Vous êtes sur la page 1sur 4

RF Distortion Analysis with Compact MOSFET Models

Peter Bendix, P. Rakers


+
, P. Wagh
+
, L. Lemaitre
+,
W. Grabinski
+
, C. C. McAndrew
+
,
X. Gu
*
and G. Gildenblat
*
LSI Logic Inc.,
+
Motorola Inc.,
*
The Pennsylvania State University
Abstract
This paper examines the relation between the structure
of a compact MOSFET model and its ability to model
harmonic distortion. It is found that non-singular behavior
at zero drain bias is essential for qualitatively correct
simulations of the third harmonic power dependence.
Specifically, nonlinear distortion analysis requires that the
Gummel symmetry condition be satisfied by the compact
model. A simple procedure to enforce the Gummel
symmetry without increasing the complexity of the model
is incorporated in an advanced surface-potential-based
MOSFET model to enable correct harmonic distortion
modeling.
Keywords: IM3, distortion, FFT, harmonic analysis, RF
CMOS, MOSFET modeling.
Introduction
The design of RF CMOS circuits for mixed signal
applications is becoming increasingly important. Distortion
is a key measure of performance for RF circuits, and for
mixer design third order intermodulation IM3 ([1], pp. 296-
297) is commonly used as a measure to evaluate distortion.
Unfortunately, simulation of IM3 can be qualitatively
wrong, and therefore cause significant errors for RF design,
because of fundamental problems with models used for
simulation. This has been shown for MESFETs in [2]. In
this paper we show that a similar problem exists in
MOSFET models; we then describe the reasons for the
problem, and an efficient approach to fix the problem.
The IM3 modeling problem is manifest in the simple
situation where a sinusoidal voltage (of zero DC bias and
swept amplitude A ) is used to drive the drain of a
transistor with the gate biased on, and a fast Fourier
transform (FFT) of the drain current is done to get the
amplitude of the various harmonics of the drain current.
The MOSFET channel in this case acts as a weakly nonlin-
ear resistor, which produces the harmonic distortion being
analyzed. Both theoretical considerations [1], [2] and
analysis of experimental data indicate that the second
harmonic should be proportional to the square of the input
signal amplitude
2
A , and the third harmonic should be
proportional to the cube of the signal amplitude
3
A . As a
result of unphysically asymmetric MOSFET model
behavior around 0 =
ds
V , all older and many modern
compact MOSFET models (including source-referenced
threshold-voltage based models like the standard Level 1,
Level 2, Level 3, BSIM, BSIM2, BSIM3, and BSIM4
models) used in commercial circuit simulators fail the
Gummel symmetry test [3] and therefore cannot be used for
IM3 and related harmonic distortion tests.
The purpose of this paper is to conclusively demonstrate
the IM3 modeling problem. For this purpose we work with
the latest generation compact MOSFET models SP [3] and
BSIM3 and BSIM4 [4].
Transient Simulation
A transient analysis for the circuit of Fig. 1 was done
using the Spectre

circuit simulator with the BSIM4 and SP


MOSFET models (BSIM4 is a threshold voltage-based
model while SP is a true surface-potential-based model).
The results are not dependent explicitly on the choice of
model parameters. The transient analyses were done for 8
values of driving sinusoid amplitude (160, 200, 240, 280,
320, 360, 400, and 440mV). The current was then analyzed
using an FFT code based on the four1 routine of [5], and
the magnitude of the fundamental, second, and third
harmonics were plotted as a function of ) log(A . In running
these simulations, tolerances and biasing were chosen
carefully to avoid anomalies in the results. In particular,
the amplitude of the sinusoid was not too large, there was a
sufficiently small timestep chosen to provide enough data
points for the specified frequency to get accurate FFTs,
and the simulation accuracy tolerances were chosen small
enough so as not to corrupt the results.
Fig. 1 Circuit for distortion analysis, proposed in [2].
Fig. 2 shows the simulation results from BSIM4. As
expected, on a dB scale the slope of the fundamental is one
and the slope of the second harmonic is two. However, the
slope of the third harmonic is two (just as for the second
harmonic) instead of the correct value of three. The reason
for this problem with BSIM4, and many other MOSFET
models, is a lack of symmetry in the model. This leads to
different left and right hand second order derivatives at
0 =
ds
V . Consequently, the second and higher order
derivatives do not exist at 0 =
ds
V .
As Fig. 3 shows, the SP model gives results identical to
what theory predicts. The slope of the third harmonic is
three, as it should be. This means SP gives correct results
for IM3 RF simulations.
+

+
~
) cos( t A V
RF
e =
g
V
b
V
d
I
2-2-1 9
IEEE 2004 CUSTOM INTEGRATED CIRCUITS CONFERENCE
0-7803-8495-4/04/$20.00 2004 IEEE.
Fig. 2 BSIM4 transient simulation distortion results for
the single transistor circuit of Fig. 1.
Fig. 3 SP transient simulation distortion results for the
single transistor circuit of Fig. 1.
Fig. 4 Harmonic balance simulation of circuit of Fig. 1.
Harmonic Balance Simulations
A single tone harmonic balance simulation of the circuit
of Fig. 1 was run in the Mica

simulator, for BSIM3,


BSIM4, and SP models. Different model parameters were
used than for the transient simulations of Fig. 2 and Fig. 3,
but the parameters used for all the harmonic balance
simulations were extracted for the same 0.18m CMOS
technology. Fig. 4 shows the results of the harmonic
balance simulations. The third harmonic component again
has an incorrect slope of two for BSIM3 and BSIM4, but
SP has the correct slope of three.
CMOS RF Mixer Simulation
A three tone intermodulation distortion harmonic
balance simulation was also run on a simple CMOS RF
mixer, to investigate the effect of the singularity at 0 =
ds
V
on a circuit, and not just a single device used as a voltage
controlled resistor. Fig. 5 shows the topology of the mixer.
Fig. 5 Simple CMOS RF mixer. Besides the RF signals
there are DC offsets to bias the MOSFETs on and avoid
forward biasing of source-bulk and drain-bulk junctions.
The local oscillator frequency
LO
f was 1GHz, and two
RF signals at =
1 RF
f 910 and =
2 RF
f 920 MHz were fed in
to the mixer and the power level ramped. The first order
intermodulation (IM1) frequencies are
LO RF
f f
1
and
LO RF
f f
2
. The second order intermodulation (IM2)
frequencies are
2 1 RF RF
f f and
2 1 RF RF
f f + . The IM3
frequencies are
LO RF RF
f f f
2 1
2 and
LO RF RF
f f f
1 2
2 (the other IM3 frequencies
LO RF RF
f f f +
2 1
2 and
LO RF RF
f f f +
2 1
2 are
considered out of band and are not analyzed).
The power levels of the IM1 at 90MHz and the second
and third intermodulation products at 10MHz and 100MHz
were recorded. As Fig. 6 shows, the IM3 slope for SP is
very nearly three (it was calculated from a linear fit to the
lowest power data points as 2.98), whereas for BSIM3 and
BSIM4 it is again close to two. As far as we are aware, this
is the first presentation of incorrect simulation of IM3 at a
circuit level, as opposed to the individual device level,
caused by the singularity at 0 =
ds
V .
16 15 14 13 12 11 10 9 8 7 6
80
70
60
50
40
30
20
10
V
RF
(dB)
I
d

(
d
B
)
Fundamental slope=1
2
nd
Harm
onic slope=2
3
rd
Harm
onic slope=2 (should be 3)
16 15 14 13 12 11 10 9 8 7 6
100
90
80
70
60
50
40
30
20
10
V
RF
(dB)
I
d

(
d
B
)
Fundamental slope=1
2
nd
Harmonic slope=2
3
rd
H
arm
onic slope=3
40 38 36 34 32 30 28 26 24 22 20
160
140
120
100
80
60
40
V
rf
(dB)
I
d

(
d
B
)
Fundamental BSIM3, BSIM4 and SP
2
nd
H
arm
onic BSIM
3, BSIM
4 and SP
3
rd
H
a
rm
o
n
ic
S
P
3
rd
H
arm
onic BSIM
3 and BSIM
4
BSIM3
BSIM4
SP
~
+
LO
V
~

LO
V
+
RF
V

RF
V
out
V
DC g
V
,
DC g
V
,
DC b
V
,
2-2-2 10
Fig. 6 Intermodulation products for the simple RF
CMOS mixer circuit of Fig. 5.
Gummel Symmetry Test
To corroborate the reason for the success or failure of
modeling IM3, the models were subjected to the Gummel
symmetry test (as described in [6]). For this test a MOSFET
is biased symmetrically with respect to source and drain,
with
x d
V V V + =
0
and
x s
V V V =
0
, see Fig. 7. This
makes the drain current
d
I an odd function of
x
V .
Consequently, all odd order derivatives of ) (
x d
V I with
respect to
x
V should be continuous at 0 =
x
V , and all even
order derivatives, including
2 2
x d
V I c c , should exist and
should be equal to zero at 0 =
x
V .
Fig. 7 Circuit for Gummel symmetry test.
Fig. 8 shows the results of this simulation for the SP,
BSIM3, and BSIM4 models. SP has the desired behavior,
with no kinks or discontinuities in
2 2
x d
V I c c at 0 =
ds
V .
BSIM3 and BSIM4 however both exhibit a singularity at
0 =
x
V . More specifically, the left and right hand
derivatives of
x d
V I c c with respect to
x
V are different at
0 =
ds
V , so
2 2
x d
V I c c does not exist at 0 =
ds
V .
Consequently BSIM3 and BSIM4, and many other
MOSFET compact models, fail to simulate third order
harmonic distortion properly.
Fig. 8 Gummel symmetry test results for the second
derivative.
Ensuring Non-Singular Behavior at 0 =
ds
V
The violation of the Gummel symmetry test in BSIM
and other MOSFET models has been traced to the use of a
source-referenced threshold voltage and to the singularity
of the velocity-field relation [6]. In body-referenced models
like SP only the second issue needs to be examined.
Over the years several physically motivated and
empirical equations have been suggested for the drift
velocity
d
v of mobile carriers in the channel of MOS
transistors. The most popular form, routinely encountered
in compact models,
(1)
n
n
c
d
E
E
E
v
1
0
1
|
|
|
.
|

\
|
|
|
.
|

\
|
+
=

(where
0
is the low-field effective mobility, E is the
lateral component of the electric field, and
c
E is the
critical field), has two drawbacks. First, the ) (E v
d
dependence for 1 = n is too weak for n-channel transistors.
Further, for 1 = n
(2)
+
= =
|
|
.
|

\
|
= =
|
|
.
|

\
|
0
2
2
0
0
2
2
2
E
d
c
E
d
dE
v d
E
dE
v d
so the 2
nd
and higher order derivatives at 0 =
ds
V do not
exist. This is the root cause of the singularity in the
) (
ds d
V I characteristics at 0 =
ds
V , i.e. why models that
use this form of velocity saturation model fail the Gummel
symmetry test [6].
As noted in [6], the expression (1) with 2 = n is more
accurate for n-channel transistors and does not have the
singularity at 0 = E . However, setting 2 = n results in a
complicated expression for the drain current. SP removes
50 45 40 35 30 25 20 15 10
200
180
160
140
120
100
80
60
40
20
V
rf
(dBm)
V
o
u
t

(
d
B
m
)
Fundamental BSIM3, BSIM4 and SP
2
nd
H
a
rm
o
n
ic B
S
IM
3
, B
S
IM
4
a
n
d
S
P
3
rd
H
a
rm
o
n
ic
S
P
3
rd
H
a
rm
o
n
ic B
S
IM
3
a
n
d
B
S
IM
4
BSIM3
BSIM4
SP
+
+

g
V
b
V
d
I
+

+

x
V V +
0 x
V V
0
10 8 6 4 2 0 2 4 6 8 10
0.4
0.3
0.2
0.1
0
0.1
0.2
0.3
0.4
V
x
(mV)

2
I
d
/

V
x 2

(
S
/
V
)
BSIM3
BSIM4
SP
2-2-3 11
this singularity via a semi-empirical correction factor
0
o .
In this approach [7], [8]
(3)
|
|
.
|

\
|
+
=
c
d
E
E
E
v
0
0
1 o

where
(4)
L E g V
V
c h ds
ds
+
=
0
o ,
L is the effective channel length, and
h
g is an empirically
adjustable parameter.
Expression (3) has several advantages for the purpose of
compact modeling. First, since
0
o does not depend on
position along the channel, it is as easy to use as (1).
Further, the
ds
V dependence of
0
o sharpens the ) (E v
d
dependence, which was the original purpose of introducing
0
o in [7], [8]. Finally, when used in an otherwise
symmetric compact model it leads to [3]
(5) 0
0
2
2
=
c
c
=
x
V
x
d
V
I
.
Hence (3) will not lead to a violation of the Gummel
symmetry test.
Conclusions
We have presented transient and harmonic balance
simulations that show that compact MOSFET models with
a singularity at 0 =
ds
V , as evidenced by failing the
Gummel symmetry test, are not able to model distortion
properly. Even for a symmetric, bulk-referenced model
formulation, the use of a velocity saturation model with a
singularity at 0 = E will still cause problems for distortion
modeling, and this can be overcome, as in SP, by using (3)
in place of (1).
As far as we are aware, this is first time that distortion
analysis results that show the problem caused by a
singularity at 0 =
ds
V have been presented for MOSFETs,
as opposed to MESFETs in [2], and the first time that such
results have been presented for a mixer circuit, as compared
to a single transistor simulation.
References
[1] T. H. Lee, The Design of CMOS Radio-Frequency
Integrated Circuits, Cambridge University Press, 1998.
[2] N. Scheinberg and A. Pinkhasov, A Computer
Simulation Model for Simulating Distortion in FET
Resistors, IEEE Trans. CAD, vol. 19, no. 9, pp. 981-
989, Sep. 2000.
[3] G. Gildenblat, T.-L. Chen, X. Gu, H. Wang, and X.
Cai, SP: An advanced Surface-Potential Based
Compact MOSFET Model, Proc. IEEE CICC, pp.
233-240, 2003.
[4] W. Liu, MOSFET Modeling for SPICE Simulation
Including BSIM3v3 and BSIM4, John Wiley and Sons,
2001.
[5] W. H. Press, B. P. Flannery, S. A. Eukolsky, and W. T.
Vettering, Numerical Recipes in C, Cambridge
University Press, 1988.
[6] K. Joardar, K. K. Gullapalli, C. C. McAndrew, M. E.
Burnham, and A. Wild, An Improved MOSFET
Model for Circuit Simulation, IEEE Trans. Electron
Devices, vol. 45, no. 1, pp. 134-148, Jan. 1998.
[7] T. Grotjohn and B. Hoefflinger, A Parametric Short-
Channel Transistor Model for Subthreshold and Strong
Inversion Current, IEEE J. Solid-State Circuits, vol. 9,
no. 1, pp. 100-112, Feb. 1984.
[8] N. D. Arora, R. Rios, C.-L. Huang, and K. Raol,
PCIM: A Physically Based Continuous Short-Channel
IGFET Model for Circuit Simulation, IEEE Trans.
Electron Devices, vol. 41, no. 6, pp. 988-997, Jun.
1994.
2-2-4 12

Vous aimerez peut-être aussi