ω ) VRF(ω )
ω ω
−ωLO 0 ωLO −ωRF 0 ωRF
VIF( ω)
0 ω
ωRF − ωLO ωLO − ωRF ωRF − ωLO ωRF + ωLO
B. Mixer Metrics
In order to evaluate the performance of mixers, several metrics are defined, they are:
conversion gain or loss, noise figure, port isolations, linearity and power consumption,
etc.
56
Conversion gain or loss is a measure of mixer efficiency, it is defined as the ratio of the
desired IF output (voltage or power) to the RF input signal value (voltage or power).
More specifically:
gain and power gain. At most a passive mixer can only provide voltage or current
gain but not power gain. The parametric converter, which is a special kind of passive
mixer, can provide power gain.
2. Noise Figure
A mixer’s noise figure (NF) is the signal-to-noise ratio (SNR) at the input (RF)
port divided by the SNR at the output (IF) port. Although the mixer’s noise figure
definition looks similar to the LNA’s noise figure definition, there are some subtle
eration frequencies. Mixers perform frequency translation, so there are two different
types of noise figures according to two different types of frequency conversion schemes,
single-side band (SSB) noise figure and double-side band(DSB) noise figure.
For the DSB noise figure, the desired signal appears at the both sides of the LO
57
frequency. Fig. 24 gives the situation of the double-side band conversion. Assume
the RF signal power for each side is Si , the noise power for each side presented at the
RF input is Ni , the conversion power gain is G, the mixer internal noise referred to
its output is Nno , the noise factor of the double-side band mixer can be calculated as
For the SSB noise figure, the desired signal only appears at one side of the LO
2S i G
Desired Signal
RF IF N no
Si Si
Ni
2N i G
f RF1 f LO f RF2 f IF
LO
frequency. Suppose the noise at image frequency has not been removed, Fig. 25
shows this scenario. Using the same notation as the calculation of DSB noise factor,
the SSB noise factor can be expressed as
It shows that the SSB noise figure is 3 dB higher than the DSB noise figure. When
designing a mixer, it should be made clear which noise figure is targeted at. The
simulator like SpectreRF gives the SSB noise figure. So if the DSB noise figure is
required, a 3-dB difference should be subtracted manually.
In practical system, the image noise in a SSB case is filtered using an image
58
S iG
Desired Signal
RF IF N no
Si
Ni
2N i G
f RF1 f LO f RF2 f IF
LO
rejection filter (IRF) before the mixer or a complex filter after the mixer (see Fig. 26).
So, the image noise experiences different gain with the noise within the signal band.
Suppose the gain of the IRF or complex filter at image frequency is Gim , the noise
factor is
Si Ni G + Ni Gim G + Nno Nno
FSSB,IRF = × = 1 + Gim +
Ni Si G Ni G
With related to SSB or DSB noise factor,
e¡ ¢1£&¤ ¥ ¢¦¨§X¤ ©ª¬«X o
*
\
l
&o Xwu e / X
®
¯
3. Port-to-Port Isolation
In reality, signals always leak through different mechanisms from one port to another.
The port-to-port isolation figure accounts for this unwanted transmission, it is defined
as the ratio of the signal power available into one port of the mixer to the measured
power level of that signal at one of the other mixer ports in a 50 Ω system.
The mixer has three ports. The desired transmission is from RF port at RF
frequency to IF port at IF frequency. The leakage can occur among all the other
The signal leaked from the LO port to the RF port at a RF frequency (LO-to-RF
leakage) will mix with the LO signal again, causing the so called self-mixing problem
in direct conversion. Due to the non-zero reverse gain (s12 ) of the LNA, the LO
leakage may even reach the antenna through the LNA, then reflected back by the
RF IF
LNA
LO
The LO-to-IF feed-through may cause desensitization of the block following the
mixer which is usually a low pass filter. In order to guarantee a good mixing per-
formance, the LO power is usually greater than the IF power. Although the LO
frequency will be in the stop-band of the filter, the large LO signal will drive the
60
filter out of its desired operation region. Whereby it is desirable to have some passive
The RF-to-LO feed-through will allow the interferer and spurs present in the RF
signal to interact with the LO. The RF-to-IF feed-through at the IF frequency may
intermodulation product.
4. Linearity Measurement
When a real mixer operates, not only do the desired tones mix (multiply) each other,
their harmonics also experience the mixing process due to the non-linear characteristic
of the mixing device. These unwanted products may fall into the spectrum band of
the signal and degrade the desired signal. The non-linear behavior of the system is
frequency dependent. We are interested in the dynamic linearity of the system, not
just the DC non-linear characteristics.
mixer’s linearity: 1-dB compression point (P1dB ), second and third order intercept
point (IP2 and IP3), spurious free dynamic range (SFDR) and compression free dy-
namic range (CFDR). The meanings of these terms for mixer are similar as for RF
spectrum from one location to another. Linear, time-invariant systems can not gener-
ate spectral components not presented in the input. The mixer must be a non-linear
or time-variant system. Virtually any nonlinear elements can be used as mixers. Some
61
nonlinearities just work better and more practical than others. The mixer can be pas-
sive or active depending on whether the mixer can provide conversion loss or gain.
signal domain.
1. Diode Mixers
Diode mixers use the non-linearity of the diode to implement the frequency conversion.
Fig. 28 shows the single-diode mixer.The single-diode mixer is the simplest and oldest
passive mixer. The output LC tank is tuned to the desired IF, and the input is the
sum of RF, LO and DC bias. This mixer can not provide any isolation between
ports, neither can it achieve conversion gain. However, at very high frequencies (e.g.
millimeter-wave band) this kind of mixer is extremely useful.
VIN VOUT
L C R
diodes work as switches, regardless of the level of the RF signal. When the diodes
are on, RF and IF are connected together, so the RF-IF isolation is poor. But the
RF signal is common-mode for the transformer, so the RF-LO isolation is excellent.
62
LO IF RF
A double-balanced diode mixer uses four diodes. Due to the symmetry of the
circuit, isolations between each pair of ports are excellent, mainly limited by the
device matching. The diode mixer is almost completely linear and the upper limit of
LO
IF
RF
This kind of mixer is more like the double-balanced diode mixer. The diodes are re-
placed by MOS transistors working as switches as shown in Fig. 31. MOS transistors
63
M1-M4 are working as switches and are driven by a large LO signal in the anti-phase.
Thus, only one diagonal pair of transistors are turned on at any given time. More
specifically, when the LO is high, M1 and M4 are on, VIF equals to VRF , and when
the LO is low, M2 and M3 are on, VIF equals to −VRF . So it is equivalent to treat
C1 L1
´µ
¶ L3 LO M1 M2 LO
V IF
°²±R³ RG C3
CL
´µ
¶ L4 LO M3 M4 LO
C2 L2
In Fig. 31, the input LC network provides matching and filtering. And due to
the reactive matching network, the voltage conversion gain can be greater than 1,
but this mixer can not provide power conversion gain. The noise figure and IIP3 are
strong functions of the LO signal strength and depend on how the MOS transistors
are switched. The implementation of a passive mixer is usually simple and there are
no stability issues involved [13].
64
3. Gilbert-cell Mixer
Gilbert cell mixers are the most popular types of integrated mixers. They can be
port isolation (40∼60 dB) can be achieved as a result of circuit symmetry. If the
input signals (RF and LO) are small enough such that all the transistors in Fig. 32(a)
are working in their linear region, then the Gilbert-cell will behave like an analog
multiplier. But this is not an efficient way to perform the RF frequency conversion
using a Gilbert-cell, it will generate a prohibitive high noise figure and has a strong
LO dependence of conversion gain.
IF
IF
M3 M4 M5 M6 IF
LO
LO LO LO
LO M3 M2 LO
M1 M2
RF
GmVRF -GmVRF
RF M1
Fig. 32. Gilbert-cell mixer (a) transistor implementation (b) working principle and (c)
single-ended version
are driven by a large enough LO signal such that they work as current commuting
65
switches. The linearity of the mixer is limited by the linearity of the V-I converter.
Additional linearization techniques are usually applied to the V-I converter to improve
the linearity of the mixer. For direct conversion and low IF, the noise figure is limited
by the flicker noise of the current switches and for higher IF, the noise figure is
limited by the thermal noise of the circuit. The transconductance conversion gain
can be expressed as
2
G c = α gm (3.6)
π
where gm is the transconductance of the V-I converter. Factor α is referred to as
switching efficiency. Because the practical mixer’s switches are usually not ideal,
smaller than its ideal value one. In some cases, single-balanced mixer can fulfill the
requirement, then half of the Gilbert-cell can be used as shown in Fig. 32(c). This
mixer consumes half of the power, has single-ended input and differential output. It
4. Sub-Sampling Mixer
shows a general structure of this kind of mixer. One big advantage of this circuit is
that the LO signal is clocked at a relative low frequency, which eases the design of LO
circuitry. But the sampler still must have a good time resolution which means the
clock absolute time jitter must be a tiny fraction of the carrier period. Due to the sub-
sampling, the noise folding effect makes the mixer present a large noise figure. The
linearity of the sub-sampling mixer is usually very high, but due to jitter noise and
thermal noise folding, its dynamic range may be inferior to other carefully designed
of the operational amplifier (Opamp) used in the sampler. The limited achievable
gain-bandwidth (GBW) of the Opamp will limit the RF input signal to a certain
range.
φ2
M6
VCM
φ3 φ 1_b M3
M1
RF φ1 M4
M2
φ 1_b M5
φ3
VCM
M7
φ2
5. Harmonic Mixer
Unlike conventional mixers, harmonic mixers mix RF signals with the second or higher
order harmonic of the LO signal. For standard mixers, the mixing product of interest
The idea of harmonic mixing is to use the even harmonics of LO for its conversion
product. The odd harmonics including its fundamental will be rejected either due to
the odd symmetry of the system or by filtering or both.
One direct benefit of this idea is that the LO can run at half rate, which makes
67
VCO design easier. But because of the harmonic mixing, the conversion gain is usually
small (several dB) and the noise figure is high. Harmonic mixers are attractive for
the direct conversion applications due to the fact that they have low self-mixing DC
offset.
Fig. 34 is one possible implementation of the harmonic mixer [14]. Two emitter-
coupled BJT pairs work as two limiters. The odd symmetry of their transfer function
suppress even order distortion including LO self-mixing. The small RF signal will
modulate the zero crossing point of the relatively large LO signal. The output of
the mixer is a rectangular wave in pulse width modulation fashion, a low pass filter
will demodulate the signal. The leaked LO signal will generate a position modulated
signal which can not be demodulated by the low pass filter. So ideally this mixer will
VCC
RL RL
IF
RF RF
LO
links between mobile PCs, mobile phones, PDAs and other portable devices. It oper-
68
ates at the 2.4 GHz Industrial Scientific Medicine(ISM) band. Its modulation format
is Gaussian frequency shift keying (GFSK) with an index of 0.28∼0.35. The data rate
is 1 Mb/s and the channel spacing is 1 MHz. Low cost and possible system-on-chip
(SoC) solutions are the most attractive features of Bluetooth technology. The digi-
tal system is an important part of the system and digital technology benefits most
by using CMOS process. So the CMOS process is a must for SoC implementation.
Comparing to the BiCMOS technology, CMOS process is preferred in the Bluetooth
level. Thanks to the recent development in the silicon fabrication technique, the
current CMOS technology is able to accommodate the high speed RF circuit design
in the GHz frequency band and hence can be applied in the Bluetooth receiver de-
sign. Therefore, TSMC 0.35 µm CMOS technology is chosen to realize the Bluetooth
receiver system.
For a fully integrated CMOS implementation, there are two architecture can-
didates, direct conversion and low-IF. Direct conversion has the most possible high
level integration and does not require image rejection. It needs less components and
can achieve low power consumption. The difficulties faced by direct conversion are
DC offset and flicker noise [15]. While DC offset can be removed by some dedicated
circuitry, the flick noise presented in the mixer prevents one from adopting this archi-
tecture. The process, TSMC 0.35 µm CMOS, used in this implementation has flicker
noise corner frequency around 1∼2 MHz. It is not possible to make the mixer’s noise
figure meet the specification for direct conversion. A low-IF receiver can also achieve
a high level integration and has a possible low power requirement with careful system
level and circuit level design [16]. For low-IF, flicker noise is less significant in the
signal band. The DC offset can be easily removed with relatively simple circuitry.
But one needs to consider the image rejection and folded-back interference problem.
69
is decided to be 2 MHz due to the high flicker noise corner frequency. The image and
Fig. 35 shows the low-IF Bluetooth receiver block diagram. It has two identical mixers
for the I and Q branches. The mixers in this receiver convert the 2.4 GHz frequency
band RF signal to a 2 MHz band IF signal, then the down-converted I and Q signals
are fed to a complex filter to make image/interferer rejection and channel selection.
The specifications for the mixer required in this receiver are listed in Table VI.
For the front-end of the Bluetooth receiver, the LNA is designed by designer Wenjun
Sheng using the conventional source inductive degeneration technique [17]. The down-
70
Complex
Filter
I Mixer
RF
Cancellation
Demodulator
Filter
DC Offset
Limiter & Binary
GFSK
LNA 0
90
RSSI Stream
Q Mixer
Synthesizer
& VCO
conversion mixer is a modified Gilbert-Cell mixer [18], as shown in Fig. 36. In order to
improve the output dynamic range, the tail current in the conventional Gilbert-cell is
removed, this also allows low voltage operation and has higher linearity. It is obvious
that without the tail current, the voltage headroom will be improved by a V dsat of a
MOS transistor if a simple current source was used. The improved linearity without
sistors in the saturation region, the input differential voltage should be within [19]
√
|Vid | < 2Vod (3.7)
where Vod is the gate over-drive voltage when Vid = 0, i.e. at quiescent DC biasing
point. But if there is no tail current source, for the same DC biasing current, the
differential input range for active operation will be expanded to
The differential pair has less linear input range due to the existence of the tail
current source. The current source makes the two transistors in the differential pair
71
ÀÂÁ¿Á
½l× ½×
Ø × Ø ×
ÀÂÁ¿Á
à ¾h¹ à ¾Ä»
ÃÅ ÃÅ
·¸¼»
ÒºÓ ÒºÔ ÒÂÕ ÒºÖ
·¸º¹
½Å(È ½lÅ(È
ÀÇÅ(È
ÉÊ`ËÌ
ÍÏÎ Ð+Ñ
interact with each other so the their active input region is limited more than two single
transistors without such interaction. However, without tail current source, the circuit
is not a truly differential structure any more which may have a poor common-mode
rejection performance. Fortunately, if the LNA’s output signals are well balanced
and/or additional loop is used to suppress the common-mode signal, this architecture
can be used.
The injected current IB , called bleeding current, allows separate control of current
72
flowing through the drive stage (M1 and M2 ) from the current switches (M3 ∼M6 ).
For the same drive bias current, bleeding reduces the current flowing through the
current switches and load resistance. Since flicker noise can be expressed as [20]
A
ID f
i2nf = Kf ∆f (3.9)
Cox Wef f Lef f f
Reducing the bias current will significantly reduce flicker noise contributed by the
switches and allow large load resistors which increases the conversion gain. Fur-
thermore, with current bleeding, the switches can work with smaller gate to source
voltage, so for a given LO signal level, smaller charges are necessary to turn on or off
RF drive stage, because bleeding reduces the bias current of the switches therefore
increasing the load impedance seen by the driving transistors. Ideally, the RF drive
stage works as a voltage-current converter, it should see as small impedance at its
ciency. Due to Miller effect, increased voltage gain from the input of the RF stage to
which is usually implemented using active devices. In this implementation, the bleed-
ing current source is formed by a PMOS current source with a large device size. This
arrangement makes the flicker noise contribution from the bleeding source smaller.
Care must be taken to make the matching between the two bleeding current sources
as good as possible. Otherwise, a large mismatch will drive the injection node to-
wards the positive or negative power supply, if the switch current is too small to set
can be used to prevent this problem [21], it will require an additional tail current
73
source which is removed for the limited voltage headroom reason. Since the load of
the mixer is resistive, the need of a common mode feedback circuit at the mixer’s
output is avoided.
The voltage conversion gain can be written as
2
Avc = gm,RF RL (3.10)
π
perfectly matched, then gm,RF = gm,M1 = gm,M2 , otherwise for the first order approx-
gm,M1 + gM,M2
gm,RF ≈ (3.11)
2
For a regular differential pair with tail current source, and using the same transistor
which is by a factor of √1 smaller than the one used here without a tail cur-
θVod
rent source. It is obvious from (3.10) and (3.12) that an increase in the drive stage
transconductance will increase the conversion gain and linearity at the same time.
But increasing gm will result in a larger power consumption. The noise analysis of
the mixer is quite involved and can be found in [23].
The mixer’s design flow is shown in Fig. 37. Because the signal reaches the mixer’s
input after being amplified by the LNA, we start with the linearity specification for
74
the RF driving stage. From (3.12), the required gate over-drive voltage V od, RF can
be calculated, then from (3.10), the device size WRF , bias current IRF and load
the LO swing should be larger than the switch differential pair’s linear operation
√
voltage range, which is 2Vod, SW . Therefore, the gate over-drive of the current
switches Vod, SW , and their size WSW can be calculated. After the hand calculations,
adjusted intensively through simulation and several iterations are required to achieve
When doing the layout of the mixer, special care should be taken to ensure the I and
Q mixers are symmetrical and the gate resistance is minimized. Matching techniques
such as common centroid and inter-digitized pattern are applied. The length of the
poly gate is kept short enough to reduce the gate resistance (large gate resistance
will degrade the noise performance). For the layout of a poly-poly capacitor, if the
bottom plate is floating, the parasitic capacitance from the bottom plate to the sub-
capacitors may be needed to prevent the circuits from oscillation. Metal should be
wide enough to carry large current. The current density allowed through metal is
about 1 mA/µm. Guarding rings are place around the circuits to improve isolations
from other blocks on the same die. The mixer uses resistors as load, guidelines for
reducing resistor mismatch are: 1) Use dummy poly lines when possible. 2) Maximize
the number of contacts per width of the device. 3) Avoid partial coverage of resistors
- Noise figure
Design Specs - IIP3
- Voltage gain
- Power consumption
RF Driving Stage:
2 V od , RF
V IIP 3 = 4 → Vod , RF
3 θ
2
AVC = g m , RF RL → WRF , I RF , RL
π
Bleeding Current
RL , I RF → I B , I SW
Current Switches
Simulation
Verification
NO YES
Meet the Layout and post
Specs ? layout simulation
NO Meet the
Specs ?
YES
Tape-out
identical resistors, resistors at the edge showed about 2% difference in value from
their nearest neighbors. 5) Gradient effects are present and do effect matching, keep
circuit is simulated using Cadence SpectreRF environment. Because the mixer will
be interfaced with LNA directly, its input impedance is first obtained around its
operation frequency. And is given to the LNA designer. The LNA and mixer co-
simulation is done to make sure they are compatible. Mixer’s separate performance
I Mixer Q Mixer
Bias
RF-LO isolation 30 dB
LO swing 0 dBm
Supply Voltage 3 V
Current consumption 5 mA
The Bluetooth receiver IC is fabricated using the TSMC 0.35 µm standard CMOS
process through MOSIS service and packaged in a 48-pin TQFP plastic package. The
die microphotograph is shown in Fig. 39, and it occupies 2.5 mm × 2.5 mm silicon
area.
with the LNA. In order to guarantee the performance, no access to the internal high
frequency nodes within the LNA and mixer is granted. Therefore, the LNA and mixer
have been tested as a single block. Together they consume 10 mA current from a 3 V
power supply.
The measured cascaded NF and voltage gain are 8.5 dB and 25 dB, respectively.
Fig. 40 shows the input S11 measurement plot of the front-end. The S11 is less than
-10 dB in the 2.4GHz band. The system IIP3 is about -10 dBm. Fig. 41 is the IIP3
78
plot of the receiver system. The IIP3 was tested using two tones 3 MHz and 6 MHz
away from the desired signal. The 3rd order intermodulation product was observed
at the output of the complex filter and referred back to the receiver input to calculate
IIP3. The achieved sensitivity is -82 dBm for the whole receiver.
79
åXé*ê
å æè
äåælç
CHAPTER IV
With the rapid development of wireless access technique, more and more devices
are Bluetooth enabled. At the same time, the wireless LAN (IEEE 802.11b, Wi-Fi)
standard comes into our everyday life. While Bluetooth targets on short range cable-
PDAs [25]. These two standards do not compete with each other, but rather comple-
ment each other [26]. They do different things and tasks. There are circumstances
where Bluetooth is preferable to Wi-Fi, for example, short range point to point or
multi-point data transfer with low power consumption, data self-synchronization and
updating. But for more complicated network functionalities and high speed data
transfer, Wi-Fi will take over. One big application is wireless Internet access. All
those functions are needed in high-end laptop and PDAs. So there are practical
needs to require both standards within one application. Combining these two stan-
dards together into one single chip will reduce the fabrication cost and add more
functionalities to the product.
Both Bluetooth and IEEE 802.11b standards work at the 2.4 GHz ISM band.
Although the base band will be quite different between these two standards, the
co-band operation makes the RF front-end having maximum compatibility. And it
building block sharing. Table VIII lists the key features of both standards.
Bluetooth Wi-Fi
As stated above, Wi-Fi belongs to the wireless local area network standard family.
combine these two standards together to make the product more competitive. In this
design, direct-conversion architecture was explored for both standards to avoid the
standards to save power and silicon area. A major concern in the low-IF architecture
is the selection of the intermediate frequency (IF). In order to benefit the most from
the low-IF architecture, the IF frequency is usually optimized to suite the modulation
format and signal bandwidth defined in a particular standard. Therefore, the IF will
vary as the receiver switches between the standards to avoid significant degradation
82
frequency is located at the IF, becomes difficult to design. In the direct conversion
receiver architecture, the channel select filter will be just a low pass filter. It does not
have a center frequency to adjust when the receiving mode changes. Furthermore, the
signal bandwidth is also different from one standard to the other. The implementation
of a bandwidth variable filter is more straightforward for low pass than that for band
pass.
Although the direct conversion receiver enjoys the high level integration and
difficulties, such as DC offset and low frequency noise. The large DC offset and low
frequency flick noise make it difficult to design a direct conversion receiver in current
CMOS technologies for narrow band systems. Additional circuits and techniques,
flick noise problem in CMOS technology for narrow band systems like Bluetooth.
BiCMOS technology, on the other hand, has much lower intrinsic DC offset and
flicker noise. Simple measures, like a 2nd or 3rd order high pass filter, can provide
enough suppression to the DC offset and flick noise. Therefore, BiCMOS process is
chosen for this dual-standard direct conversion receiver to relax design complexity for
this receiver, the front-end blocks, LNA and mixer, are completely shared between
Bluetooth and Wi-Fi standards. The received signal in the Wi-Fi standard can be
10 dBm stronger than the Bluetooth signal. To accommodate the larger dynamic
range of the 802.11b received signal, the LNA has two gain modes. It switches
An integer-N frequency synthesizer is used to generate two times the desired local
the coupling from the power amplifier to the VCO. The output of VCO is then passed
through a divided-by-two circuit to generate quadrature LO signals for the I and Q
gain amplifier and sent to analog-to-digital converter (ADC). The ADC is a parallel
pipelined structure. For Bluetooth, its sampling rate is 10 MHz, and has an 11-bit
resolution. For Wi-Fi mode, the sampling rate is 44 MHz, and its resolution reduces
to 8-bit. DC offsets cancelation technique is used to remove both static and dynamic
offsets [27].
15dB 18dB
I Mixer LPF
RF ADC
Filter φ
φ Gain Signal Level
LNA 0
Control Measurement
φ
90
φ
ADC
Q Mixer LPF
Synthesizer
Attenuator(-15dB) & VCO
For this design, a fully-differential structure is used to reject the common-mode noise
coming from the digital parts and other sources as much as possible. The differential
structure can also isolate the influence of bond-wire. For a single-ended circuit, bond-
84
wire for signal grounds directly becomes part of the circuit and must be modeled well
bond-wire is in series with the tail current and is in both DC and common-mode
path, so it does not or has little effect on the circuit differential operation. Thus its
twice of the power as single-ended one for the same performance, but the improved
isolation and reliability justifies the power increase.
For a differential LNA, the desired operation mode is the differential mode. But
the common mode stability. Fig. 43 shows the source degenerated differential LNA
input stage, its differential half-circuit and its common mode equivalent half-circuit.
In the figure, Lg is the gate inductor, Ls is the source degeneration inductor, they
together resonate with gate capacitor C1 . C1 is the total capacitance between the
gate and source, it is composed of gate-source overlapping capacitance Cov , gate-
source channel capacitance and additional capacitance added externally. 2C2 is the
parasitic capacitor of the current source. For the differential operation, this capacitor
does not appear in the half circuit. For the common mode, half of this capacitor will
present in series with the inductor Ls . The resistance of the current source is assumed
to be sufficiently large.
For the differential half-circuit, the input impedance [10] can be expressed as
1 gm
Zin,dif f = jω (Lg + Ls ) + + Ls (4.1)
jωC1 C1
1 gm
At the resonate frequency ω = √ , it is a pure positive resistance L,
C1 s
where
(Lg +Ls )C1
ïhð1ð
ïhð1ð
õoö õoö ïhð1ð
õoö
òô òô
õoö òô
õo÷ õo÷
òô õo÷
! " # $
õo÷ òXó
ñòXó ! % & &
C 1 + C 2 gm gm
Zin,com = jω (Lg + Ls ) + + Ls − 2 (4.2)
jωC1 C2 C1 ω C1 C2
It can be seen that if this real part goes negative, then there exists potential instability.
One may want to increase the capacitance of C2 in order to keep Rin,com greater than
zero. This is not the right way. A large common-mode capacitance will disturb the
differential operation and even worse, increase the common-mode gain. We should
keep this capacitance as small as possible. For an ideal tail current source of the
differential pair, the common-mode gain is infinity, so the oscillation will not be
able to sustain. Of course one should check the sign of (4.3) and further evaluate if
this negative resistance will cancel the source impedance’s real part thus causing a
problem.
86
Another concern is the interface between the LNA and mixer. We have a direct
conversion receiver. There is no image rejection filter between the LNA and mixer,
so the LNA does not need to drive 50 Ω impedance and the mixer does not have
to provide 50 Ω input matching. This is another degree of freedom for the direct
conversion front-end design. The design strategy is that the mixer will be designed
first. Then the mixer input impedance becomes known and will be driven by the
LNA. Another way is to have a rough estimate of the impedance level of the mixer
input and then the LNA and mixer can be designed at the same time by different
designers. The LNA and mixer must be brought together to verify their interface
behavior. For off-chip interconnection, the effect of bond wires on the circuit must
be modeled and considered with the circuit design.
BIASING CIRCUIT
RF_IN I+
3GHz I-
1:1 BALUN LNA
Q-
Q+
Quadrature Buffers
LO
C. Circuits Implementations
The dual-mode RF front-end includes the LNA, mixer, buffers and their biasing
circuit. Fig. 44 gives the block diagram of the RF front-end. This is a direct conversion
front-end, so no image rejection filter (IRF) is required between the LNA and mixer.
Table IX lists the specifications of the LNA and mixer for the Bluetooth and
Wi-Fi operation. Notice that Wi-Fi has the most stringent requirement. So the RF
front-end is shared by both the Bluetooth and Wi-Fi modes and targeted at the most
stringent specifications in the table.
1. LNA Implementation
The SiGe BiCMOS technology is used to implement the front-end. The LNA is given
in Fig. 45. It is an inductive degenerated differential structure and can provide two
gain steps of 15dB and -15dB. Gain control is implemented by a differential attenuator
built around the LNA as indicated in the dashed box in Fig. 45. The attenuator
is formed using NMOS transistor M5 ∼ M9 . For high gain mode, all the NMOS
transistors will be turned off by connecting their gates to ground. Thus the normal
operation of the LNA will not be affected. For low gain operation, these transistors are
driven into their triode region and the LNA’s bias current Itail will be cut off. Under
this mode, the LNA itself will not consume current and is by-passed. The resistor
VDD
Ld Ld
Cd Cd
LNA bypass switches and attenuator
Vo+
Vo-
LNA_cas_bias Q1 Q2
M5 M7
Bond wire
Cm M9
Vin+
M6 M8
M1 M2
Vin- LNA_bypass
Rb Rb
Rb Rb Ls Ls
Cm Vbb
i_tail
LNA_rf_bias
M3 M4
All the matching conditions are established on-chip using the inductive source
degeneration technique [10]. The only required off-chip components are a 1:1 RF
BALUN to convert the single-ended signal to differential and two additional chip-
transistors are more linear than bipolar transistors for the same current consump-
tion [28]. The cascoded bipolar transistors Q1 and Q2 provide stable and matched
bias voltage to the drain of NMOS M1 and M2 . Here bipolar is preferred because the
base-emitter voltage change is much smaller than the gate-source voltage change for
the same bias current variation. The transconductance available from NPN bipolar is
much larger than the one from NMOS for the same current. Thus voltage at the drain
of M1 and M2 does not change significantly which minimizes the voltage gain from
89
the gate to the drain of M1 or M2 , therefore the Miller effect is reduced and reverse
isolation is improved. This can be seen more clearly through the input impedance of
α 0 RL
Zcas,in = re + (4.4)
gm,bjt rce
α0 1
where re = gm,bjt
≈ gm,bjt
. α0 is the emitter to collector current ratio and is very near
unity for a modern NPN bipolar process. RL is the equivalent output load impedance.
If the cascoded transistor is replaced by its MOS counterpart, its input impedance
will be
0 1 RL
Zcas,in = + (4.5)
gm,mos + gmb,mos (gm,mos + gmb,mos ) rds
where gm,mos and gmb,mos are the transconductance of a MOS transistor from gate
and bulk terminal to drain terminal respectively. rds is the drain output impedance.
so
0
Zcas,in < Zcas,in
. The gain from input of M1 (M2 ) to its drain is proportional to the cascoded tran-
sistors input impedance, therefore using bipolar as cascoded transistor this gain is
Degenerative inductor Ls (1 nH) and load inductor Ld (3 nH) are on-chip planar
spiral inductors. They are formed using the top layer metal (analog metal, AM) and
there is a deep trench lattice pattern beneath them to reduce metal loss and substrate
90
loss respectively. Bond wires at the input are used to form the input biasing network.
The use of bond wire has two benefits. First, its quality factor is much larger than
on-chip inductors and thus, contributes less noise the overall noise factor. Second, it
does not occupy die area and makes the chip more compact.
A systematic design procedure was followed from the beginning and the circuit
was finely tuned through extensive simulations. Under matching conditions, there
are several equations established for the input matching network. The operation
frequency of the LNA relates the inductance and capacitance in the matching network
as
1
ωo = p (4.6)
(Lg + Ls ) Ct
In addition to the MOS transistor’s intrinsic gate-source capacitance Cgs , another
capacitor Cmim is added between the gate and source for noise consideration (not
shown in Fig. 45). Ct is the total capacitance between the gate and source and
gm
Rs = Ls = 50Ω (4.8)
Ct
1 1
Q= = (4.9)
2ωo Ct Rs 2ωo Ls gm
Now the proper Q value should be determined and the considerations about
input Q is elaborated as follows. The RF filter before the LNA requires a certain load
91
termination, which is the input impedance of the LNA, for maximum power transfer
and low sensitivity, and its performance is only guaranteed for a given tolerance for
this termination impedance, for example, between 25 Ω and 100 Ω. The variation in
the reactance part of the matching network can also cause serious variation of the
RF filter gain and pass band ripple. In order to maintain small variations in the
impedance and reactance of the matching network, its Q value can not be too large,
although a large Q is beneficial for low current consumption. Also large Q tends to
degrade the linearity of the input stage, because the Q is also the voltage gain of the
matching network, the input referred IIP3 will be reduced by a factor of Q. So the Q
calculated using
1
Ct = (4.10)
2ωo Rs Q
As stated above, the Ct is composed of gate-source capacitance Cgs of M1 (M2 )
and an MiM capacitance Cmim . The addition of Cmim helps to optimize the noise
performance and will be explained in further detail [11].
Fig. 46 shows the differential half circuit schematic (a) and small signal noise
equivalent circuit (b) of the LNA. Here the cascoded bipolar transistors have minor
influence on the noise behavior of the LNA, therefore its contribution to the total
The capacitance Ct affects the output noise current through the gate induced
(ωCgs )2
i2n,g = 4kT β ∆f (4.11)
gdo
where gdo is the transistor M1 ’s output conductance for zero drain-source voltage, β
92
2655
*:9 ;
7 -87
2<9
2/34+
D EKF H9
;1A
D EGF 9Q
M I
;1A O
')(
*,+ 2 AGB
D EGF H B D EKF A -JI
P D EKF
*,+ -/.10 . N 7
L 2 A
B
.
;CB ;1B
=>@? =SR?
4
is the gate induced current noise factor and is about 15
for long channel devices.
From the above discussions about the choice of the input quality factor, it is
transistor’s gate-source capacitance Cgs , the gate induced noise current will be quite
large, since the gate induced noise grows with the square of Cgs , which can be easily
seen from (4.11). Splitting Ct into Cgs and Cmim decouples Q from Cgs which allows
The effect of adding Cmim can be further explained as follows. The output noise
current due to gate induced noise current in,g is
gm jRs ωo Ct − 1
in,o,g = in,g (4.12)
jωo Ct j2Rs ωo Ct
93
Cgs Cgs
P ≡ = (4.14)
Ct Cgs + Cmim
and keeping Ct fixed, the output noise current originated from the gate induced noise
1
current is reduced by a factor of P
.
3 a 3 1
F = 1 + aQ2 W 2 + W 2 + bQ−2 W − 2 (4.15)
4
where a and b are constant determined by the length of the transistor M1 , process
parameters such as µef f Cox and bias current.
For a fixed Q, there exists an optimal value for W , which can be obtained by
taking the first order derivative of (4.15) in respect to W and equating it to zero.
r
Ab 5 1
Wopt = 4 (4.16)
2Q2 6 3 ωo Rs Cox L
Now that the size of the transistor is known from (4.16), Cgs can be calculated
from
2
Cgs = Cox Wopt L (4.17)
3
Ct has already been fixed by Q, so the required additional capacitance, Cmim , can
be found through (4.7). The gm of M1 can be fixed from the gain specification of
the LNA, then the degeneration inductance Ls will be obtained by (4.8). Finally the
94
After fine tuning through simulation, the width of M1 and M2 was fixed to be
96 µm which was laid out by 24 fingers. The additional capacitance Cmim is 277 fF and
Cgs is 140 fF for the size of the transistor mentioned above. The transconductance
at the operating frequency of 2.4 GHz band is about 11 mA/V. The degeneration
inductance Ls is 1 nH and the required gate inductance Lg is about 10 nH, which will
be implemented by the bonding wire and off-chip surface mount inductors. It can be
verified that the input impedance is about 26 Ω for the half-circuit shown in Fig. 46.
LW LW
Gain S.W.
Q T ,Q U
M T ,MU
LV LV
G S G S G
Fig. 47 is the layout. Its area is 570 × 580 µm2 . In the layout, interdigitate
95
and common-centroid techniques are used for transistor M1 and M2 to achieve good
matching [31]. The differential input pads are formed into a G-S-G-S-G pattern to
0.8
0.6
Coupling factor
0.4
0.2
-0.2
0 0.2 0.4 0.6 0.8 1.0 1.2 1.4
Normalized spacing (µm)
One can follow the LNA design flow given in Fig. 15, Chapter II. A single-ended
LNA is designed first then transformed to the differential form by duplicating it and
It is worth mention that, for the differential structure, the two source degenera-
tion inductors and two load inductors should be placed apart enough to avoid mutual
coupling. In order to obtain a quantitative knowledge about the coupling effect, the
mutual coupling factor k between two inductors in adjacent metal layers is simulated
using ASITIC [32]. The studied inductors are octagon spirals with a radius of 100µm,
metal spacing of 2 µm and metal width of 8 µm. One of the inductors is laid out using
metal layer 4, the other using metal layer 3. These two inductors are first put one on
top of the other, then they are shifted apart. Mutual coupling factors (k = √M ,
L 1 L2
96
M is the mutual inductance between L1 and L2 ) are calculated for different shifting
and are plotted in Fig. 48. The x-axis is the normalized spacing and is defined by the
spacing from center to center of the two spirals divided by the diameter of the spiral.
When the two spirals are completely overlapped, k is greater than 0.9 showing strong
coupling. Coupling is reduced with the increment of normalized spacing. There exists
a specific point where k equals to zero. After this point, k will increase with opposite
polarity and will reach an extremum. When the normalized spacing is 1, the two
spirals are shifted apart without any overlapping. The absolute value of the coupling
factor k for this configuration is about 0.03, which is already small enough. With the
two spirals shifting further apart, the coupling factor becomes gradually smaller.
It can be seen from the above study that as long as the two spirals are put apart
greater than 1.2 times of their diameter, the coupling can be reduced to a negligible
value. In Fig. 47, the minimum normalized spacing among the four inductors is 2.6.
Table X summarizes the simulation results of the LNA.
S21 15/-15 dB
NF 1.6 dB
IIP3 -3 dBm
S11 <-20 dB
Pd 16 mW
97
2. Mixer Implementation
The mixer shown in Fig. 49 is a fully differential Gilbert-cell based structure with I
and Q branches sharing the same RF drive stage, therefore eliminating the RF drive
stage mismatch compared to the conventional two separated I/Q mixers. The cur-
rent commutating switches are NPN bipolar transistors which require less LO power
than NMOS transistor switch pairs. This relaxes the required LO cross-coupling and
isolation performance. Bipolar switching pair also has lower flicker noise than their
required in order to achieve symmetrical on-off switching. This helps to minimize the
second-order distortion caused by a non-ideal LO signal duty circle [33]. The RF driv-
ing stage uses NMOS transistors for high linearity. A further detailed discussion [34]
cdd
]^
_ `Ga
[\ eu u eu u [\
X h j b
b b blbm bonbp bqrbs
eu u eu u
When there is no RF signal, due to the large LO signal, the voltage waveform
98
at the drain of M1 or M2 has a significant swing and it changes four times faster
than that of the LO signal. Comparing with the case of using two separate mixers,
the voltage swing at the drain of M1 or M2 is also much smaller. This is shown
in Fig. 50. Therefore, a signal with four times the LO frequency will leak into the
mixer’s RF input port through capacitive coupling (Cgd , Cgs ). Because this leakage
is at high frequency and its amplitude is about 10 times smaller than that of the two
separate mixer configuration, it will not cause problem. Due to device mismatch,
this high frequency signal will also appear at IF port but will be filtered out by the
load capacitance. Fig. 51 is the mixer layout view. Table XI summarizes the mixer’s
simulation results.
Conversion Gain 19 dB
DSB. NF 10.6 dB
IIP3 +2 dBm
Pd 8.8 mW
100
The design flow of the mixer can be generally followed as shown in Fig. 37,
Chapter III. The differences are 1) the current switching pairs are formed by bipolar
transistors here, their sizes are chosen just large enough to accommodate the current
following through them, which helps to reduce the parasitics at the drain of the RF
The RF driving stage of the LNA and mixer are all NMOS transistors. The threshold
mobility decreases, so does the gm . This can be seen more easily by considering
the MOS gm expression in the first order I-V approximation, i.e. the square law
ID = Kn (VGS − Vth )2 as
gm = 2Kn (VGS − Vth )
√ (4.19)
= 2 K n ID
where Kn = 21 µ (T ) Cox LW
ef f
. So It can be derived from (4.19) and (4.18) that
r
ID
gm ∝ (4.20)
T 3/2
Á Î Á/Æ Á Ì
Á Ë
ÁÃÂÅ ÁÃÂxÄ
Á Í ÇÈÉÈ
ÔÕÖ×
Q1 (Q1A and Q1B in parallel), Q2 , M1 ∼ M4 and REE form the PTAT current
source. Q1A , Q1B and Q2 have the same emitter area AE . The current mirror formed
current flowing through Q2 . This current is noted as IP T AT and assumes the satura-
tion current density of Q1 and Q2 is Jo , then the base-emitter voltage of transistor
kT IP T AT
VBE1 = ln (4.21)
q 2AE Jo
kT IP T AT
VBE2 = ln (4.22)
q A E Jo
The voltage developed across resistor REE can be obtained by subtracting (4.21)
102
from (4.22)
kT
∆VBE = VBE2 − VBE1 = ln 2 (4.23)
q
This voltage is proportional to absolute temperature. If the temperature coefficient
Vt
IP T AT = ln 2 (4.24)
REE
kT
where Vt = q
. For the REE value shown in Fig. 52, the nominal current at 298 K
ambient temperature is about 52 µA. All the other currents and voltages are derived
from this source. VLN A bypass is the control terminal to turn off LNA and put it into
attenuation mode.
From the above discussions, the biasing circuit can be designed as following.
First the current mirroring ratios are chosen. These ratios are usually less than 10.
Too small value will increase the power consumption of the biasing circuit, while too
large value will degrade the current mirroring accuracy. For the LNA, this ratio is
4 and for mixer, this ratio is 8. Secondly, the PTAT current IP T AT is known from
the current of the LNA and mixer, and the mirroring ratios. The required emitter
resistor REE , can be calculated from (4.24). The third step is to chose proper value of
resistors to generate all the required bias voltage. Finally, a start-up circuit is added
to ensure the correct operation status of the PTAT current source and a control circuit
is also inserted to turn on and off the LNA bias current. Table XII shows the nominal
The RF front-end was fabricated using IBM 0.25 µm SiGe BiCMOS technology through
MOSIS. A die photo is shown in Fig. 53. The area of RF front-end is 740 × 770 µm2 ,
not including bond pads. Deep trench and substrate contact rings are placed around
the front-end as indicated in Fig. 54. There are two layers of deep trench. The
substrate contact ring is connect to a quiet ground. This arrangement improves the
Substrate
contact ring
Fig. 54. Substrate noise isolation by deep trenches and guard rings
104
The testing board photo is shown in Fig. 55. This is the board for the whole
dual-mode Bluetooth/Wi-Fi receiver with accessible mixer output. The LNA input
SMA connector is put as close to the chip as possible. This is why there is a cut-in at
the left edge of the board. The PCB is fabricated using FR4 material with a thickness
of 0.031”. This FR4 material is relatively less costly than dedicated high frequency
lamination materials, and the thin thickness of the board makes the 50 Ωand 25 Ω
transmission line width manageable.
Mixer I and Q
branch output
Differential to single-
ended buffers
The input impedance matching condition was checked using a network S-parameter
analyzer HP8719ES. The testing setup is presented in Fig. 56, and the testing results
are shown in Fig. 57 for high gain mode and Fig. 58 for low gain mode. In both cases
!
"$#&%'(*),+-/.
021354
76
The intermodulation performance of the front-end was tested within the whole
receiver using the two-tone test method. Fig. 59 shows the instrumental setup. Two
signal generators SMIQ03 are used to generate the two testing tones. The two tones
are then combined by a power combiner and fed into the LNA’s input. The output
spectrum is observed by the spectrum analyzer FSEB30 from the VGA output with
at 12 MHz and 25 MHz away from same side of the LO frequency respectively when
the LNA is in high gain mode. The measured IIP3 is -13 dBm.
Fig. 61 shows the IIP2 plot. The two tones are applied at 12.2 MHz and 12.8 MHz
away from same side of LO tone and the measured IIP2 is 10 dBm. These values
meet the system specifications and indicate that the front-end performed as designed.
107
89:7;*<=
>?A@
B*CED!F,GIH
89:7;*<=
SEDATU?G
V DWXC,Y ZE?G
J 8KL=A<
M/NOQPR>
60
IIP3=-13dBm
40
20
Output (dBm)
-20
-40
-60
-80
-50 -45 -40 -35 -30 -25 -20 -15 -10
Fig. 60. IIP3 plot for 2-tone test at 12MHz and 25MHz offset
108
80
60
IIP2=10dBm
40
Output (dBm)
20
-20
-40
-60
Fig. 61. IIP2 plot for 2-tone test at 12.2MHz and 12.8MHz offset
The I and Q branch matching performance testing setup is shown in Fig. 62.
The input signal is swept to cover from 1 MHz to 10 MHz IF frequency range. The I
and Q branch’s amplitude and phase difference is observed by putting vector network
analyzer HP89140 into vector mode and looking at the amplitude and phase response.
[\U]7^*_`
s2tvu w,x!yzj gh!ijlknm,o,prq
abEc,d2e3f _
] t&myzjx,yzj
The measured mismatch between I and Q outputs of the mixer across 10 MHz
109
IF frequency range is shown in Fig. 63. It shows that the amplitude mismatch is less
than 1.2 dB, and phase mismatch is within 3.8 degrees. In the Bluetooth mode (up
to 1 MHz) the phase mismatch is as large as 3.5 degrees and the amplitude mismatch
0.96 dB. For the Wi-Fi mode (up to 6MHz), the phase mismatch is smaller than
3.5 degrees and the amplitude mismatch smaller than 1 dB. The demodulator’s SNR
degradation due to the I/Q mismatch is less than 0.2 dB in both cases.
1.12 4
1.1
Amplitude mismatch (dB)
3.5
1.08
The noise figure and conversion gain measurement can be performed by using the
spectrum analyzer FSEB30 from Rohde & Schwarz and noise source NC346B form
Noise/COM. The testing setup is shown in Fig. 64. Before the actual measurement,
the instruments are setup for calibration, then the front-end test board is inserted
into the testing chain. Testing is automated by software supplied with the spectrum
analyzer. The measured conversion gain is about 33 dB and noise figure is 5.5 dB
across 10 MHz IF frequency range. Table XIII summarizes the measurement results.
110
$EAAz
¥¢ ¡¢ £
¨ ® ª ¤ ¥ ¦§
¢
© ®¯ ¤ ¨ © ª
« ¨ ª ¨ ¢ «¬
¤ ©
3A A
{/|
}~*3z|$
}|,
|$
 3A A °±z²³´zµn¶n· ¸
ÁÃ Ä º Å Ä Á Ã
Á Â ÇÈ ¼
¾À½ ¿ ¼ ú Æ Â
ÆÁ Ç
¹ º» ¼ ½ º
$AEAn
{|
}~U}/!
}}2|
Fig. 64. Testing setup for noise figure and conversion gain
Voltage gain 33 dB
Vdd 2.5 V
Current 13.6 mA
IIP2 10 dBm
NF 5.5 dB
CHAPTER V
technique makes the CMOS process much cheaper than the others. The system-
on-chip target, also demands CMOS technology. For the same current consumption,
NMOS transistors are more linear than bipolar transistors. Still, the linearity of MOS
such as CDMA/AMPS. For example [36], in the IS-98 CDMA standard, the IIP3 of
the LNA is set by the single-tone desensitization requirement:
where LT X−RX is the duplexer TX-RX isolation in TX band and LT X is the duplexer
TX-antenna insertion loss. Typically, LT X−RX is about 53 dB and LT X is about
2.7 dB, which requires the LNA’s IIP3 to be better than +9.2 dBm.
Linearity based on negative feedback is not suitable for high frequency applica-
tions due to stability issues and gain reduction. Thus, a lot of linearization techniques
that focus on linearizing MOSFET transistors are introduced. The basic idea of lin-
The conventional way involves MOS transistors working in triode or weak inversion
systems with memory effect. Based on the non-linearity analysis, the techniques for
linearizing the LNA will be introduced, and then the proposed linearization technique
using hybrid transistors, i.e. using both MOS and BJT is discussed in details.
A. Non-Linearity Analysis
For a memoryless non-linear circuit or system, its transfer function can be represented
by Taylor series:
+∞
X
y = ak x k
k=0 (5.2)
= a 0 + a 1 x + a 2 x2 + a 3 x3 + . . .
The system non-linearity is usually characterized through the two-tone test. Sup-
pose the two testing signals are two sinusoids with the same amplitudes and different
(5.2), it is easy to shown that the linear term (x1 + x2 ) of the two-tone test is
1 2
A2 + 2
A cos 2ω1 t + 12 A2 cos 2ω2 t
(5.4)
2 2
+ A cos (ω1 − ω2 ) t + A cos (ω1 + ω2 ) t
113
the 3rd-order term (x1 + x2 )3 in the Taylor series. But the higher odd-order terms
also have contribution, especially when 3rd order cancellation techniques are used,
Higher than 5th-order non-linearities usually have less contributions and will generally
be ignored.
above. For the 3rd-order and 5th-order terms, only the fundamental and intermodula-
tion components are shown in the figure. Notice that odd-order nonlinearities generate
spurs around the fundamental and compress or expand the fundamental amplitude.
Even-order nonlinearities can generate DC and low frequency components which may
1 1
a1 (x1 + x2 ) × a1 A
ω1 ω 2
a2 ( x1 + x2 )
1 1 2 1
1/2 1/2
2nd order × a2 A 2
ω2 − ω1 2ω1 2ω 2
DC ω2 + ω1
9/4 9/4
3rd-order:
a3 (x1 + x2 )
3 Fundamental and
3/4 3/4 intermod components × a3 A 3
2ω1 − ω 2 2ω 2 − ω1
ω1 ω 2
25/4 25/4
a5 ( x1 + x2 ) 5th-order:
5
Fundamental and
intermod components
25/8 25/8
× a5 A 5
5/8 5/8
the magnitude and phase of Hk (ω1 , ω2 , · · · , ωk ) [37]. The Taylor series only shows
the low frequency effect (or the amplitude) of circuit non-linearity for a circuit with
memory, while Volterra series contains both amplitude and phase information. For
a general non-linear system, the locations of frequency components are the same as
depicted in Fig. 65, but their amplitudes and phases will be functions of frequency ω 1
and ω2 . The rest of the text in this section will obverse several non-linearity effects
For a fully-differential circuit, assume its two inputs are x1 = x and x2 = −x, then
and
yo2 (x2 ) = −H1 ◦ x + H2 ◦ x2 − H3 ◦ x3 + H4 ◦ x4 − H5 ◦ x5 (5.9)
yo = yo1 − yo2 = 2 H1 ◦ x + H3 ◦ x3 + H5 ◦ x5
¡ ¢
(5.10)
116
So a perfect fully-differential circuit does not have even-order terms in its differ-
ential output. But in reality, there will be asymmetry in the circuit. This can be the
mismatch between two signal paths or the input signals itself are not fully-balanced.
This effect can always be modeled by a DC offset ∆ at the input signal as x1 = x + ∆
and x2 = −x. Under this condition, the differential output can be shown to be
yo ≈ 2 (H1 ◦ x + H3 ◦ x3 + H5 ◦ x5 )
(5.11)
2 3 4
+ H1 ∆ + (2H2 ◦ x + 3H3 ◦ x + 4H4 ◦ x + 5H5 ◦ x )
It can be seen that input signal offset or system mismatch will cause the dif-
ferential output not only having DC offset H1 ∆ but also the even-order terms. The
IM3 is directly generated by 3rd-order non-linearity of the system but higher odd-
order nonlinearities also have contribution to IM3. For example, in Fig. 65 the fifth-
order nonlinearity term can also generate an IM3 term and cause the IM3 curve to
of the IM3 terms are proportional to the 3rd power of the input amplitude. If the
signal amplitude is significantly large, however, 5th-order distortion will start to affect
the IM3 responses. When some of the 3rd-order non-linearity cancellation techniques
are used to reduce the 3rd-order distortion, the IM3 curve usually deviates from
the 3:1 slope for a moderately large input signal. The curve will compress or expand
determined by the phase of the fifth-order term. If the phases of the 3rd and 5th order
coefficients are coherent, 5th-order distortion will expand the IM3 curve, whereas if
117
with the fundamentals again to generate 3rd-order terms. The IM3 term contributed
by the 3rd-order distortion and second-order distortion can be written using Volterra
series coefficients sun as
3
IM 3 (2ω1 − ω2 ) = H
4 3
(ω1 , ω1 , −ω2 ) X (ω1 ) X (ω1 ) X ∗ (ω2 )
3
IM 3 (2ω2 − ω1 ) = H
4 3
(ω2 , ω2 , −ω1 ) X (ω2 ) X (ω2 ) X ∗ (ω1 )
two tones in frequency domain, k = 1, 2. X (2ω1 ), X (2ω2 ) and X (ω2 − ω1 ) are the
second-order intermodulation terms feeded back to the input of the system. X ∗ (ω) is
the complex conjugate of X (ω). Fig. 66 shows how the envelope terms and harmonics
Quite often one may observe that the IM3 response at 2ω1 − ω2 and 2ω2 − ω1
are asymmetric in amplitudes. Suppose equal amplitudes in the two-tone test input
signals at ω1 and ω2 , non-linear terms caused by 3rd-order distortion H3 in the first
term of (5.12) and (5.13) usually match each other. Asymmetry arises from the facts
that i) the envelop term X (ω2 − ω1 ) appears in opposite phase in the lower and upper
sidebands, which is clear from the complex conjugate operation in the second term of
(5.12) and (5.13), and ii) the response of fundamental tones may be different. This
118
DC ω2 + ω1
ω1 − ω 2 ω 2 − ω1 2ω1 2ω 2
− ω2 − ω1 ω1 ω2
2ω1 − ω 2 2ω 2 − ω1
ω1 ω 2
usually is not the case because a flat pass band is generally desired for fundamentals.
But the 2nd-order harmonics X (2ω1 ) and X (2ω2 ) in the third term of (5.12) and
(5.13) are already far away from each other and further away from the pass band, so
their response can be quite different.
of its transconductance, but its output impedance is also non-linear and contributes
1
rout = (5.14)
λID
119
1
where λ = VA
, VA is Early voltage. ID is drain current. If the drain current contains
1
rout = (5.15)
λ (IDQ + io )
Assume rout dominates the output AC impedance, thus the output voltage due to
this impedance is
io 1 io /IDQ
vo = io rout = = (5.16)
λ (IDQ + io ) λ 1 + io /IDQ
¯ ¯
x ¯ io ¯
It is known for |x| < 1, 1+x
= x − x2 + x3 − x4 + x5 + . . .. Usually¯ IDQ ¯ < 1, thus
à !
1 1
vo ≈ r o io − i2o + 2
i3o = β1 io + β2 i2o + β3 i3o (5.17)
IDQ IDQ
1
where ro = λIDQ
, β1 = ro , β2 = −λro2 , β3 = λ2 ro3 . The output AC current io can be
or
Compare (5.19) and (5.20) the total nonlinearity term may be improved or de-
graded depending on the sign and relative value of the terms contributed from output
non-linearity behavior.
Û Þ Ú Ó × Ý Ù Ôà Óß Ý Ü Ô
Õ Ø ÀÕ Ö ØÕ
ÐÒÍÏ
Ñ Î ËÌ ä
â á ã Ê É
At high frequency, the load impedance will be dominated by the load capacitance,
thus memory effect can not be ignored. The output voltage should be expressed using
Volterra series:
Fig. 67 illustrates the large signal model of the circuit under consideration. For
simplicity and clarity, it is assumed that the MOS transistor is a long channel device
where Vod = Vgs0 − Vth , Vth is transistor’s threshold voltage. The output voltage for
121
sinusoid inputs is
H 0 + H 1 ◦ vi +
where
D0 = −KVod2 ZL (5.25)
D1 = −2KVod ZL (5.26)
D2 = −KZL (5.27)
In order to find the Volterra kernel Hk , equate the same order term of vi in both
sides of (5.24) and use the following relationships:
1
KVod2 RL ¿ (5.28)
λ
1
ωCL À (5.29)
RL
gm = 2KVod (5.30)
go = KλVod2 (5.31)
(5.28) holds because MOS transistor’s Early voltage is usually much greater than
its output DC voltage. (5.29) means the circuit’s output impedance is dominated
122
by the load capacitance at high frequency and the circuit has strong memory effect.
(5.30) and (5.31) are the linear small signal transconductance and output conductance
D0
H0 = ≈ −KVod2 RL (1 − go RL ) (5.32)
1 + λD0
1 + λH0 gm
H1 (ω) = D1 ≈ − (1 − go RL ) (5.33)
1 + λD0 jωCL
D2 + λ (D1 H1 + D2 H0 ) K (1 − go RL )
H2 (ω1 , ω2 ) = ≈− (5.34)
1 + λD0 j (ω1 + ω2 ) CL
λ (D1 H2 + D2 H1 ) Kλgm Z3 (ω1 , ω2 , ω3 )
H3 (ω1 , ω2 , ω3 ) = ≈− (5.35)
1 + λD0 (ω1 + ω2 + ω3 ) CL2
³ ´
where Z3 (ω1 , ω2 , ω3 ) = 13 ω1 +ω
1
2
+ 1
ω1 +ω3
+ 1
ω2 +ω3
+ 1
ω1
+ 1
ω2
+ 1
ω3
.
Here the DC term is dropped and it is assumed RL À ro . Notice that due to the
output capacitive loading, both the amplitude and phase of the non-linear term are
quite different from the case without memory effect.
ture. It handles the trade-off among input impedance matching, noise figure and gain
gracefully. This text will study the 3rd-order non-linearity of this type of LNA.
Figs. 68 are the schematic view (a) and equivalent circuit for non-linearity analy-
sis (b).
Ld Rd
Cd
Rs Lg C gd ids = f (vgs ) Z3
Lg vs
Z1 C gs
vgs Ld Rd
Rs Cd
vs Ls
Z2
(a) (b)
gm of the LNA core device, here the MOS transistor. The drain AC current can be
2 3
ids = f (vgs ) = gm vgs + g2 vgs + g3 vgs (5.37)
The distortion analysis for BJT was studied in [38], [39] using Volterra series. The
equations can be adapted for MOS transistors by noting that MOS transistors have
source resistance. The IM D3 at drain node and the input referred IP3 power can be
expressed as
3
IM D3 = · |H (ω)| · |A1 (ω)|3 · |ε (∆ω, 2ω)| A2s (5.38)
4
and
1
IIP 3 (2ω2 − ω1 ) = (5.39)
6Rs · |H (ω)| · |A1 (ω)|3 · |ε (∆ω, 2ω)|
124
1 1 + jωCgd Z3 (ω)
A1 (ω) = (5.41)
gm + g (ω) Zx (ω)
ε (∆ω, 2ω) = g3 − goB (∆ω, 2ω) (5.42)
· ¸
2 2 2 1
goB (∆ω, 2ω) = g2 + (5.43)
3 gm + g (∆ω) gm + g (2ω)
1 + jωCgd [Z1 (ω) + Z3 (ω)] + jωCgs [Z1 (ω) + Zx (ω)]
g (ω) = (5.44)
Zx (ω)
Zx (ω) = Z2 (ω) + jωCgd [Z1 (ω) Z2 (ω) + Z1 (ω) Z3 (ω) + Z2 (ω) Z3 (ω)] (5.45)
In the above equations, H (ω) relates the equivalent input IM3 voltage to the IM3
response of the drain current non-linear terms. A1 (ω) is the linear transfer function
from the input voltage vs to the gate-source voltage vgs . ε (∆ω, 2ω) shows how the
non-linear terms in (5.37) contribute to the 3rd-order distortion. The first term in
(5.42) comes from the 3rd-order non-linearity and the second term comes from the
2nd-order non-linearity as explained in (5.12) and (5.13). The 2nd-order feedback
paths here include the gate-drain capacitor and the degeneration inductor.
|H (ω)| and |A1 (ω)| depend on the in-band source and load impedances which are
usually selected to provide desired gain, noise figure and impedance match. Therefore
lower distortion is achieved by reducing |ε (∆ω, 2ω)|. If g3 dominates non-linearity
which is generally the case, reducing it can significantly improve IIP3. In a bipo-
lar LNA, out-of band termination or matching is usually used to make the term
|ε (∆ω, 2ω)| small thus IMD3 will be reduced. This is possible because g3 and goB
have the same sign for bipolar transistor. |ε (∆ω, 2ω)| is the difference between these
different signs. More specifically, in the saturation region, since g3 is negative while
goB is positive, the only way to reduce |ε (∆ω, 2ω)| is to make both g3 and goB small
values.
So when g3 has already been reduced by the third-order cancellation technique,
it is also important to reduce goB by keeping g (∆ω) and g (2ω) large. It is assumed in
the above discussions that ∆ω is at a very low frequency. Therefore for the inductive
source-degenerated LNA, Z1 (∆ω) ≈ Rs , Z2 (∆ω) ≈ 0 and Z3 ∆ω ≈ 0. So g (∆ω) has
a very large value with respect to gm , and goB (∆ω, 2ω) can be approximated by
2 g22
goB (∆ω, 2ω) ≈ goB (2ω) = (5.46)
3 gm + g (2ω)
If the load LC tank has a high enough quality factor, then Z3 (2ω) is a small quan-
tity and Z1 (2ω) = Rs + j2ωLg , Z2 (2ω) = j2ωLs . Under input impedance match
gm
condition, ωT Ls = L
Cgs s
= Rs holds. Putting all these considerations together, it
1
can be shown that |g (2ω)| is proportional to ωLs
. So in order to have small goB ,
one must select a small Ls , which means less degeneration. This is contrary to the
about 2 ∼ 3 times larger. Therefore, IIP3 can have an improvement by at least 3 dB.
This |g (2ω)| increment assumes that the out-of-band termination network is added
directly at the gate of the MOS transistor, thus it has to be implemented on the chip.
tank resonant at ω, provides a very small impedance path to ground at frequency 2ω.
Its performance will be limited by the quality of inductors available in the process
and it will introduce additional noise, and probably affect the in-band impedance
match. So it is usually not desirable for the LNA design. Off-chip termination using
126
low loss quarter-wave transmission lines is another way, but the improvement will be
relatively small and is not worth the complexity, and added cost. To conclude, in
order to make a linear LNA, the 3rd-order coefficient g3 of the intrinsic transistor’s
Taylor series should be small and if possible, adding an out-of-band termination net-
work and/or using a smaller degeneration inductor will also help. The rest of this
chapter will study the techniques to provide a more linear LNA core circuit.
The low noise amplifier (LNA) has stringent requirements on operation frequency,
noise, and linearity. Therefore, a LNA usually uses a minimum number of transis-
tors. For example, a single-ended LNA usually contains only one or two transistors in
its main signal path. In order to improve the LNA’s linearity, more components have
to be added into the signal path. Resistive degeneration and shunt-series feedback
are traditionally used, but they introduce additional noise and require more power. A
LNA’s core devices. There is no theoretical analysis done on the previously reported
MGT. This section will explore the nature of the multi-gated linearization technique
and the later sections will present a different way to implement the 3rd-order inter-
modulation cancellation by using a hybrid structure, i.e. combining the MOS and
than 3rd-order:
2 3
ids (vgs ) = gm vgs + g2 vgs + g3 vgs (5.47)
127
In order to obtain a large IIP3, gm should be kept almost unchanged or even larger
The transfer characteristics of a short channel MOS transistor that is valid in all
operation regions can be expressed as [45]
χ2
iDS = K (5.48)
1 + θχ
where
³ Vgs −Vth ´
χ = 2ηφt ln 1 + e 2ηφt (5.49)
kT
Vth is the threshold voltage, φt is the thermal voltage q
, µe is effective mobility
channel length and is independent of body effect. η is the rate of exponential increase
of drain current with gate-source voltage in the sub-threshold region and the size of
the moderate inversion region, which has values between 1 and 2 [45]. In moderate
Here the gate-source voltage is expressed as the sum of the a DC bias voltage Vgs0
and small signal AC voltage vgs , and Vod = Vgs0 − Vth is the gate source over-drive
voltage.
Expanding (5.50) using the Taylor series in terms of vgs and neglecting the DC
component and components higher than 3rd-order, the coefficients in (5.47) are de-
termined by
KVod (2 + θVod ) K θK
gm = 2 , g2 = 3 , g3 = − (5.51)
(1 + θVod ) (1 + θVod ) (1 + θVod )4
128
The low frequency 3rd-order intercept point for the transistor in strong inversion is
4 Vod
A2IIP 3,strong = (2 + θVod ) (1 + θVod )2 (5.52)
3 θ
8 Vod
A2IIP 3,strong > (5.53)
3 θ
vgs
ids = Is0 e ηφt (5.55)
Vod
where Is0 = K (2ηφt )2 e ηφt . Treated the same way as in the strong inversion case, the
Taylor series coefficients of (5.54) are
√
AIIP 3,weak = 2 2ηφt (5.57)
Fig. 69 shows the second and third-order terms of a NMOS transistor versus its
gate-source biasing levels. It can be easily seen that between the moderate/strong
and weak inversion, the 2nd-order term has the same sign, and both are positive.
The 3rd-order term has a different sign. In the moderate or strong inversion, it has
negative sign, and in the weak inversion it has a positive sign. In the moderate
inversion region, there is a point where the 3rd order term is zero. Of course, it is
very hard to bias a transistor exactly at this optimal point for minimum 3rd-order
129
V
æ$ç íè æ$ç æ
ì
÷ ü æ$ç è
ê gþ g æ$ç æ
õ
úû æ$ç èæ æ$ç æê ü
gÿ
øù ÷ æ$ç éÀë
æ$ç æ
ó úû
æ$ç æ
ì
æ$ç æë øù
g
ý ü æ$ç æè æ$ç æ
í
g
úû å æ$ç æë
øù ý å æ$ç é7æ æ$ç æè
å æ$ç éÀê æ$ç æ3é
g
å æ$ç è
è æ$ç æ
æ
æ3ç èîæïæ3ç èîìðæ3ç í
êñæ3ç ëëòæ$ç ó
èðæ$ç êîæôæ$ç êîìïæ$ç õ
êöæ3ç ì
ë
V
non-linearity. The alternative is to use two or more transistors and bias them at
different inversion regions and trim the size of the transistors such that the positive
term will cancel the negative term. This way, the minimum 3rd-order distortion is
achieved. [40]-[44] shows the idea of using multiple NMOS transistors connected in
parallel form and biased at different gate drive levels to achieve an extremely linear
The multi-gated configuration using two NMOS transistors is shown in Fig. 70.
Transistor M1 is the main transistor working in the strong inversion region. Transistor
Fig. 71 shows the linearity (IIP3 and IIP2) versus bias voltage VBaux of the
auxiliary transistor. The main transistor is biased at 0.74 V. Notice that when the
bias voltage of the auxiliary transistor is around 0.5 ∼ 0.55 V , the combined device
response has the best IIP3. The IIP3 improvement is about 10 dB compared to only
having transistor M1 . Note that the linearity measurement here is for multi-gated
130
RB
VBaux
M2
Cc
RFIN Cc
M1
VBmain
RB
transistors only. The observed output is the combined drain current. The actual
circuit using the multi-gated core will have less IIP3. The current consumption only
increases slightly because M2 is operating in the weak inversion.
16 16
15 15
14 14
IIP3
13 13
IIP2
12 12
IIP3 (dBm)
IIP2 (dBm)
11 11
10 10
9 9
8 8
VBmain=0.74V
7 7
6 6
0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7
Auxilliary gate bias, VBaux (V)
Fig. 71. NMOS multi-gated transistor linearity v.s. auxiliary bias voltage
Notice that the IIP2 almost has its worst value as demonstrated in Fig. 71. This
is due to the fact that for different gate biases, although the 3rd-order terms can have
different signs, so they can be canceled out by combining the current, the 2nd-order
131
terms will always have the same signs. By adding the output currents of the main and
the IIP2.
VDD
RB M2
V Baux
VB
M3
Cc ids,p
io
RFIN Cc ids,n
VBmain M1
RB
+,+.- /
)( )(
%& ' %& '
$ ! *
"#" +,+.-10 "#"
!
24365 7983:<;>=5 3 ?@3:BADC EFG
HBIJK5 L 5 36M N9863:O;>=65 3 ?PQGSRUT<VXWYG[Z
multi-gated transistor, CMGT), IIP3 and IIP2 can be improved at the same time.
Fig. 72 illustrates this configuration. When the AC signal of the NMOS transistor
132
will decrease. So if the input for M1 is vgs , the input for M2 will be −vgs . Thus if we
ignore the DC and higher order distortion terms, the output current of NMOS M 1
and PMOS M2 can be written as
2 3
ids, n = gm, n vgs + g2, n vgs + g3, n vgs (5.58)
2 3
ids, p = −gm, p vgs + g2, p vgs − g3, p vgs (5.59)
The overall output current of the CMGT is the difference between the currents of
those two transistors:
io = ids, n − ids, p
(5.60)
2 3
= (gm, n + gm, p ) vgs + (g2, n − g2, p ) vgs + (g3, n + g3, p ) vgs
It is observed that as the total transconductance increases, the IM2 term de-
creases because g2, n and g2, p have the same sign, and the IM3 term decreases because
g3, n and g3, p have different signs as shown in (5.51) and (5.56). Fig. 73 shows the
IIP3 and IIP2 curves of the CMGT configuration. Because PMOS transistors have
as significant as the NMOS MGT. It is also clear that the IIP2 and IIP3 do not share
the same optimal bias voltage. This is due to the different non-linear characteristics
of NMOS and PMOS transistors. If the future process can match the non-linear be-
havior and speed of NMOS and PMOS transistors, the linearity improvement will be
more significant.
Instead of making g3 small by using MGT and goB small by out-of-band termi-
nation, [47] proposed a method to change the phase and amplitude of the MGT’s
in-band g3 by tapping the degeneration inductor into the source of the transistor
working in the weak inversion. The consequence is that g3 is modified to match the
133
Using a MOS transistor biased at weak inversion may have potential speed limita-
tions [48]. A new implementation method to cancel the IM3 term is proposed. The
goal is to keep the main transistor M1 in the strong or moderate inversion where its
IM3 has a negative sign and another transistor is added to provide a positive signed
IM3. Instead of using the MOS transistor M2 biased in its weak inversion, a bipolar
transistor Q2 is used. Fig. 74 depicts the configuration of the proposed LNA lin-
earization method (Hybrid LNA). This implementation requires both MOS and BJT
available in the process. BiCMOS is a natural choice, but in a specific RF CMOS
not compete with that in a BiCMOS process, for linearization purposes, it may be
good enough. For example, in the TSMC 0.18 µm RF CMOS process, a 2 µm × 2 µm
and its βac is 22. A NMOS RF transistor with dimensions 185 µm × 0.18 µm biased
where VBEQ is the base-emitter bias voltage, Iso is the saturation current and IQ =
VBEQ
Iso e φt . Note that (5.61) resembles (5.55), so its 3rd-order term coefficient can be
VDD
Ld RFOUT
VBNPN
RB2
RFIN Q2
Bipolar replaces NMOS
M1 at weak inversion.
re
RB1
VBMOS
Bond wire
sion is
θK
g3, mos = − (5.63)
(1 + θVef f )4
It is observed from (5.62) and (5.63) that g3, bjt and g3, mos have different signs. If
their magnitudes are matched, the overall 3rd-order term can be canceled. Usually
the absolute value of the bipolar’s IM3 coefficient g3, bjt is much larger than that of the
NMOS transistor’s IM3 coefficient g3, mos . Therefore, g3, bjt needs to be scaled down
show that the memory effect in the degenerated BJT is weak, so it does not change the
phase of the non-linear terms. For this purpose, Volterra analysis of the degenerated
BJT is carried out.
135
Where all the signal quantities are assumed to be in the sinusoidal form, e.g., i o means
Aio cos (ωt + φ). In order to calculate Volterra kernel Bk , a large signal high frequency
model is shown in Fig. 75. In this model, only the collector current non-linearity is
considered. The base current ib and resistor rb (as shown in the dashed box) are
ignored and all the capacitors are assumed linear. Base-collector capacitor Cµ will be
combined with the MOS transistor’s gate-drain capacitor Cgd , so it is excluded from
the model. The difference between the emitter and collector current is also ignored,
and Re is the degeneration resistor. A more strict analysis can be found in [38].
vi Cµ
io
rb
ib vπ
Cπ
v −v
ve io = I Q exp i e
φt
Re
where ge = Re−1 . Solving for ve and substituting it into (5.65) and replacing io by its
B0 + B1 ◦ vi + B2 ◦ vi2 + B3 ◦ vi3 + · · · =
n
IQ 1 + I1t [−B0 + (ge − B1 ) ◦ vi − B2 ◦ vi2 − B3 ◦ vi3 + · · ·]
(5.67)
2
+ 2I12 [−B0 + (ge − B1 ) ◦ vi − B2 ◦ vi2 − B3 ◦ vi3 + · · ·]
t
o
3
+ 6I13 [−B0 + (ge − B1 ) ◦ vi − B2 ◦ vi2 − B3 ◦ vi3 + · · ·] + · · ·
t
B0
−φ
B0 = IQ e t ge (5.68)
where It0 = φt ge .
holds. Therefore, the phase angle of B1 is very small (no greater than 1 degree) and
For the worst case ω1 = ω2 = ω3 , the phase angle of B3 is no greater than 6 degrees.
Thus B3 can also be regarded as frequency independent. It should also be noticed
that if the degeneration resistance is chosen right, B3 can become zero or change
polarity.
The above derivations have justified the memory effect in the resistive-degenerated
BJT is very weak and can be ignored. Resistive degeneration can be used to scale
the magnitude of the 3rd-order coefficient to match that of MOSFET. In this case
(5.62) corresponds to
3
1 gm
g3, bjt = (1 − 2gm Re ) (5.73)
6IQ2 (1 + gm Re )5
due to the extremely weak frequency independence. Because the theoretical analysis
does not consider the BJT’s base and emitter’s extrinsic resistance, so the theoretical
curves are shifted from the simulated ones, and the non-linearity predicted by the
theoretical curves is a little bit larger, but the trend is well predicted. Fig. 77 shows
the 3rd-order cancellation effect of the proposed hybrid configuration. The MOS and
138
BJT are biased separately and then their bias voltages are swept and their output
0.25
Theoretical
0.20
0.15
DC & 3 GHz
0.10
g3,NPN (A/V3)
3 GHz
0.05
DC
0.00
Simulated
-0.05
-0.10
-0.15
0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95 1.00
Bias voltage (V)
The input matching network can be designed using the inductive source-degeneration
technique [10]. The degeneration inductor will be implemented using bond wire.
The gate inductance will be implemented in part from the bond wire and the other
part from off-chip surface mount inductor. The on-chip inductor is usually not used
at input because its size is usually large, so it will consume too much chip area.
Another reason to put the gate inductor off-chip is that the on-chip inductor does not
have a good quality factor, so the loss of the on-chip inductor will degrade the noise
performance of the low noise amplifier.
Fig. 78 is the small signal equivalent circuit for the input impedance calcula-
139
0.34
0.26
0.18
g 3,NPN
0.10
g3 (A/V3)
0.02
-0.06
g 3,MOS g3
-0.14
-0.22
-0.30
-0.16 -0.12 -0.08 -0.04 0.00 0.04 0.08 0.12 0.16
Input voltage (V)
Fig. 77. 3rd-order terms of the BJT, NMOS and their combination
tion. Ct accounts for the total capacitance between the gate and source of the MOS
transistor M1 in Fig. 74.
Ct = Cgs1 + Cπ (5.74)
expressed as
1 IQ
gπ = = (5.75)
rπ βφt
The bipolar’s emitter degeneration resistor Re is about 20 ∼ 40 Ω. This is a relatively
small value and will be ignored in the following analysis for simplicity.
as
µ ¶
gm gπ L s 1 gm gπ
Zin ≈ jω Ls + Lg + 2 2 + + Ls + 2 2 (5.76)
ω Ct jωCt Ct ω Ct
It is observed that the bipolar transistor introduces a shunted RC network between
the gate and source of the MOS transistor as shown in Fig. 78. The base-emitter
140
Lg g mvgs
gπ vgs
Ct
Z in
Ls
capacitance Cπ will shift the input matching frequency to a lower frequency. The
resistance rπ in parallel with this additional capacitor is at the magnitude of several
kilo-ohms. This resistance will vertically shift the S11 curve upward. Fig. 79 shows
the simulated S11 plot with and without the bipolar transistor activated. Therefore,
considering the effects of bipolar input impedance on the total input impedance, the
\.cb
\ `a
jDkltvu w1opq r s
\` `
\` _
\` ^
\ `b
\] a
jDklmXn opq r s
\ ]`
\ ]_
\ ]^
`Kd b ae`Kd b fe`Kd g ae`Qd g fh]6d a ai]6d a fh]Kd c1ah]6d cfi]6d `a
x1y z{6|z}~U,Q
U
Due to the low bias current, the bipolar contributes a small amount of noise to
the whole circuit. Simulation shows that the bipolar transistor adds less than 2.4%
to the overall noise of the circuit while the MOS transistor contributes about 14%.
Table XIV lists the noise contribution ratios of different devices at 3 GHz. Here the
input matching network is designed to achieve the best impedance match. The noise
figure calculated from the values given in the table is about 2.2 dB.
Table XIV. Device noise contribution ratios of single-ended hybrid LNA († signal gen-
erator’s internal resistance)
Rs† 60%
used. It is necessary to verify that the NPN transistor is fast enough to work at the
RF frequency. Table XV lists the simulated fT of the NPN transistor and NMOS
transistor used in Fig. 74 against process corners. It can be seen that the NPN
transistor almost has the same speed as the NMOS transistor in FF and SS corners.
Figs. 80 gives the IIP3 of the NMOS-NPN combination with different bias condi-
tions of the MOS and bipolar device. Two operation frequencies are given: 2.4 GHz
and 3 GHz. The two test tones are placed 2 MHz apart. The two plots in Figs. 80
for different frequencies are almost the same. This means that IIP3 is not sensitive
142
TT 30GHz 28GHz
FF 53GHz 35GHz
SS 21GHz 22GHz
It is observed from the figures that if the MOS transistor is biased at 0.685 V
and the bipolar transistor is biased at 0.8 V, the configuration has the optimal IIP3
condition. If the MOS transistor gate is biased too low or too high, there will exist
two IIP3 maxima, one for a lower bipolar base bias and the other for a higher bias.
For Vbmos = 0.685 V , if due to process variations, its value changes to 0.635 V or
0.700 V, the IIP3 does not vary significantly. If Vbnpn is around 0.8 V, the IIP3 curves
is pretty flat for Vbmos = 0.635 V and Vbmos = 0.685 V . Notice that for Vbmos = 0.7 V ,
the IIP3 at Vbnpn = 0.8 V has a local minima, so the variation of Vbnpn within about
±30 mV will not degrade the IIP3.
Fig. 81 is the IIP3 versus bipolar base bias voltage against process corners. Here
the MOS transistor is biased at 0.685 V. Process corners will make the optimal bias
condition change, especially for the FF corner. The optimal bias voltage is shifted to
a lower value for the FF corner. An On-chip corner dependent biasing scheme can be
how the IIP3 profile changes with different bias voltage of the NPN transistor under
different temperatures. The optimal biasing point of the bipolar transistor shifts with
temperature as shown in Fig. 83. This biasing profile can be realized by deriving the
143
1
OY .¡¢.£ ¤.¢O¢Y
OY .¡¢.£ ¤.¥O¦Y
OY .¡¢.£ ¤.§O¦Y
© ª «X¬K®
Ø× OY .¡¢.£ ¨<¢O¢Y
ÔÕÖ
Ó
ÑÒÑ
1 1 11
ËDÌ.Í<Î.ÍKÏ,ËQÐ
´³
º»O¼½O¾<¿À<Á Â<À<À º
²± º»O¼½O¾<¿À<Á Â<Ã<Ä º
².° º»O¼½O¾<¿À<Á Â<Å<Ä º
Ç1µ ³ÈSÉÊ º»O¼½O¾<¿À<Á ÆOÀ<À º
Ø× ².¯
ÔÕÖ ².´
Ó
ÑÒÑ ²³
±
°
¯
³µ ¶³µ ¶ ´³µ ¶ ¯³µ ¶!°³µ ¶±h³µ ±·³1µ ±!´¸³µ ± ¯³1µ ± °³µ ±±³1µ ¹
ËDÌ.Í<Î.ÍKÏ,ËQÐ
18
SS
TT
16
FF
14
IIP3(dBm)
12
10
4
0.7 0.74 0.78 0.82 0.86 0.9
Vbnpn (V)
base bias voltage from a PTAT current source running through a resistor. The NMOS
transistor should be biased using a constant -gm biasing circuit to keep the gm of the
24
Temp=-10 Temp=0
22 Temp=10 Temp=30
Temp=50 Temp=80
20
Temp=100
18
16
IIP3(dBm)
14
12
10
8
6
4
0.6 0.64 0.68 0.72 0.76 0.8 0.84 0.88
Vbnpn (V)
Ù1Ú à!â
Ù1Ú àá
ôõö
ó Ù1Ú Ý à
î
ñò ð
Ù1Ú Ý ß
íî ïð
Ù1Ú Ý!Þ
é ëì
éê Ù1Ú ÛÜ
Ù1Ú Û Û
ã áÙÙháÙÞÙåäÙæâ1Ùåß ÙæÛ1ÙæÝÙà1ÙçÜÙèáÙÙ
÷Qøùûúøü ýþ ÿ ü ø
This single-ended hybrid LNA’s design flow is depicted in Fig. 84. To begin with,
BJT is added to the circuit. The BJT and degeneration resistor’s size and bias cur-
rent can be determined from the MOS transistor and BJT’s 3rd order coefficient
simulation plot. After optimal conditions are found, the input impedance matching
the load should be fine-tuned to make gain return to the specification. The overall
simulation verification is then carried out and the design usually needs to iterate to
The IIP2 shown Fig. 71 almost has its worst value for the best IIP3. This is due to
the fact that for different gate bias, although the 3rd-order terms can have different
signs, so they can be canceled out by combining the current, the 2nd-order terms,
however, will always have the same signs. By adding the output current of the main
and auxiliary transistor, the 2nd-order terms will be added constructively, which will
deteriorate IIP2. The single-ended hybrid LNA configuration also has this problem.
In order to simultaneously keep the 2nd-order performance and provide 3rd-order
effect of a differential pair. Usually Cp is proportional to the size of the tail current
source and thus, to the tail current. gm is also proportional to the tail current.
146
- Noise figure
Design Specs - IIP3
- Voltage gain
- Power consumption
NF, Gain
Adjust Lg, LS and Load
Simulation
Verification
NO YES
Meet the Layout and post
Specs ? layout simulation
NO Meet the
Specs ?
YES
Tape-out
Therefore, ωp for MOS is at the same order as that for the BJT. The additional
phase shift in H3 for MOS is almost the same as that of BJT, i.e. for a bipolar
or MOS differential pair, their 3rd-order terms will have the same sign. However, as
mentioned before, for the same current consumption, bipolar’s 3rd-order term is much
larger than that of the MOS transistor. So with reduced bias current and resistive
emitter degeneration, bipolar’s 3rd-order term can be made to match that of the MOS
transistor. By subtracting the output of the MOS and bipolar differential pair, the
3rd-order term can be canceled without significantly reducing the overall gain.
VDD
LD LD
VOUT+ VOUT-
M1 M2
VIN- VIN+
Q1 Q2
re re
IEE ISS
In the ideal case, where the matching between transistors is perfect, there should
practise, mismatch exists, and can be modeled as a biasing offset. The proposed
differential method can not make the 2nd-order non-linearity substantially smaller
than a single differential pair, but it can expand the input range for a small 2nd-order
term, thus becoming more tolerant to the bias offset or device mismatch. On the other
hand, if single-ended output is chosen from the differential structure, simulation shows
and bipolar differential pairs together with overall non-linearity terms. The 3rd-order
cancellation and 2nd-order range expansion are easily identified from the plots.
120 1.20
Extended flatness 0.80 BJT+NMOS
80 for small IM2
0.40
40 BJT+NMOS
IM2 (mA/V2)
IM3 (A/V3)
0.00
0
-0.40
-40 BJT NMOS
-0.80
BJT
-80 -1.20
NMOS
-120 -1.60
-300 -220 -140 -60 20 100 180 260 -300 -200 -100 0 100 200 300
Fig. 86. (a) 2nd-order input range expansion and (b) 3rd-order cancellation of the
proposed differential LNA
differential pairs. The output currents coming from the bipolar differential pair is
subtracted from that of the MOS differential pair. While in the single-ended struc-
ture, the output currents generated from the MOS and bipolar transistors are con-
structively combined. This subtle difference makes the input matching of differential
probably introduce an additional differential phase shift between the MOS and bipo-
lar differential pairs which will deviate the non-linearity cancellation effect. So the
contributes about 15% to the overall noise of the circuit. Table XVI shows how
the noise contribution ratios break-down into different components. The noise figure
Rs† 42%
The differential hybrid LNA’s design procedure is similar to the single-ended one
previously shown in Fig. 84.
The single-ended and differential LNA’s were designed using TSMC 0.18 µm RF
CMOS process. Testing chips were fabricated through MOSIS MEP program and
packed in a QFN package. The die photomicrographs of these two circuits are shown
in Figs. 87. The active size is 390 µm × 390 µm for the single-ended LNA and is
(a) (b)
Fig. 87. Die photomicrographs of hybrid linearized LNA’s (a) single-ended (b) differ-
ential
Fig. 88 shows the IIP3 plot of a single-ended linearized LNA using the lineariza-
tion technique depicted in Fig. 74. Out-of-band termination is realized using the
tuned LC tank as the load [49]. The two tones are put at 2700 MHz and 2701 MHz
respectively. The S-parameter measurement plots are shown in Fig. 89. The S11 is
better than -10 dB from 2.6 GHz to 2.9 GHz. Due to the single-ended structure and
lack of cascoded transistor, the reverse isolation (S12) is not very good. The power
gain is about 6.4 dB, noise figure is measured to be 2.1 dB around 2.7 GHz. It draws
centered around 2.5 GHz. Fig. 90 is the measured IIP3 plots of the circuit with
and without the cancellation bipolar pair activated. The two testing tones are put
at 2500 MHz and 2501 MHz respectively. When the bipolar pair is enabled, the
fundamental term is reduced by about 2 dB and the IM3 term is reduced by 12.5 dB,
therefore about 5 dB IIP3 improvement can be obtained with the bipolar cancellation
pair. Fig. 91 is the measured S-parameters of the differential LNA. The S11 is better
than -9 dB from 2.4 GHz to 2.6 GHz. The circuit’s reverse isolation measured by S12
151
ED
"! # $
ABC
?@>
; <=
; <:
9:
%&'(*)+'-,.0/12435687
J
S I G
K
p L
mno l
k jh
F K
ij S II
g
fgh
de 4F J
S GG
FG-I
S GI
FG-H
IM N IM NN IOM H I M HPN I M Q IM QN IM J IOM JN IM R
T-UVWX-V"Y-Z\[^]4_+`Pacb
40
IIP3 with BJT disabled: 7.3 dBm
20
-20
-40
-100
-30 -25 -20 -15 -10 -5 0 5 10 15
Input Power (dBm)
Fig. 90. IIP3 of differential LNA with and without bipolar cancellation pair activated
xv
x-t w+x
v
t
\q v ww
qx-t xx
qxv
q\wt
q\wv
x-w
q\ut
q\uv
qsrOt
wOy uzwy uvzwy r{wy rOv|w y v}wOy vPv~wOy ~wy Pvzw y
+
is much better than that of the single-ended LNA. The power gain is about 10 dB.
Noise figure is measured to be 3.4 dB around 2.5 GHz. This differential LNA draws
11 mA from a single 1.8 V power supply. Because the differential structure inherently
suppresses the second-order non-linear term, out-of-band termination is not required
Table XVII compares the proposed technique with reported high linearity LNA’s.
The figure of merit (FOM) [50] is defined as
PIIP 3 × G
F OM = (5.78)
(F − 1) PD
where PIIP 3 , G, F and PD are the input-referred 3rd-order intercept point, power
gain, noise factor and power dissipation respectively.
The proposed linearized LNAs can work at a higher frequency and achieve an
improved figure-of-merit. This is due to the fact that the emitter-degenerated BJT
can be designed without the BJTs first, and then adding the BJTs for linearization.
154
Note that the proposed design is implemented in CMOS technology, the bipolar
device has a limited performance. One may wonder what if a true BiCMOS technol-
ogy is used. To this end, a single-ended hybrid LNA is simulated using IBM 0.25µm
BiCMOS technology. Thanks to the superior bipolar transistor available in the BiC-
MOS process, the noise figure of this LNA is less than 1.2 dB at 3.0 GHz with the
power gain of 9.5 dB and IIP3 of 12 dBm. The power dissipation is 6.6 mW. Thus
the BiCMOS design has a figure-of-merit 67, which is much better than its CMOS
counterpart.
155
CHAPTER VI
Nowadays, wireless technologies are advancing faster then ever. The diverse range
interest. Being the first active block in the receiver chain, the low noise amplifier
(LNA) has to have multi-band capabilities. Most of the current wireless systems are
narrow band. A wide band LNA can be used to cover a wide frequency range, but it
consumes a large amount of power and at the same time amplifies interferences which
are not on the actual reception band. Therefore, one would like the LNA’s transfer
function curve to looks like the superposition of two or more narrow band LNAs. In
order for the front-end LNA to cover more than one frequency band, the matching
network must be able to provide degrees of freedom equal to the number of bands
covered. Methods already reported in the literatures are: (i) device switching [33], (ii)
concurrent matching [51], and (iii) device switching plus concurrent matching [52].
The switched method can provide optimal design for every band while the concurrent
approach can receive more than one band at the same time.
Current wireless mobile terminals provide dual- or multi-band operations mainly
will be enabled to access different systems providing various services. Fig. 92 shows
the mobile station receiver band in the spectrum for a clear review of frequency
allocations. The frequency numbers in the figure are shown in MHz. Numbers shown
156
§¨ª©¬« ¶ «¨
«-®4¯ ¦*°c±c²*£ ½¾"«-® ±\¢µµ ©c¾"«-® ±¤-µµ ¿ ¾½P¨§
«-® ¤ ¯ ³³ ³³ ³³ ³³ ³³ ³³ ³³ ³³ ¶ ³³ ³³ ³³ «³³ ³³ ³³ ¨ ³³ ³³ ³³ ³³ ³³ ³³ ³³ ³³ ³³ ³³
·
¡¡ ¡¡ ¡¡ ¡¡ ¡¡ ¡¡ ¡¡ ¡¡ ¡¡ Às¡¡ ¡¡ Á ¡¡ ¡¡ ¡¡ ¡¡ ¡¡ ¡¡ ¡¡ ¡¡ ³ ³ ³ ³ ³ ³ ³ ³ ³ ³ ³Âs³ ³Á ³ ³ ³ ³ ³ ³ ³ ³ ³ ³ ³ ¸¸ ¸¸ ¸¸ ¸¸ ¸¸ ¸¸ ¸¸ ¸¸ ¸¸ ¸¸ ¸¸ ¸¸ ¸¸ ø¸ ¸¸ Á ¸¸ ¸¸ ¸¸ ¸¸ ¸¸ ¸¸ ¸¸ ¸¸ ¸¸ ¸¸ ¸¸ ¸¸ ¸¸ ¸¸ ¹¹ ¹¹ ¹¹ ¹¹ ¹¹ ¹¹ ¹¹ ¹¹ ¹¹ ¹¹ ¹¹ ¹¹ ¹¹ ¹¹ ¹¹ ¹¹Ä¹¹ ¹¹Å ¹¹ ¹¹ ¹¹ ¹¹ ¹¹ ¹¹ ¹¹ ¹¹ ¹¹ ¹¹ ¹¹ ¹¹ ¹¹ ¹¹ ¹¹ ¹¹ º º º º º º º º º º º º º º º º º ĺ º Å º º º º º º º º º º º º º º º
¡¢¡ ¡ £*¡ ¡ ¡¤¥ ¡ ¡ ¡ ¡ ¡ ¡¢*¡ ¡ ¤*¡ ¡ ¡¦ ¡ ¡ ¤´ ¯ ¤-£µ ¸±\¸ ¸ ¢¸ ¸ µ¸ ¸ ¸ ¸ ¸ ¸ ¸ ¸ ¸ ¸ ¸ ¸ ¸ ¸ ±\¸ ¸ ¸¢¸ ¸ ¢¸ ¸ ¸µ^¸ ¸ ¹ ±¹ ¹ ¤¹ ¹ ¹ ²*¹ ¹ ¹µ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹±¹ ¹ ¤-¹ ¹ £¹ ¹ ¹ µ» ¹ ¹ ºº ºº ´ºº ºº ºº ±ºº ºº ºº±\ºº ºº µ ºº ºº ºº ºº ºº ºº ºº ºº ºº ºº ºº ºº ºº ´ºº ºº ºº ±cºº ºº ¼*ºº ºº ºº µ ºº ºº ºº
¯
in the rectangle representing the band are its bandwidth. A dual-band receiver front-
end typically consists of a dual-band antenna, followed by a monolithic dual-band
filter and dual-band LNA ate provides gain and impedance match at two bands.
An input impedance match method using mutual coupled inductors will be stud-
ied first and measurement results for an LNA working in the 900 MHz GSM band will
be presented. Then the extension of a concurrent dual-band LNA will be proposed.
The advantage of concurrent operation is it does not need any dual-band switch or
duplexer and has maximum front-end circuit sharing which reduces the silicon area.
Concurrent reception is more desired when two bands provide different services such
matching, noise figure and gain specifications. A cascoded structure is usually used
in the LNA to reduce Miller effect on input impedance and improve reverse isolation.
ÐÒÑ ÍÏÎ
ËÌ
ËÇ Í Ê Ó
ÆÈÇ*ÉÊ Ô
ÕªÖ× Ó
ËÉ
1. Input Impedance
Fig. 93 shows the idea of using mutual inductance in the LNA input matching network.
Its small signal circuit for input impedance calculation is depicted in Fig. 94. It is
easily seen that the input impedance Zin can be written as the sum of inductor LG ’s
impedance sLG and the impedance ZX looking into the circuit after inductor LG :
VG1
ZX = Ii
ñ ðï
Ý Ù õôó Ý à
ìë
Þ ò
îí
ØÚÙÛ-Ü ê è é ç ø æ äªáªå âsã á
ß
û úù ò
ñ ö÷ Ý Û
M
k=√ (6.3)
L S LM
A typical range for the k-factor achievable in silicon designs is 0.6 ≤ k ≤ 0.95. Z in is
given by
1 gm1
Zin = s (LG + LS ) + + (LS ± M ) (6.4)
sCGS1 CGS1
For the mutual coupling polarity shown in Fig. 94 by the asterisk, minus sign
will be assigned to the third term in (6.4). This impedance expression resembles the
one without mutual inductance coupling except that the real part is modified by the
1
ωo = p (6.5)
(LG + LS ) CGS1
gm1
where ωT = CGS1
.
159
1
ωT ∝ µ (VGS1 − Vth1 ) (6.7)
L2
The device channel length L is usually chosen to be process minimum for high fre-
quency performance. CGS1 is calculated from the optimum input quality factor QI =
1
ωo CGS1 Rs
. Therefore the device width W is obtained by resolvingCGS1 = 23 W LCox .
Gate over drive voltage VGS1 − Vth1 can be fixed by power consumption constraint
ID ∝ W
L
(VGS1 − Vth1 )2 or linearity requirement IIP 3 ∝ VGS1 − Vth1 .
For a conventional inductive degenerated LNA, difficulty may arise from using
high bias level to obtain high input linearity. Input match condition requires R s =
ωT LS , where RS is usually 50 Ω or 75 Ω. For high gate bias level, VGS1 − Vth1 is
large, therefore ωT is also large. This situation will probably require a very small
by the designer and is affected by the placement of the die in the package which is
unknown at the very beginning of the design stage. Small LS will also require large
because LG is series directly with the gate. The presence of the mutual inductance
offers another degree of freedom for input impedance matching. By choosing the
negative sign in (6.4), Rs = ωT (LS − M ), larger LS can be used. If the design is
targeting minimum current consumption and if VGS1 − Vth1 is small, a large LS may
be required. In this case, the positive sign in (6.4) can be used to enable smaller
degeneration inductance LS .
160
CGD1
LG LM
*
+ CGS1 ro1
Zo VGS1 I o1 CGS2
_ gm1VGS1 1
g m2
Is *
LS
'
Zin2 Zo1 Zin2
2. Interstage Impedance
In Fig. 93, cascoded transistor M2 reduces Miller effect and provides output-input
shows the small signal equivalent circuit for inter-stage impedance calculation. Z o is
1
the source impedance. gm2
in parallel with CGS2 represents the input impedance of
the voltage gain from the gate of M1 to its drain terminal will be decreased due to a
reduced load impedance, thus reducing the Miller feedback.
It can be shown that M1 ’s drain current Io1 and source current Is have the fol-
lowing relationship
Is = F1 Io1 (6.8)
161
1
sLG ± sM + sCGS1
+ Zo
F1 = 1 (6.9)
sLG + sLs + sCGS1
+ Zo
At input matched condition, F1 can be simplified as
ωo
F1 = 1 − j (6.10)
ωT
ωo2 ωo
Zo1 = 2ro1 + L s + j Zo (6.11)
ωT ωT
gm2 0
where ωT 2 = CGS2
. By choosing a proper ωT 2 , Zin2 can have only the resistive part
1
and its value is smaller than gm2
. The effect is that there will be more current pumped
into the cascoded stage thus improving efficiency.
It can be shown that the transconductance of the first stage does not get affected
By choosing negative polarity of the inductive coupling, the voltage gain from M 1 ’s
gate to its drain is
¶2
ω2
µ ¶ µ
1 1 ωo
AV 1 = Gm1 2ro1 + o Ls k ´2 − ωT M (6.14)
ωT gm2 ωT
³
ωo
1+ ωT 2
This gain can be made smaller than that without mutual inductive coupling therefore
ý
üýþÿ
ÿ
þ
3. Effective Transconductance
The effective transconductance from the input terminal to the current flowing through
inductor LM can be found by observing the small signal equivalent circuit in Fig. 96.
VGS1 = QI Vi
Is = I o + I g
(6.15)
Ig = jωo CGS1 VCGS1
where
1
ωoG = √ (6.17)
LG CGS1
and
1
ωoS = √ (6.18)
LS CGS1
ωo2 2
It is easy to shown from (6.5), (6.18) and (6.17) that 2
ωoG
+ ωω2o = 1, so Gm is further
oS
163
LM LM
M1 * M1 *
LG vn2 LG
ind
2
1
CGS1 CGS1
in2
ing
2 * *
1 LS LS
Fig. 97. Equivalent input noise sources of the inductive coupled LNA
simplified into
µ ¶
ωT 1
Gm = −j (6.19)
ωo Zo
It is seen that the phase of effective transconductance is -90 degrees. In order to
4. Noise Analysis
Since in an LNA with the cascoded structure, the cascoded MOSFET has a little
effect on the noise performance of the whole circuit [55], its noise contribution will
be ignored in the following noise analysis. M1 ’s drain noise current i2nd1 and induced
gate noise current i2ng1 will be represented by a noise voltage vn2 and noise current i2n
as shown in Fig. 97.
ωo
in = ing1 + ind1 j (6.20)
ωT
164
and
ing1
vn = − (6.21)
jωo Cgs1
Since the mutual inductance does not appear explicitly in the above equations, it
does not make the expression forms of input equivalent noise sources different from a
αδ
Rn = (6.23)
5gm1
Gc ≈ 0 (6.24)
and à r !
|c| 5γ
Bc = −ωo Cgs1 1+ (6.25)
α δ
The minimum noise figure obtained for this circuit configuration is
2 ωo
q
γδ 1 − |c|2
¡ ¢
Fmim = 1 + √ (6.26)
5 ωT
and à r !
|c| 5γ
Bs = Bopt = ωo Cgs1 1+ (6.28)
α δ
1 1
Under perfect impedance match condition, Gs = Zo
= Rs
and Bs = 0, the noise figure
is à !
µ ¶2 r
αδ 1 ωo αδ γ 4
F =1+ + gm1 Rs + + |c| γδ (6.29)
5 gm1 Rs ωT 5 α 5
165
³ ´2
ωo
Several observations can be obtained from (6.29). The ωT
term shows the noise
figure will increase with operation frequency if other parameters keep constant. If
there is no correlation between the gate and drain noise current, i.e. c = 0, the noise
figure could be smaller. There is an optimal value of gm1 Rs product which makes the
flowchart. It is very similar with the design flow in Fig. 15, Chapter II. The major
difference is the input impedance matching involves mutual inductance and the design
VDD
LD1 RD
VB CC2
M2 LD2
CC3
LM
RF OUT
CC1 LG M1 *
M3
RF IN k
CGS1
RS
*
LS
Output Buffer
1
Determine Iden, Q and C gs =
2QRsωo
Calculate Device Size
I DS = W × I den
Rs
Ls ± M =
ωT
Calculate Lg, LS, M 1
Lg = 2 − Ls
and Required Load ωo C gs
ωo
RL = AV Rs
ωT
Simulation
Verification
F ↓→ ↑, Q ↑, I den ↓
ωT
ωo NO Meet the
IIP3 ↑→ Vgs − Vth ↑→ LS − M Specs ?
0.35µm CMOS technology and fabricated through MOSIS service. Its die micro-
The mutual coupled inductors LM and Ls was implemented on-chip by two in-
terleaved square spirals. The inductor LG was formed by the bond wire and off-chip
surface mount inductor. Multiple bond pads were used for ground connections to
reduce the ground inductance. The LNA occupies 700 µm × 500 µm active silicon
area.
Fig. 101 shows the measured small signal performance of the proposed LNA. The
gain (S21) is 17 dB at 960 MHz. The noise figure is 3.4 dB which seems a little higher
for GSM application. This is because it includes the output buffer formed by M3 ,
Rs and LD2 . Bonding pads and inferior quality factors of inductors also makes the
NF larger. Simulation shows that the LNA with better inductors and bonding pads
removed can have a noise figure about 1.4 dB. The LNA is tested within a plastic
package and soldered on a PCB board. The S12 is measured for the whole setup,
168
so the reverse isolation for the LNA itself should be better than what is shown in
Fig. 101. The measured IIP3 plot is illustrated in Fig. 102, which is -5.1 dBm. The
18 10
14
10 S21 9
6
8
2
-2 7
S11, S21 and S12 (dB)
Fig. 101. Measured small signal performance of the mutual-coupled degenerated LNA
Table XVIII summarizes the measurement results together with other reported
different working frequencies [33]. It basically attempts to merge two or more designs
into one compact circuit by sharing components. An issue of this method is that the
169
15
-45
-60
-75
-90
-45 -40 -35 -30 -25 -20 -15 -10 -5 0
Input power (dBm)
shut-off components may affect the active ones due to the parasitics introduced by
turned-off devices. The concurrent operation LNA can work at two or more modes
at the same time, but usually it is hard to provide optimal working condition for
both [51]. Introducing switches at the output of concurrent matched LNA can pro-
vide additional degrees of freedom for more frequency bands [52]. Because the switch
is at output, it will not affect the input matching condition. But linearity of the
switch and parasitics may offset the linearity and output working condition of the
circuit.
From (6.4) one can find that without disturbing other factors, using a different
LG value can provide input match around different operation frequency ωo . A strait
forward implementation is to switch in or out more inductance using switches. This
approach suffers from the added parasitic capacitance and resistance of the switches
and increased noise due to switch resistance. Without using switches, an LC parallel
network can be put in series with LG as shown in Fig. 103. The parallel LC network
(LB k CB ) will present inductive or capacitive impedance below or above its intrinsic
resonant frequency (ωB = √ 1 ), thus modify the effective inductance series with
L B CB
the gate. Mutual coupling between LG and LB can further modify the impedance
0
and also reduce the required value of inductance thus save area. LG represents the
size of LG can be used. The frequency point where the second term of (6.31) changes
171
+-4"4
%04 # 564
+-, 1)
(*) %042)
13
%'&
, ( #
% % .
(*3
.
12# . "!$# / 798=<?>0@
%0, /?A 5 !
.
798-:; % !
from inductive to capacitive does not depends on the mutual coupling. Therefore
by changing CB , the impedance property cross-over point can be changed and by
adjusting the coupling strength, proper matching at different frequency points can be
realized. Fig. 99 can be followed as the design procedure of the dual-band LNA by
considering the discussions about input impedance matching.
The dual-band LNA is design using the TSMC 0.35µm CMOS technology. Fig. 104(a)
is the simulated S11 and S21 plot of the dual-band LNA. The LNA’s input is matched
to 50 Ω at 900 MHz GSM band and 1800 MHz DCS-1800 band. The S21 for these
Fig. 104(b) shows the noise performance. The noise figure at the two bands
almost equals to the minimum noise figure, which means the input matching network
is also optimized for noise matching. The noise figure is 0.8 dB for 900 MHz band and
172
1.6 dB for 1800 MHz band. Fig. 105 shows the input-referred third-order intercept
point (IIP3) plots for both bands. In the 900 MHz GSM band, the IIP3 is -5.5 dBm.
For DCS-1800, the IIP3 is -2.5 dBm. A Q value of 10 for the inductors is assumed in
the simulation.
and the proposed dual-band LNA. It shows that the inductive-coupled LNA gives
comparable results to the literature.
Table XIX. Reported LNA performance in for cellular applications († indicates the
value is for the whole front-end)
CC
D$E
DGF
D$H l6o?m
jk I
hi
C
d eg
BC
d e
e f
BI
B D$H
lnm[m
B DGF
B D$E
B CC
IJHHLKHHMEHHMNHHOD$HHHPDD$HHPDGCJHHQD$RHHSDGFJHHPD$THHQDGIJHHPD$KHHQD$EHHPD$NHHUCJHHH
VXW Y[Z]\ ^`_ acb
(a)
v
u
t
s
r
q
p
v0pcpxwcpcpxycpcp{zcp'p|qJpcpcp}qcqpcpUq2r0p0pUq"scpcp~q2t0p0pUq"ucpcpq2v0pcp}q"wcp'p}q"ycpcpq"zcpcpr0p'pcp
X[
] `60
(b)
Fig. 104. Simulated small signal performance of the dual-band LNA (a) S11 and S21
(b) noise figure and minimum noise figure
174
20
IIP3=−5.5dBm
−20
Output power (dBm)
−40
−60
−80
−100
−40 −35 −30 −25 −20 −15 −10 −5 0
Input power (dBm)
(a)
40
20
0
Output power (dBm)
IIP3=−2.5dBm
−20
−40
−60
−80
−100
−40 −30 −20 −10 0 10
Input power (dBm)
(b)
Fig. 105. Simulated IIP3 of the dual-band LNA (a) GSM-900 (b) DCS-1800
175
CHAPTER VII
Ultra-wide band (UWB) systems are being considered excellent candidates for the
future-generation short-range, high-throughput wireless communications. It is emerg-
ing as a solution for the IEEE 802.15.3a (TG3a) standard [59] and will complement
(110, 200, 480 Mbps) and reasonable low power consumption. Therefore, UWB re-
quires CMOS design in order to achieve low power and low cost integration with other
devices, and to fulfill the vision of integrated connectivity [60].
Ultra-wide band systems transmit signals that demonstrate extremely low power-
spectral-density and occupy very wide bandwidth. The UWB signal bandwidth is
greater than 20% of its center frequency and must have a minimum value of 500 MHz
the time domain, which enables a path diversity gain by using a RAKE receiver.
Therefore UWB signals are immune to multi-path fading. According to Shannon
channel capacity theorem, the information a channel can carry [61] (channel capacity,
C) is
µ ¶
P
C = B log2 1 + (7.1)
BN0
176
where B is the channel bandwidth in hertz (Hz), P is the received signal power in
watt (W), N0 is the noise power spectral density in watt per hertz (W/Hz). In
order to increase the channel capacity, one can increase the transmitted power or
the signal bandwidth. But the channel capacity has a linear increase in bandwidth
bandwidth than higher signal power. With increased bandwidth, the transmitted
power can be reduced thus reducing the interference to other systems. As an overall
consequence, UWB can provide very high data rates at limited range using very low
signal power. Due to its extremely low power spectral density, UWB can co-exist with
other standards operating in its frequency bands. Fig. 106 shows the UWB spectrum
compared with the 802.11a signal spectrum.
[¤X« [c®`¯c¤[¤±°²¡?£´³
¨¨¨¨¨
¨¨¨¨¨
¨¨¨¨¨
¨¨¨¨¨ ¹cº¥º¸» ®[¼½¾ ©]¿ À ÁÂÀ ½
¨¨¨¨¨
¨¨¨¨¨ ¯ ¬ Ã Ä ÂÅ ·ÁÇÆ °²¡?£´³
¨¨¨¨¨
¨¨¨¨¨ µ¥¶¸· ¯$ª © ¢¡?£´³
¨¨¨¨¨
§ § § § § § § § § § § § § § § § § § § § § ¨§ ¨§ ¨§ ¨§ ¨§ § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § §
§ § § § § § § § § § § § § § § § § § § § § ¨§ ¨§ ¨§ ¨§ ¨§ § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § §
§ § § § § § § § § § § § § § § § § § § § § ¨§ ¨§ ¨§ ¨§ ¨§ § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § § §
* ¢¡£ © ª[« ©`¬=© X« © -¡?£ c¤¥ ¦] -¡?£ ¹ ¼È9ÉÊËÈÌËÍJÎ
The bandwidth resources available for UWB can be used two different ways.
Impulse radio was the original approach to UWB realization. It communicates with
gigahertz. Data could be modulated using either pulse amplitude modulation (PAM)
A more recent approach to UWB is a multi-banded system where the UWB fre-
quency band from 3.1 GHz to 10.6 GHz is divided into several smaller bands. Each of
these bands has a bandwidth greater than 500MHz to comply with the FCC definition
of UWB. Several companies like Intel, Texas Instruments and Time Domain support
this approach. The multi-banded approach has a much greater flexibility in coex-
istence with other wireless systems and is based on more conventional technologies.
The official UWB standard is still under development by several companies that have
already provided their proposals to the 802.15.3a task group. As an example, Texas
each band. The OFDM carries are generated using an 128-point IFFT/FFT.
Due to the wide bandwidth of UWB signal, direct conversion may be the best
receiver architecture. If an IF is to be used, it should be at least higher than 250
MHz. This high IF will increase the complexity, power consumption and cost of the
baseband circuit. The wide band nature of UWB signal makes the DC offset and low
frequency noise less problematic than that in a narrow band system. Fig. 107 depicts
a conceptual UWB receiver block diagram.
ø0ùú ó ÷ ö õ ô
â
ä´å[æ çËè[é éÄêËë
ìîí ×Ûï í ×
Ü[Ý?Þ àáß ðòñ9×ñ
ã
Ï?Ð Ñ[ÒÔÓ6ÑÕ Ñ[ÖØ× ä´å[æ çËè[é éÄêËë
ÙÚ Õ ×ÛÑÐ
ûÿ ýü
þ
The wide spread spectrum of UWB signal makes its dynamic range (10-12 dB)
much smaller than that of its narrow band counterpart (50 dB for Bluetooth and 66 dB
for IEEE 802.11b). It may be suitable to employ the so called software-defined radio
architecture where the ADC is put directly after the LNA. But for current technology,
a high sampling rate ADC is still a big challenge, so the frequency conversion scheme
In conventional circuit design, it is well known that there exists a fundamental limi-
i.e. each element only accounts for one dominant physical or electrical effect. Re-
sistor represents the heat dissipation, inductor represents energy storage of magnetic
field and capacitor represents energy storage of electric field. But in the real world,
especially at high frequency, when the dimensions of a circuit is comparable to the
lossy metal line at 30 GHz has all of the effects of heat dissipation, magnetic and
electric energy storage. No single one of the effects overwhelms others. Thus the
circuit will be treated as distributed. Distributed circuits have the potential of large
bandwidth and the gain will not trade with bandwidth but rather with time delay.
Transmission line (T-line) is the key element of distributed circuits. Fig. 108 is the
series inductance is due to the magnetic field effects and the shunt capacitance is
due to electric field coupling between the signal and ground lines. The loss in the
transmission media is modeled by series and shunt resistor. The R, L and C constants
are defined as per unit length circuit parameters. Note that this model only applies
Assuming steady state operation, V (z, t) = V (z) ejωt , I (z, t) = I (z) ejωt . V (z)
and I (z) are voltage phasor and current phasor respectively. It can be shown that
the voltage and current along the line fulfill the following equations
dV (z)
= (R + jωL) I (z) (7.2)
dz
dI (z)
= (G + jωC) V (z) (7.3)
dz
or
d2 V (z)
= γ 2 V (z) (7.4)
dz
d2 I (z)
= γ 2 I (z) (7.5)
dz
p
where γ = (R + jωL) (G + jωC) = α + jβ. γ is known as propagation constant, α
2π
is known as attenuation constant, β is usually called the wave number, and β = λ
.
180
√
β = ω LC (7.10)
ω 1
vp = =√ (7.11)
β LC
For low loss transmission line, it is generally assumed that R ¿ ωL and G ¿ ωC,
so Zc is approximately to be r
L
Zc ∼
= (7.12)
C
which is the same as the characteristic impedance of a lossless line. The propagation
p
γ= RG − ω 2 LC + jω (RC + LG)
Due to the low loss assumption, the RG term can be dropped off from the above equa-
181
Based on low loss assumption, only keep the first two terms of (7.13),
1√ √
µ ¶
R G
γ∼
= α + jβ = LC + + jω LC (7.14)
2 L C
The phase constant β is the same as the wave number in the lossless case. But the
R GZc
α= + (7.15)
2Zc 2
jω ∼ jω 1 ∼ 1
vp = = √ √ ¢=√ ¢√ =√
γ jω LC + 2 LC R
1
+G 1 R G
¡ ¡
L C
LC + 2j ωL
+ ωC
LC LC
(7.16)
a matched line, there is no reflected signal, the voltage along the line at point z + ∆z
can be expressed as
So the attenuation is
¯ ¯2
P (z + ∆z) ¯¯ V (z + ∆z) ¯¯
=¯ = e2α∆z
P (z) V (z) ¯
182
Expressed in dB form, it is
∆PdB
µ ¶
10 R
= 20α log10 e = + GZc (7.17)
∆z ln (10) Zc
lines.
√ √
β ω LC ω LC
vp √1 √1
LC LC
used to measure these values [63]. The S-parameter matrix for a lossy unmatched
where l is the physical length of the line, Ds = 2Zc Zo cosh γl + (Zc2 + Zo2 ) sinh γl. It
can be shown that s
(1 + s11 )2 − s221
Zc = Z o (7.19)
(1 − s11 )2 − s221
¶−1
1 − s211 + s221
µ
e−γl = ±K (7.20)
2s21
183
q
1 2
where K = 2s21
(s211 − s221 + 1) − 4s211 .
calculated by
R = < {γZc } (7.21)
½ ¾
γ
G=< (7.22)
Zc
L = = {γZc } ω −1 (7.23)
½ ¾
γ
C== ω −1 (7.24)
Zc
The key to successful design of distributed circuits is the modeling of transmission line
used in the circuits. One can have two ways to arrive at the desired circuit. One is
have a well modeled transmission line, then design the circuit based on that line. The
other way is to build a general macro model of transmission lines, design the circuit
using that model then implement the line to map the macro model. Both methods
require a good modeling. There are two types of transmission lines which are suitable
to be integrated on silicon, they are coplanar stripeline and micro-strip line. Among
them, coplanar stripline seems more suitable for the purpose. The following sessions
investigate these two different lines.
The structure of the silicon and coplanar stripline is demonstrated in Fig. 109 and
Fig. 110. They are the structure dimensional drawing and the rendered 3D drawing
in HFSS respectively. The substrate structure of the TSMC 0.18 µm CMOS process
is very complicated, the structure shown here is a simplified version. Some of the
layers are lumped together and their parameters are averaged over their thickness.
The spacing and width of the lines (M6) are chosen to be 5 µm [64]. The total length
184
of the line is 500 µm. The dielectric thickness between the bottom of metal 6 and
substrate under field oxide (FOX) is 8.15 µm. Using the line width comparable to the
thickness of the dielectric, the electric-field (E-field) will only penetrate the surface
of the lossy silicon substrate and at the same time large metal surface area is used
to conducting current. So loss due to lossy silicon substrate and metal skin effect is
For metal 1 and metal 6 the skin depth is about 0.6 µm at 30 GHz.
Permitivity
es s +0 7
ckn 0E
Thi ty 2.4
ivi
uct
400
AIR(1.0) ond
0 .6
0
M 6C
PASS3(7.9)
5
PASS1_2(4.2) 2 .6
M6 2.34
5
IMD5b(4.2) 5 5 5 0 .3
8
IMD5a(3.7) 1 .1
2 m
IMD1_4(3.77) 5 .5 m .c
ty 8oh
ti v i
0
1 .1 s
FOXILD(4.0) si
Re
c on
Sili
0
SILICON(12) 30
500 Unit: µ m
NOT DRAWN
80 ON SCALE
There exist two modes of operation for the coplanar stripline, even-mode and
odd-mode. Fig. 111 depicts the field distribution of these two operation modes. The
desired propagation mode is the odd-mode. It has better field confinement than that
of even mode, therefore less loss. Simulation shows that at 30 GHz the loss of the
The micro-stripline structure is depicted in Fig. 112. The ground plane is formed
185
File Edit View Coordinates Lines Surfaces Solids Arrange Options Window Help Plug-Ins
File Edit View Coordinates Geometry Data Plot Radiation Options Window Help
tsmc018_1
File Edit View Coordinates Geometry Data Plot Radiation Options Window Help
tsmc018_1
(a)
Waiting for hardcopy output to finish...
(b)
Waiting for hardcopy output to finish...
Fig. 111. Coplanar stripline operation mode (a) Even mode (b) Odd mode
186
by metal 1. The signal line is formed by metal 6, which is the highest level and thickest
metal available in this technology. There is only one operation mode for the micro-
stripline. Fig. 113 shows the field distribution. For line width from 5 µm to 25 µm,
the loss is approximately 0.44 dB/mm at 30 GHz. The smaller loss is due to better
Now comes the question: Which type of transmission line is more suitable for
silicon integration? Note that the transmission line will be loaded by transistors, thus
lowering its effective characteristic impedance. If the loaded input and/or output line
is required to be 50 Ω, then the line itself should be greater than 50 Ω. Fig. 114 shows
coplanar stripline can be controlled from around 66 Ω to 124 Ω. On the other hand,
the line width of micro-stripline is the only changeable variable by the designer. It
is observed that the wider the line, the smaller the impedance. In order to keep the
impedance greater than 50 Ω before loading, the line width should be smaller than
7 µm. For a larger impedance value, the line width will be even smaller. This will
File Edit View Coordinates Geometry Data Plot Radiation Options Window Help
mstrip 187
increase the line loss due to its high resistance. So the coplanar stripline is preferred
Waiting for hardcopy output to finish...
for distributed circuit design on silicon.
planar transmission lines. Artificial line uses lumped inductors and capacitors to
emulate the behavior of a real transmission lines. It generally requires high quality
inductors and capacitors which are usually not readily available in regular silicon
process, especially for inductors. Artificial line also has intrinsic cut-off characteris-
tics which will limit the operation bandwidth. Planar transmission lines discussed
in the previous section may be a better choice for large bandwidth and high opera-
tion frequency. The distributed LNA is implemented by periodically loading planar
transmission lines with active devices usually MOSFET transistors. Fig. 115 is the
schematic representation of a distributed LNA. Two transmission lines, the gate line
188
100
14
90
80
70
60
4 8 12 16 20 24 28
Line spacing (µ m)
(a)
65
60
55
50
Zo (Ohms)
45
40
35
30
25
20
4 9 14 19 24 29
(b)
and drain line are coupled unilaterally by MOS transistors. The portion between
the dashed lines is a unit section or cell. The amplifier shown in this figure has four
sections. Fig. 116 is the equivalent circuit for the unit section. For simplicity, the
T-lines are assumed to be lossless, the MOS transistor is modeled by its gate-source
& ( ' 6
5
+ ) . ,
* -
4 2
3
% $ 8 1 /
# & 7 0 !
"
Due to the periodic loading nature of the T-lines, Cgs and Cds can be treated
as uniformly distributed along the length (lg for gate line, ld for drain line) of the
unit cell. If the intrinsic gate line and drain have Lg , Cg and Ld , Cd as their dis-
A
>@? X Y
] Z\[ Z
9:9:9:9:99 ;:;:;:;:;;
9:9:9:9:99 ;:;:;:;:;;
9:9:9:9:99 ;:;:;:;:;;
9:9:9 ;:;:;
H K
FG IJ
R S
W
TVU T
<:<:<:<:<< =:=:=:=:==
<:<:<:<:<< =:=:=:=:==
<:<:<:<:<< =:=:=:=:==
N Q
E
BDC
LM OP
the length of the drain line. Under matched conditions both lines are terminated by
β d ld = β g lg = θ (7.32)
191
N gm Zcd −jN θ
Av = − e (7.33)
2
where N is the number of unit sections. It is observed that within the valid frequency
range of uniform loading assumptions made in (7.26) through (7.29), the amplifier has
a gain which is increasing linearly with the total number of sections and a constant
group delay which is also a linear function of N. So increase gain by adding more
sections will increase the time delay between input and output. The power gain can
be written
N 2 gm
2
Zcd Zcg
G= (7.34)
4
In practice, losses exist both in the transmission lines and the active loading
MOS transistors. The gate line is then loaded by Cgs and Rgs in series, and the drain
line is loaded by Cds and Rds in parallel. In silicon implementation, losses come from
the T-lines may also need to be considered. Therefore the propagation constants of
loaded gate line and drain line will have a real part αg and αd respectively. Imaginary
components will appear in the loaded characteristic impedances but their contribution
is usually small within the useful frequency range. Under the phase synchronization
Thus the gain does not increase monotonically with N and at a particular fre-
ln (αd ld ) − ln (αg lg )
Nopt, LN A = (7.36)
α d ld − α g lg
In a practical amplifier, the major losses is due to the gate loading, an upper
192
2
N≤ (7.37)
Rgs ω2C d
gs Zc
process. The T-lines are formed by coplanar striplines. This circuit draws 45 mA
from 1.2 V power supply. Fig. 117 is the layout view of the distributed LNA. Fig. 118
is the S-parameters plot. The flat gain bandwidth is from DC to 18 GHz with 2 dB
ripple. From DC to 14 GHz, the input and output return losses are better than 10 dB.
w:u
w:r
£
u
¡ ¢
r `
pu
p w:r
p w:u
p tvr
p tvu
p qsr
w q u x yzwvw{w|q}w:u}w|x~w|y{t
wtjqtjutDx
`
Dv
Fig. 119 is the noise figure plot of this 5-section distributed LNA. From 8 GHz to
193
20 GHz, the noise figure is better than 6 dB. A detailed noise analysis will be carried
out in the following text to show the noise performance. The loss due to T-lines
loading and T-line loss itself will be ignored for simplicity. Before the noise figure
can be calculated, one must obtain the noise contribution due to MOS transistors
and termination resistors ( ZsT and ZdT ) as well as the noise power available from the
source generator. The noise of load impedance ZdL belongs to the next stage, so it is
not included in the noise figure calculation.
¨¦
¨¥
Ð ¨¤
ÍÎ Ï
¨f¨
È
ÉÆ ÊËÌ §
Æ ÇÈ
ÄÅ ¦
¤
© ª « ¬¨®¯¨D©°¨jª±¨j«°¨@¬²©`®³©f©¯©fª´©
«
µ
¶·`¸
¹`·º`»@¼¾½n¿ÁÀÂ@Ã
The noise contributed from the transistors in different unit sections is assumed to
be uncorrelated with each other. Thus one can first calculate the noise power delivered
into the load due to the transistor in one unit section, then sum the noise power over
the whole N sections to obtain the total noise power generated by transistors in the
distributed circuit. For this purpose, Fig. 120 shows the noise model of a unit section
The major noise sources are the MOS transistor’s drain noise current ind and
194
åë
é ì êæè ç î
ï
í
×:×:×:×:×× Ø:Ø:Ø:Ø:ØØ
×:×:×:×:×× Ø:Ø:Ø:Ø:ØØ
×:×:×:×:×× Ø:Ø:Ø:Ø:ØØ
×:×:× Ü Ú Ù Ø:Ø:Ø Ü Ú Ù
Û Û
à Ý
Þ:ß
ä á
âã
Ö Ô Ó Ö Ô Ó
Ñ:Ñ:Ñ:Ñ:ÑÑ Õ Ò:Ò:Ò:Ò:ÒÒ Õ
Ñ:Ñ:Ñ:Ñ:ÑÑ Ò:Ò:Ò:Ò:ÒÒ
Ñ:Ñ:Ñ:Ñ:ÑÑ Ò:Ò:Ò:Ò:ÒÒ
Ñ:Ñ:Ñ Ò:Ò:Ò
ð ô³õ ñ ò øù
ö ó ÷
γ
i2nd = 4kT gm ∆f (7.38)
α
and
ω 2 Cgs
2
i2ng = 4kT δα ∆f (7.39)
5gm
respectively. The drain noise sources will be treated as uniformly distributed over the
unit section. That is to say the drain T-line is periodically driven by noise current
ind
ld
. The noise voltage developed across the load ZdL due to the drain noise source of
When the gate noise current is injected into the k-th section gate T-line at
position (k − 1) lg , the voltage generated across the load ZdL of the drain T-line due
195
The noise contributed from the k-th section gate noise is the sum of the forward and
reverse voltage:
+ −
vng, k = vng, k + vng, k
·
sin(k− 21 )θ −j (k− 1 )θ
¸ (7.43)
e−j (N −k+ 2 )θ
1
1
g Z g Z di N − k + 12 +
¡ ¢
= 4 m c c ng θ
e 2
Adding (7.40) and (7.43) together, the output noise voltage due to transistors in
the k-th section will be obtained as
" #
sin 2θ 1
ind + M (k)gm Zcg ing Zcd e−jθ(N −k+ 2 )
1
vno, k = (7.44)
θ 4
where
sin k − 21 θ −j (k− 1 )θ
µ ¶ ¡ ¢
1
M (k) = N − k + + e 2 (7.45)
2 θ
It is known that ing and ind are correlated with correlation coefficient c which is
purely imaginary and equals j0.395 for long channel devices. ing can be decomposed
into a component ingc which is fully correlated with ind and another component ingu
which is uncorrelated with ind :
where
i2ngu = 4kT Gu ∆f (7.47)
196
and
ω 2 Cgs
2 ¡
1 − |c|2
¢
Gu = δα (7.49)
5gm
√ √
ingc ingc i∗nd ing i∗nd i2ng i2
Fc ≡ ind = i i∗ = q 2 √ q 2 = c q 2ng
nd nd ind i2ng ind ind
q (7.50)
δ ωCgs
= j |c| α 5γ gm
Using the above notations, the noise power delivered to the load by transistors
The output noise power contributed by the drain T-line termination impedance
ZdT is
PndT = kT ∆f (7.52)
The gate T-line termination impedance ZgT contributes to the output noise power
in the form of
¯ sin N θ ¯2
¯ ¯
PnsT = kT G ¯¯ ¯ ∆f
Nθ ¯
The noise power available from the signal source is
Pns = kT ∆f (7.53)
In order to gain insights, a simplified version of (7.55) is needed. For large number of
sections, one can assume that the first term in (7.45) dominates, so two summations
2 N3
ΣN
k=1 |M (k)| ∼
3
This essentially removes the correlation between the gate and drain noise current
sources. The second and third term in (7.55) are small quantities for a large N,
gm
where ωT = Cgs
.
The sinc function in (7.56) dominates the low frequency portion of the noise
figure. Between DC and the first null, the sinc is a descending function of frequency.
198
Thus the noise figure decreases with frequency increasing as shown in Fig. 119. When
frequency increases the gate noise contribution denoted by the ω 2 term in (7.56) be-
comes more significant and at the same time the gain drops from its ideal value,
therefore noise figure goes up again with frequency increasing. [67] gives the noise
figure analysis for distributed amplifiers using artificial transmission lines. The sim-
plified form of noise figure is almost the same as derived here except that there is no
sinc function in [67]. For the exact noise figure calculation, [67] did not consider the
fact that the forward and reverse propagated gate noise currents are actually fully
correlated.
Distributed mixers are constructed the same way as distributed amplifiers. Three
types of distributed mixers are studied in [68]. Cascoded mixer mimics the dual-gate
FET mixer. Matrix mixer utilizes three tiers of transistor to implement RF ampli-
fication, mixing and IF amplification. Balanced mixers use CG-CS FET pairs. By
injecting signals from different terminals of a FET, different mixing modes can be ob-
tained, [69] studied four different distributed mixing modes of FET. It demonstrated
Fig. 121 is a typical distributed mixer block diagram. Ignore losses introduced by
the T-lines and resistive loading by the transistors, assume the loaded characteristic
ZcRF and βRF for RF line, ZcLO and βLO for LO line, ZcIF and βIF for IF line. And
further assume all the T-lines are terminated by their characteristic impedances. The
199
where VRF is the signal incidence into the RF port. The conversion transconductance
gmc of a unit section is uniformly spread over the unit length lIF of IF line. The LO
signal traveling along the LO T-line can be modeled by the phase-shift of conversion
dist
transconductance. Therefore the distributed conversion transconductance gm (xLO )
is given by
dist gmc −jβLO xLO
gm (xLO ) = e (7.60)
lIF
& $%
ü
-+, -+, -+,
*() ' *() '
#
!"
ECD H B 55 55 B >< = ;
22 22 F1G ?A@ 55 55 ?A@
22 22 44 44 .. .. 55 33 33 33
/10 22 44 44 .. .. 33 33 33
44 .. 333
>
<= ; : 6
798
þþ þþ
ýý ýý ýý þþ þþ
û
ü ýý ýý ýý ÿÿ ÿÿ ÿÿ úú úú úú
ÿÿ ÿÿ ÿÿ úú úú úú
L
The total voltage developed across the IF load impedance ZIF can be calculated
by
N lIF
1
Z
dist ∗
VIF = gm (xLO ) vRF (xRF ) ZcIF e−jβIF (N lIF −xIF ) dxIF (7.61)
2 0
∗
where down-conversion is assumed, N is the number of sections, and vRF (xRF ) rep-
200
where Θ = θLO − θRF − θIF . If (LLO , CLO ), (LRF , CRF ) and (LIF , CIF ) are the
distributed parameters of the loaded LO, RF and IF T-lines respectively, then one
can write:
p
θLO = βLO lLO = ωLO lLO LLO CLO (7.65)
p
θRF = βRF lRF = ωRF lRF LRF CRF (7.66)
p
θIF = βIF lIF = ωIF lIF LIF CIF (7.67)
such that
p p p
lLO LLO CLO = lRF LRF CRF = lIF LIF CIF (7.68)
Notice that ωIF = ωLO − ωRF and combined with (7.68), then Θ = 0 and the
voltage conversion gain has constant amplitude and group delay over a valid frequency
range:
N
Avc = gmc ZcIF e−jN θIF (7.69)
2
Similar to the distributed LNA voltage gain, the conversion gain and group delay is
Consider the loss in the circuit, αRF , αLO and αIF are the attenuation constants
of loaded RF, LO and IF T-lines. Furthermore, assume that the loss in the LO signal
approximately
−N (αIF lIF ) ¯2
g 2 Z IF Z RF
¯ −N (α l +α l ) ¯
¯e RF RF LO LO
− e
Gc ≈ mc c c ¯ ¯ (7.70)
4 ¯ αRF lRF + αLO lLO − αIF lIF ¯
If the LO signal is large enough such that its attenuation along the line will
not affect the conversion transconductance, then the power conversion gain is not
dependent on αLO :
−N αIF lIF ¯2
g 2 Z IF Z RF
¯ −N α l ¯
¯e RF RF
− e
Gc ≈ mc c c ¯ ¯ (7.71)
4 ¯ αRF lRF − αIF lIF ¯
This resembles the distributed LNA power gain equation (7.35), so the optimal num-
ber of sections is
ln (αRF lRF ) − ln (αIF lIF )
Nopt, mixer = (7.72)
αRF lRF − αIF lIF
and if the loss is dominated by the RF T-line loading due to the MOS transistor’s
gate resistance, (7.37) can also be applied to determine the upper bound of N by
In the wide-band LNA design, the input impedance matching is another major prob-
lem additional to bandwidth. There are several ways to provide matching. One
solution is to use common gate or common base as input stage. Another way is to
use feedback. These two methods can only work for a moderately wide bandwidth.
using artificial or real transmission lines have good wide band matching property.
It provides 50 Ohm input and output intrinsically. Even for non-distributed circuit,
transmission line matching network can also achieve relatively large bandwidth for
prescribed matching requirement. For on-chip implementation, especially at lower fre-
quency range (several GHz) transmission line’s dimension will be prohibitively large.
Fig. 122 shows the detailed procedure to match a specific input impedance (s11 )
over a specific frequency band to around the center of Smith chart using lumped
components.
First, for a start point, the input impedance of a unilateral MOS transistor is
2
Cgs = Cox W Lov + Cox W Lef f (7.73)
3
where Lov is the gate-source overlap length, Lef f is the effective gate length which
equals L − 2Lov . L is the drawn length of the gate. The resistance is due to poly
resistance Rpoly and non-quasi-state (NQS) gate resistance Rnqs . The poly resistance
comes from the physical resistance of gate poly material and will contributes noise.
In order to reduce the gate poly resistance, multiple transistors usually connect in
parallel instead of using just one big transistor. Suppose a transistor comprises num-
ber of n smaller transistors with channel length L and channel width W , then the
gate poly resistance can be calculated from
Rsh, poly W
Rg, poly = (7.74)
12n2 L
203
L2
Step1
TLine L2
Cc Cc
M1 M1
L1
C1 RB RB
S11
S11
VB VB
Finish
L2
Cc
M1
L1
C1 RB
S11
VB Cc
Step3 M1
L2 RB
Cc S11
M1
VB
Start
L1
S11 C1 RB
Step2 VB
where Rsh, poly is the gate poly sheet resistance. Here it is assumed the both ends of
The NQS effect can be modeled by a resistor Rnqs in series with the gate capacitor
Cgs [70]. Note that resistor Rnqs does not contribute to noise, because it is not a
1
physical resistor. The value of Rnqs is approximately 5gm
[70].
The start point impedance is plotted in the Smith chart over the desired fre-
quency range. The impedance curve is located at position Start in Fig. 122. Then
1. Series a proper value of inductance such that the conductance at the frequency
edges have the same real part. The impedance curve should move to the position
2. Parallel an inductor and capacitor tank to bring the frequency edges close to
each other. The impedance contour resembles a circle and moves to the position
3. Series an inductor again to move the center of the impedance contour to a pure
resistance point as at the position Step3 in Fig. 122. This inductor can be
used to rotate the impedance contour to the center of the Smith Chart, as shown at
the position Finish in Fig. 122. The output impedance matching can be achieved
similarly if required.
205
cussed in the previous section is shown in Fig. 123. This circuit is simulated using
TSMC 0.18 µm CMOS technology. It draws 5 mA from a 1.8 V power supply.
IK]
L
z {
w
z~} LON
^
{
S TWVYX Z [YX]\^V]_`[
U w
Z XVbaWcWdfe cgc]e hYaji e a][
w|{
z~}
x
LONQP R
wyx
z x LMJ
kmlonYpbqsr tvu
IKJ
Fig. 124 gives the S-parameters and noise figure plots from 2.8 GHz to 5.0 GHz.
Table XXI summarizes the post-layout simulation results of this wide-band LNA. This
LNA can cover the first 3 bands (from 3.168 GHz to 4.752 GHz) of Texas Instruments
OFDM UWB proposal and can be used in its Mode 1 UWB devices [71].
206
20 1.6
15
S21
10 1.5
5
0 1.4
NF
-5
S22
-10 1.3
-15
-20 S11 1.2
-25
-30 1.1
-35 S12
-40 1
2.8 3 3.2 3.4 3.6 3.8 4 4.2 4.4 4.6 4.8 5
Frequency (GHz)
S21 > 12 dB
Power 9 mW
207
CHAPTER VIII
In the dissertation, LNA design aspects and the major design considerations were
first reviewed. In order to implement a good design, a lot of information from the
This makes the seemingly simple circuit of a LNA actually the trickiest part in the
front end of a receiver system. Two design approaches exist. One is the microwave
approach, which first selects a specific microwave transistor according to the data
sheet obtained from manufacture. The choice of device relies on its gain, noise and
linearity capability. Then trade-offs are made for gain, impedance match and noise
match. The matching network at input and output will finally be implemented con-
sidering the above merits and stability. The way RF IC designers are adopted is more
complicated. In IC design, one has the freedom to change the geometry of the devices
as well as bias condition. The microwave design target is to find an optimal matching
network for a specific device, but the IC designers want to find an optimal structure
for all possible devices. Of course, this is not always possible, but the freedom of
varying device sizes really gives more room for IC designers. Another consideration is
that if the LNA does not require driving off-chip filters, it is not necessary to match
its output to 50 ohm or even there is no need to realize matching network between
the output of the LNA and the input of the next stage (usually mixers). Of course,
with three different frequencies. Mixer translates the amplified signal form the LNA
to a frequency band suitable for the baseband circuit to process. Conversion gain
208
shows the mixer’s frequency translation efficiency. A mixer is more noisy than a
LNA. Especially for low-IF applications, low frequency flicker noise dominates the
mixer’s noise performance. It is vital to choose a proper IF frequency such that the
MOS transistor’s flicker noise corner is lower than the IF frequency. Mixers in sili-
to reduce the flicker noise effect of the current switching pairs. Measurements show
that the mixer worked properly with other blocks in the whole system.
induced noise reduction and bipolar cascoded techniques and the down-conversion
mixer has combined RF driver and bipolar current switching pairs. A PTAT biasing
circuit is also included for stable bias condition setup. The dual-mode receiver with
this front-end has sensitivity for Bluetooth mode -91 dBm and for WiFi mode (11
More research effort was exerted on LNA design techniques. With the devel-
opment of the IC fabrication, bipolar transistor will be readily available in CMOS
process. For the LNA design, the possibility of using both MOS transistor and bipolar
transistor was explored. This hybrid structure was inspired by the multi-gated MOS
configuration for linearization purpose. It is the first structure that MOS transistor
and BJT are working together at the signal path. A mutual inductive degenerated
LNA designed for dual-band application was studied to show the benefits of mutual
inductive coupling in the LNA design.
UWB technology can provide costumers more bandwidth, high speed and versa-
tile services. It has enormous potential applications, especially in the wireless PAN
209
due to their inherent wide-band nature. For lower frequency bands, lumped-matched
wide-band LNA occupies less silicon area and is more power efficient. A successful
design of a UWB system will require multi-disciplinary knowledge, such as channel
measurements, antenna theory, RF and high speed analog design, software devel-
opment and system simulation. For the RF front-end portion, the LNA/antenna
co-design may be needed for optimal performance.
for a low-IF Bluetooth receiver and a direct conversion Bluetooth/WiFi dual mode
and mixer, and the proposed procedure to achieve wide band impedance match using
lumped components.
210
REFERENCES
[2] R. Ludwig and P. Bretchko, RF Circuit Design Theory and Applications. Upper
Saddle River, N.J.: Prentice Hall, book and CD-ROM ed., Nov. 1999.
[3] T. H. Lee, The Design of CMOS Radio Frequency Integrated Circuits. New York:
[4] B. Razavi, RF Microelectronics. Upper Saddle River, NJ: Prentice Hall, 1998.
[5] J. Rogers and C. Plett, Radio Frequency Integrated Circuit Design. Boston, MA:
for RF applications,” IEEE Int. Symp. on Circuits and Systems, vol. 4, pp. 349–
352, May 1998.
CMOS LNAs,” IEEE Int. Solid-State Circuits Conf., vol. 1, pp. 406–407, Feb.
2002.
[10] D. K. Shaeffer and T. H. Lee, “A 1.5V, 1.5GHz CMOS low noise amplifier,”
and Digital Signal Processing, vol. 48, pp. 835–841, Sept. 2001.
[13] V. Geffroy, G. D. Astis, and E. Bergeault, “RF mixers using standard digi-
tal CMOS 0.35µm process,” IEEE MTT-S Int. Microwave Symp. Dig., vol. 1,
[14] T. Yamaji and H. Tanimoto, “A 2 GHz balanced harmonic mixer for direct-
conversion receivers,” IEEE Custom Integrated Circuits Conf., pp. 193–196, May
1997.
front ends of fully integrated receivers,” IEEE Trans. On Circuits and Systems-
II: Analog and Digital Signal Processing, vol. 45, pp. 269–282, Mar. 1998.
[18] S.-G. Lee and J.-K. Choi, “Current-reuse bleeding mixer,” Electronics Letters,
[20] H. Aoki and M. Shimasue, “Channel width and length dependent flicker noise
cuits and Systems-II: Analog and Digital Signal Processing, vol. 46, pp. 315–325,
March 1999.
ers,” IEEE J. of Solid-State Circuits, vol. 34, pp. 772–783, June 1999.
[24] R. Shepherd, “Bluetooth wireless technology in the home,” Electronics and Com-
munication Engineering Journal, vol. 13, pp. 195–203, Oct. 2001.
[25] P. S. Henry and H. Luo, “WiFi: what’s next?,” IEEE Communications Magazine,
[26] D. Coursey, “Bluetooth vs. WiFi: Why it’s NOT a death match.”
http://techupdate.zdnet.com/techupdate/stories/main/0,14179,2868374,00.html,
May 2002.
213
receiver in 0.25µm BiCMOS,” IEEE Int. Solid-State Circuits Conf., vol. 47,
pp. 270–271, 527, Feb. 2004.
and SiGe LNA’s and mixers for wireless LAN application,” IEEE Conf. on Cus-
tom Integrated Circuits, pp. 531–534, May 2001.
[29] Q. Huang, P. Orsatti, and F. Piazza, “Broadband, 0.25 µm CMOS LNAs with
sub-2dB NF for GSM applications,” Proc. of the IEEE Custom Integrated Cir-
[30] Y. Cheng, M. Chan, K. Hui, M. chie Jeng, Z. Liu, J. Huang, K. Chen, J. Chen,
[31] A. Hastings, The Art of Analog Layout. Upper Saddle River, NJ: Prentice-Hall,
Inc., 2001.
[32] A. M. Niknejad, “ASITIC: Analysis and simulation of spiral inductors and trans-
formers for ICs.” http://rfic.eecs.berkeley.edu/∼niknejad/asitic.html.
mixer for a WCDMA receiver,” IEEE J. of Solid-State Circuits, vol. 38, pp. 1045–
1050, Jun. 2003.
214
[35] D. Johns and K. Martin, Analog Integrated Circuit Design. New York: John
[37] B. Leung, VLSI for Wireless Communication. Upper Saddle River, N.J.: Pren-
[40] B. Kim, J.-S. Ko, and K. Lee, “Highly linear CMOS RF MMIC amplifier using
multiple gated transistors and its volterra series analysis,” IEEE MTT-S Int.
[41] T. W. Kim, B. Kim, and K. Lee, “Highly linear receiver front-end adopting
MOSFET transconductance linearization by multiple gated transistors,” IEEE
[42] B. Kim, J.-S. Ko, and K. Lee, “A new linearization technique for MOSFET RF
amplifier using multiple gated transistors,” IEEE Microwave and Guided Wave
[43] T. W. Kim, B. Kim, I. Nam, B. Ko, and K. Lee, “A low-power highly linear
ment (revised),” IEEE Microwave and Wireless Components Letters, vol. 13,
gated transistors,” Proc. of 4th Int. Conf. on ASIC, pp. 670–672, Oct. 2001.
[46] T. W. Kim, B. Kim, and K. Lee, “Highly linear RF CMOS amplifier and mixer
adopting MOSFET transconductance linearization by multiple gated transis-
tors,” IEEE Radio Frequency Integrated Circuits Symp., pp. 107–110, Jun. 2003.
[47] V. Aparin and L. E. Larson, “Modified derivative superposition method for lin-
earizing fet Low noise amplifiers,” IEEE Radio Frequency Integrated Circuits
(RFIC) Symp., June 2004.
[48] S. Yan and E. Sánchez-Sinencio, “Low voltage analog circuit design techniques:
[49] Y. Ding and R. Harjani, “A +18dBm IIP3 LNA in 0.35µm CMOS,” IEEE ISSCC
[50] V. Aparin, E. Zeisel, and P. Gazzerro, “Highly linearSiGe BiCMOS LNA and
[53] J. Long, N. Badr, and R. Weber, “A 2.4GHz sub-1 dB CMOS low noise amplifier
with on-chip interstage inductor and parallel intrinsic capacitor,” IEEE Radio
[54] W.-S. Kim, X. Li, and M. Ismail, “A 2.4 GHz CMOS low noise amplifier using an
[55] W. Guo and D. Huang, “The noise and linearity optimization for a 1.9-GHz
CMOS low noise amplifier,” IEEE Asia-Pacific Conf. on ASIC, pp. 253–257,
Aug. 2002.
[58] K. Sharaf and H. ElHak, “A compact approach for the design of a dual-band
low-noise amplifier,” Proc. of the 44th IEEE 2001 Midwest Symp. on Circuits
217
[59] “IEEE 802.15WPAN high rate alternative PHY task group 3a (TG3a).”
http://www.ieee802.org/15/pub/TG3a.html.
Radio Frequency Integrated Circuits (RFIC) Symp., pp. 497–500, Jun. 2003.
[62] M. Z. Win and R. A. Scholtz, “Impulse radio: How it works,” IEEE Communi-
cations Letters, vol. 2, pp. 36–38, Feb. 1998.
oscillator design,” IEEE J. of Solid-State Circuits, vol. 36, pp. 1480–1488, Oct.
2001.
House, 1993.
[67] C. S. Aitchison, “The intrinsic noise figure of the MESFET distributed ampli-
fier,” IEEE Trans. on Microwave Theory and Techniques, vol. 33, pp. 460–466,
Jun. 1985.
218
[70] Y. Tsividis, Operation and modeling of the MOS transistor. New York: McGraw
Hill, second ed., 1999.
[71] A. Batra, Texas Instruments et al., “Multi-band OFDM physical layer proposal,”
[72] M. Schetzen, The Volterra and Wiener Theories of Nonlinear Systems. New
APPENDIX A
CONSTANT CIRCLES
where z is the point on the circle, c is a complex constant number, b is a real constant
Adding |c|2 to both sides of (A.1) leads to |z|2 − c ∗ z − cz ∗ + |c|2 = b + |c|2 . Since
q
2 2 2 2 2
|z| − c ∗ z − cz ∗ + |c| = |z − c| , then |z − c| = b + |c| and |z − c| = b + |c|2 .
Because b + |c|2 > 0, it is the equation for a circle with center located at point c and
q
having radius r = b + |c|2 .
Taking the input mismatch factor (??) for example. We will prove that for fixed value
of s11 and Gs , Γs resides on a circle and find its center and radius.
1 − |s11 |2 1 − |Γs |2
¡ ¢¡ ¢
gs =
|1 − s11 Γs |2
Finally,
Notice that gs is a real number and 0 < gs < 1, according to (A.1), (A.2)
(1 − gs ) − |s11 |2
b=
1 − |s11 |2 (1 − gs )
The proof for output constance circle can be carried out similarly.
We want to show that (2.66) is a circle equation for variable Γs . Rewriting (2.66) as
and then
at last
¶∗
N − |Γopt |2
µ µ ¶
2 Γopt Γopt
|Γs | − Γs − Γ∗s = (A.4)
1+N 1+N 1+N
According to (A.1), (A.4) describes a circle with
Γopt
c=
1+N
221
APPENDIX B
Vito Volterra first studied the functional series named after him in the 1880s as a
Wiener applied the Volterra theory to the analysis of a series RLC circuit with a non-
linear resistor to a white Gaussian excitation. The systematic study of the application
of the Volterra series to non-linear system was conducted by J. F. Barrett in 1957.
D. A. George extended Barrett’s work and developed a system algebra and used
A linear, causal and time-invariant system with memory can be described by the
convolution integral
Z ∞
y (t) = h (τ ) x (t − τ ) dτ (B.1)
−∞
where h (t) is the impulse response of the system. A non-linear system without
memory can be represented using a Taylor series
A Volterra series combines (B.1) and (B.2) to describe a non-linear system with
223
memory [72]
in which
Z ∞ Z ∞
Hn [x (t)] = ··· hn (τ1 , · · · , τn ) x (t − τ1 ) · · · x (t − τn ) dτ1 · · · dτn (B.4)
−∞ −∞
is called the n-th order Volterra operator. hn (τ1 , · · · , τn )is the n-th order Volterra
1 X a
h(s)
n (τ1 , · · · , τn ) = h (τ1 , · · · , τn ) (B.6)
n! P n
where the summation runs through all possible permutations of the n τ ’s.
To gain more insights, one would like to study the non-linear system in fre-
quency domain. The basic integral transform, Fourier transform can be extended
to the Volterra series representation. For an n-th order kernel hn (τ1 , · · · , τn ), its
n-dimensional Fourier transform is
Z ∞ Z ∞
Hn (jω1 , · · · , jωn ) = ··· hn (τ1 , · · · , τn ) ej(ω1 τ1 +··· +ωn τn ) dτ1 · · · dτn (B.7)
−∞ −∞
If the input’s Fourier transform is X (jω), then the Fourier transform of the n-th
224
1
R∞ R∞
Yn (jω) = (2π)n−1 −∞
··· −∞
Hn (jω − jν1 , jν1 − jν2 , jν2 − jν3 , · · · , jνn−1 )
Y (f ) = H1 (f ) X (f )
R∞
+ −∞ H2 (f − f1 , f1 ) X (f − f1 ) X (f1 ) df1
R∞ (B.9)
+ −∞ H3 (f − f1 , f1 − f2 , f2 ) X (f − f1 ) X (f1 − f2 ) X (f2 ) df1 df2
+ ···
In a general non-linear system or circuit, the memory effect and non-linearity are not
distinctly separated, which makes the circuit representation rather complicated even
the following filters. For example, a lot of circuits can be approximated as a non-
linear memoryless gain stage with input and output filters, as shown in Fig. 125. The
non-linear gain stage is represented by the Taylor series as in (B.1). The first order
response of this system is simply
The second and third order responses are depicted in Fig. 126. For the second order
response, the input filter is evaluated at two different frequencies, then the response
are multiplied and further scaled by the second order gain and output filter response
225
Memory
but linear
X Y
F ( jω ) A(x ) G ( jω )
Non-linear
but memoryless
frequency components of the non-linear system. Suppose the input has m frequency
components
where ωp1 , ωp1 , . . ., ωpn choose from all the possible n out of m permutations of ±ω1 ,
±ω2 , . . ., ±ωm . X n represents the corresponding amplitude Xp1 Xp2 · · · Xpn . The op-
erator ◦ generates all the items cos (ωp1 + ωp2 + · · · + ωpn ) and scale their magnitudes
226
F ( jω 1 )
K2 G ( jω 1 + jω 2 )
F ( jω 2 )
F ( jω 1 )
F ( jω 2 ) K3 G ( jω 1 + jω 2 + jω 3 )
F ( jω 3 )
with |Hn (jωp1 , jωp2 , · · · jωpn )| and modify their phases with ∠Hn (jωp1 , jωp2 , · · · jωpn ).
The short-hand notation (B.14) is usually used for actual circuit non-linear analysis.
In this section, a detailed procedure for deriving a MOS differential pair’s Volterra
series is given as an example [37]. The MOS differential under study is shown in
Fig. 127. This seemingly simple circuit can not be represented using the circuit model
representation given in Fig. 125, so full Volterra expansion has to be employed. The
small signal input to the differential pair is assumed to have only differential voltage
vd . The input common-mode voltage provides DC bias. The small signal voltages
vd
applied to the gates of M1 and M2 are 2
and − v2d respectively. The tail current source
is assumed to be ideal and provides DC bias current Iss . The small signal voltage at
227
id
vd vd
2 2
M1 M2
vs
Iss Cs
the common-source node is vs . The only memory effect is introduced by the parasitic
capacitance Cs at node vs . All other capacitances are ignored for simplicity. The
MOS transistors are modeled as square law devices and
The goal is to obtain the Volterra series expansion for the small signal drain
this end, the Volterra series of the common-source voltage vs will be determined first,
i.e. vs will be expanded into
vs = G1 ◦ vd + G2 ◦ vd2 + G3 ◦ vd3 + · · ·
(B.16)
= vs1 + vs2 + vs3 + · · ·
where vsk = Gk ◦ vdk , is the k-th order term of the Volterra series of vs . Applying KCL
dVs K£
(Vgs1 − Vt )2 + (Vgs2 − Vt )2
¤
Cs + Iss = (B.17)
dt 2
228
where Vs , Vgs1 and Vgs2 can be written into DC and signal small signal terms
Vs = V S + v s (B.18)
and further express vgs1 and vgs2 in terms of differential input voltage vd and common-
source voltage vs
vd
vgs1 = − vs (B.21)
2
vd
vgs2 = − − vs (B.22)
2
Substituting from (B.18) to (B.22) into (B.17) and simplifying it,
dvs K
Cs + 2gm vs − Kvs2 = vd2 (B.23)
dt 4
K 2
(jωCs + 2gm ) (vs1 + vs2 + vs3 + · · ·) − K (vs1 + vs2 + vs3 + · · ·)2 = v (B.24)
4 d
The Volterra kernel Gk can be obtained by equating the same order term of vd
at both sides of (B.24). Keeping only the first order terms
Hence
G1 (ω) = 0 (B.26)
K 2
(jωCs + 2gm ) (vs2 + vs3 + · · ·) − K (vs2 + vs3 + · · ·)2 = v (B.27)
4 d
229
K 2
[j (ω1 + ω2 ) Cs + 2gm ] G2 (ω1 , ω2 ) ◦ vd2 = v (B.28)
4 d
Thus
K
4
G2 (ω1 , ω2 ) = (B.29)
j (ω1 + ω2 ) Cs + 2gm
Cs is parasitic capacitance, so usually (ω1 + ω2 ) Cs ¿ 2gm . (B.29) can be simplified
by taking its Taylor expansion and only retain the first two terms
· ¸
K Cs
G2 (ω1 , ω2 ) ≈ 1 − j (ω1 + ω2 ) (B.30)
8gm 2gm
G3 (ω1 , ω2 , ω3 ) = 0 (B.32)
Now, vs has been expanded into Volterra series up to the third order. Note that
Because of the operator ◦, (B.33) actually consists of four terms corresponding four
different cases: ω1 , ω2 = ±ωa , ±ωb .
Remember that the final goal is to obtain the Volterra series expansion for id in
terms of vd , i.e.
id = H1 ◦ vd + H2 ◦ vd2 + H3 ◦ vd3 + · · ·
(B.34)
= id1 + id2 + id3 + · · ·
230
Using the MOS device equation, the small-signal output current id can be related to
K ³ vd ´2 ³v
d
´
id = − v s + gm − vs (B.35)
2 2 2
1
id1 = gm vd (B.37)
2
Thus
1
H1 (ω) = gm (B.38)
2
This is the transconductance of the differential pair for a single-ended output.
Isolating the second order terms in (B.36)
K 2
id2 (ω1 , ω2 ) = v − gm G2 (ω1 , ω2 ) ◦ vd2 (B.39)
8 d
So " #
K 1 K Cs
H2 (ω1 , ω2 ) = 1− Cs
≈ j (ω1 + ω2 ) (B.40)
8 1 + j (ω1 + ω2 ) 2gm 16 gm
To obtain the third order kernel, separating the 3rd order terms from (B.36)
K2
· ¸
Cs
H3 (ω1 , ω2 , ω3 ) ≈ − 1 − j (ω1 + ω2 + ω3 ) (B.42)
16gm 3gm
This leads to the end of the Volterra series expansion of drain current id in terms
of differential input vs . Table XXII compares the results obtained from the above
analysis with low frequency Taylor series analysis. It is seen that Taylor series does
not reveal the high frequency second-order nonlinearity for a differential pair.
H0 + H1 ◦ vi + H2 ◦ vi2 + H3 ◦ vi3 + · · · =
n
IQ 1 + I1t [−H0 + (ge − H1 ) ◦ vi − H2 ◦ vi2 − H3 ◦ vi3 + · · ·]
(B.43)
2
+ 2I12 [−H0 + (ge − H1 ) ◦ vi − H2 ◦ vi2 − H3 ◦ vi3 + · · ·]
t
o
3
+ 6I13 [−H0 + (ge − H1 ) ◦ vi − H2 ◦ vi2 − H3 ◦ vi3 + · · ·] + · · ·
t
232
In the above expression, the frequency variables of Hk are dropped for conciseness.
hold:
µ ¶
ω
It (ω) = It0 1+j (B.44)
ωπ
It0 = φt ge (B.45)
ge
ωπ = (B.46)
Cπ
1
Re = (B.47)
ge
The Volterra kernels can be calculated by equating the same order terms from both
sides of (B.43). The zero-th order or DC term will be calculated first, then the first
H0 H0 H2 H3 H
− 0
=1− + 02 − 03 + · · · = e Ito (B.48)
IQ It0 2It0 6It0
H0 IQ
It0
and It0
are much less than unity values, high order terms in (B.48) can be ignored
and
IQ
H0 ≈ IQ
≈ IQ (B.49)
1+ Ito
H1 1
IQ
= It1
(ge − H1 ) + I12 [−H0 (ge − H1 )] + 1 2
3 H0
2It1
(ge − H1 ) + · · ·
h t1 i
H02
= 1
It1
(ge − H1 ) 1 + −H It1
0
+ 2I 2 + · · · (B.50)
t1
H0
1 −I
= It1
(ge − H1 ) e t1
³ ´
where It1 = It (ω) = It0 1 + j ωωπ . Solving for H1 from (B.50)
H0
−I
I Q ge e t1
H1 = H0 (B.51)
−I
It1 + IQ e t1
233
Note that
H0 IQ IQ
≈ = = g m Re (B.52)
It0 It0 φ t ge
H0
−I
For intermediate frequency of operation ω ¿ ωπ and also notice e t0 ≈ 1,
H0
H0 − H H
e
−I
t1 =e (
It0 1+j ωω
π ) ≈ e− It00 · ej It00 ωωπ ≈ ej ωωπ gm Re (B.53)
ω
For DC and low frequency operations ωπ
∼ 0,
gm
H1 ≈ (B.55)
1 + g m Re
This is the same as using Taylor series expansion for the circuit’s transconductance.
H2
IQ
= −H2
It2
+ 1
2 (ge
2It2
− H1 )2 + 1
2 H0 H2
It2
(B.56)
−H02
−H0
3 (ge
2It2
− H1 )2 + 3 H2
2It2
+ ···
³ ´
ω1 +ω2
where It2 = It (ω1 + ω2 ) = It0 1 + j ωπ , and
1©
(ge − H1 )2 = [ge − H1 (ω1 )]2 + [ge − H1 (ω2 )]2
ª
(B.57)
2
H0 1 H02
µ ¶ µ ¶
It2 1 2 H0
+1− + 2
− · · · H2 = (ge − H1 ) 1 − + ··· (B.58)
IQ It2 2 It2 2It2 It2
234
or equivalently
µ ¶
It2 H
− 0 1 H
− 0
+ e It2 H2 = (ge − H1 )2 e It2 (B.59)
IQ 2It2
and solving for H2
H0
−
IQ (ge − H1 )2 e It2
H2 = (B.60)
2It2 I + I e− IHt20
t2 Q
H3 −H3 1 1
IQ
= It3
+ 2 H0 H3
It3
− 2 (ge
It3
− H1 ) H2
(B.63)
+ 1
3 (ge
6It3
− H1 )3 − 1 2
3 H0 H3
2It3
+ ···
³ ´
ω1 +ω2 +ω3
where It3 = It (ω1 + ω2 + ω3 ) = It0 1 + j ωπ , and
1
(ge − H1 ) H2 = 3
{[ge − H1 (ω1 )] H2 (ω2 , ω3 )
1©
(ge − H1 )3 = [ge − H1 (ω1 )]3 + [ge − H1 (ω2 )]3 + [ge − H1 (ω3 )]3
ª
(B.65)
3
Rearranging the terms in (B.63) yields
H02
µ ¶
It3 H0 1 1
+1− + 2 − · · · H3 ≈ 2 (ge − H1 )3 − (ge − H1 ) H2 (B.66)
IQ It3 2It3 6It3 It3
235
or equivalently
µ ¶
It3 H
− 0 1 1
+ e It3 H3 ≈ 2
(ge − H1 )3 − (ge − H1 ) H2 (B.67)
IQ 6It3 It3
Solving for H3
IQ (ge − H1 )3 − 6It3 (ge − H1 ) H2
H3 ≈ 2 H (B.68)
6It3 I + I e It3
− 0
t3 Q
Higher order kernels can be calculated by repeating the above procedures. For
quick reference, the results in this example are summarized in Table XXIII.
H0 IQ IQ
H
− 0 ³ ´
IQ ge e It1 gm
H1 H
− 0 1+gm Re
1 − j ωωπ
It1 +IQ e It1
H0
− h i
IQ (ge −H1 )2 e It2 2
gm
H2 2It2 H
− 0
1
1 − j 2(ω1ω+ω
2IQ (1+gm Re )3 π
2)
VITA
Chunyu Xin was born in Tianjin, China. He received his B.S. and M.S. degrees from
Nankai University, Tianjin, China in 1997 and 1999 respectively. He is presently work-
ing toward his Ph.D. degree under the supervision of Dr. Edgar Sánchez-Sinencio in
the Analog and Mixed Signal Center, Texas A&M University. From September 2001
to August 2002 he worked for Motorola in Austin, TX for his internship. His re-
search interests are focused on integrated radio frequency circuits design for wireless
receivers. He held a Motorola fellowship in 2000. He has been an IEEE student mem-
ber since 2001. He can be reached through the Department of Electrical Engineering,