Documente Academic
Documente Profesional
Documente Cultură
532-539,
Aug 1992.
A newer (latest revision 2004) paper, A VLSI Device Model for Analog Circuits, is
also available. The newer paper substantially supersedes this 1992 publication.
CMOS Device Modeling for Subthreshold Circuits
Michael D. Godfrey*
Information Systems Laboratory
Electrical Engineering Department
Stanford University
Abstract
There is a need for a simple model of MOS device behavior that covers the sub-
threshold regime and the transition to above threshold. The need for such a model
will increase as increasing use is made of subthreshold operation for analog computa-
tion and for very low-power applications.
This paper explores such models and proposes a formulation which appears to
provide results as good as process variation permits, and which is well-suited to
ecient computation. The exponential dependence of source-drain current on gate
voltage in subthreshold implies that current values may be very sensitive to variation
in parameter values (particularly those that appear in exponents). This problem is
investigated, particularly with respect to threshold voltage, I
0
and .
* Research supported by a Grant from Apple Computer, Inc.
CMOS Device Modeling for Subthreshold Circuits
Michael D. Godfrey
1. Introduction
The increasing use of standard CMOS processes for construction of circuits which
operate in the subthreshold (weak inversion) regime leads to the need for a reliable
model of device behavior in this regime. The model should be reasonably accurate
with respect to device scale for present and future technologies and with respect
to a reasonable range of values of technology parameters. The form of the model
should lend itself to ecient computation and should t into suitable circuit models.
Ideally, the model should cover above threshold operation, but it is essential that
it yield accurate results in the transition between the two regimes. More precisely,
the results should be accurate for the circuit variables which signicantly aect the
behavior of typical constructions. For example, it should be possible to accurately
compute gain and conductance. Tsividis [6] has pointed out that obtaining such
accuracy is not as straightforward as might be thought, even when IV curves appear
to be highly accurate.
It would be particularly useful if the model had broad predictive power, i.e. it
would normally get fairly accurate results, given measured parameters from the pro-
cessing run (not necessarily the wafer or chip or device), without the need for ad-
justments of uncontrolled parameters.
There has been a substantial amount of modeling work which has addressed be-
havior in subthreshold (cf. [3, 4, 5, 6, 8, 9, 10 and 11]). This work provides a wealth
of background and useful analysis. For example, the problem of a discontinuity such
that the derivative of current as a function of voltage takes on the wrong sign in the
threshold transition was noted by Antognetti et.al. [10]. This problem was addressed
by Sheu et.al. in [11]. Sheu et.al. also provide a model which extends into subthresh-
old. However, their main interest is in improved model behavior above threshold.
The work reported by C.C. Enz, in his Ph.D. thesis [1], is focussed on subthreshold
behavior. However, his derivation depends on results from above threshold operation.
His model has the additional feature that it provides an ecient computational form
that covers subthreshold, inversion, and the transition between regimes. It does not
provide a model of operation in the linear regime.
This paper reports on experience in using the model reported by Enz, and work
that has been carried out to develop the model so that it yields accurate results
for, for example, present MOSIS-provided processes. The main eort in doing this
was to develop the model so that it is based on device parameters that are valid in
the subthreshold regime. Previous work has tended to use measurements done in
strong inversion and extrapolate them to subthreshold, and to force a correspondence
between approximations above and below threshold. A result of this approach has
been that the model works well for the square-law regime, but shows large errors
when used in subthreshold. The key parameters that require careful treatment are
2 CMOS Device Modeling for Subthreshold Circuits
the threshold voltage, V
t
, and the pre-exponential constant, I
0
. V
t
is famous for its
many conicting denitions [6]. I
0
is well-known as the place where all the left-
overs are dumped. In order to expect reasonable predictive power, both of these
parameters need precise, measurement-based, denitions.
2. Classical Derivations
The classical model equations are:
I
ds
=
(V
gs
V
t
)V
ds
V
2
ds
2
S
2
V
gs
V
t
2
(1 + V
ds
)
0 < V
ds
< V
gs
V
t
; linear
(1)
0 < V
gs
V
t
< V
ds
; strong inversion
where I
ds
is the drain-source current, S = channel width to length ratio (W/L),
= C
ox
where is the carrier mobility, C
ox
= /t
ox
where is the permitivity of
the gate insulator and t
ox
is the thickness of the gate insulator, V
gs
is the gate-source
voltage, V
ds
is the drain-source voltage, V
t
is the threshold voltage, and is the linear
channel length modulation (Early) factor.
The corresponding, simplied, result for the subthreshold saturation regime is:
I
ds
= I
0
kT
q
2
e
qVg
kT
e
qVs
kT
e
qV
d
kT
+ g
d
V
ds
(2)
where I
0
is the pre-exponential constant (to be discussed below), V
g
is the gate
voltage, V
s
is source voltage, V
d
is drain voltage all relative to substrate, is the
body eect ( =
s
V
g
, where
s
is the surface potential), k is Boltzmanns constant,
T is temperature, q is the charge on an electron, and g
d
=
I
V
d
is the coecient
for channel length modulation (Early) eect. Note that the eects of channel length
modulation appear to be dierent between above and below threshold. Therefore,
there is likely not a simple relationship between and g
d
.
The problem now is to try to match up the strong inversion (square law) expres-
sion and that for subthreshold. This looks easier than it is. However, note that if,
in the square law expression, (V
gs
V
t
) is divided by kT/2q (the justication for the
factor 2 will be explained below) then the constant terms:
I
0,sq
= S2
kT
q
2
square-law
(3)
and
I
0,sb
= I
0
kT
q
2
subthreshold
(4)
at least have the same dimensions of current. A preview of the problem of near-
threshold operation is that I
0
is not likely to be equal to S2 since its dependence
on the gate-channel capacitance may be dierent from that of .
CMOS Device Modeling for Subthreshold Circuits 3
Another problem is that the simplifying assumptions, typically linearization about
some operating point by rst-order Taylor Series approximation, in both above and
subthreshold may not be suciently accurate approximations to be useful in the
transition region. In addition, the denition and measurement of V
t
has added con-
siderable confusion and is critical to accurate results. It is important to note that
V
t
is not a physical constant, but rather it is a point value which is intended to dis-
tinguish between two regimes which have no discontinuous boundary. Unfortunately,
a number of theoretical and empirical denitions of V
t
are in use, and they produce
wildly diering results. The usual formal denition is that V
t
is the gate voltage such
that the current due to drift is equal to the current due to diusion. However, this
is not a state that is directly measurable and therefore various indirect methods are
used for measurement. Tsividis [6] makes this problem abundantly clear and shows
specic examples of the resulting errors. The consequences of this problem for the
Enz-Vittoz-Oguey model are discussed below.
3. The Problems of the Transition Region
The basic problem of the transition region is that there is no clear understanding
of the interaction of the potentials in the channel in the situation where part of
the current is due to drift and part due to diusion. The relationship between V
p
,
pincho voltage, and V
t
is unclear in this region. There is a tendency to use V
p
and
V
t
interchangeably in much of the literature. Specically, equations similar to those
given above, but as a function of V
p
instead of V
t
, are sometimes used. This is the
case in the Enz-Vittoz-Oguey derivation below.
One approximation for V
p
is V
p
= (V
g
V
t
). If the subthreshold equation is
written using V
p
in place of V
t
, and this approximation for V
p
is substituted then a
term of the form e
V
t
appears in the pre-exponential I
0
term. The dependence of the
pre-exponential term on V
t
can have a very large eect on device behavior and on any
understanding of the sources of variation in currents and in conductance. However,
the standard subthreshold model does not give rise to a V
t
term in the pre-exponential
constant.
The transition region is larger than one might guess. The strong inversion model
predicts zero current for any gate voltage below V
t
, with the current dropping to zero
as the square of the gate to threshold voltage dierence. The subthreshold model
predicts a log-linear current for any gate voltage. The slope of the IV curve will start
to diverge from log-linear well before the threshold voltage. Thus, there is a region of
width on the order of 0.5 volt in which neither model provides anywhere near correct
behavior. Since the error rst appears as a change in the slope of the IV curves, it
greatly aects all behavior that depends on the derivative of current with respect to
voltage.
4 CMOS Device Modeling for Subthreshold Circuits
4. The Enz/Vittoz Model
The model described by Enz is based on prior work by Vittoz and Oguey, and was
developed under Vittozs supervision. A key mathematical approximation used in
the model is due to Oguey (see below). Therefore, we will refer to the model as the
Enz/Vittoz/Oguey (EVO) Model. In the following Sections we develop the model
from device equations, following the derivations in Enz, but also pointing out some
additional consequences of the derivation and approximations.
In is essential to keep in mind that, for our purposes, the model should behave
well for devices in subthreshold. Thus, gate voltages are intended to be well below
threshold voltage and typical I
ds
values are around or below 10
9
A.
4.1 Weak Inversion
The equation for drain current in weak inversion used by Enz in [1] takes the form:
I
d
= K
W
SU
2
T
e
(VpVs)
U
T
e
(VpV
d
)
U
T
(5)
where:
variable denition dimensions
S = channel width/length ratio (W/L) dimensionless
U
T
=
kT
q
= V
k
= 0.025v at room temperature V
I
d
drain current A
= C
ox
50A/V
2
A/V
2
K
W
Adjustment factor dimensionless
n = 1/ 1.4 Charge eect due to substrate dimensionless
V
T0
Threshold voltage (ref. substrate) V
V
p
=
1
n
(V
g
V
T0
) Pincho voltage V
V
s
Source voltage (ref. substrate) V
V
d
Drain voltage(ref. substrate) V
4.2 Strong Inversion
In strong inversion current ow is governed by drift. In order to match the functional
form required for the approximation (explained below) Enz derives the following
square law equation:
I
d
=
nS
2
((V
p
V
s
)
2
(V
p
V
d
)
2
). (6)
However, he points out, but does not comment on, the fact that this equation is valid
only in the conduction regime dened by:
V
s
and V
d
< V
p
.
In order to make this equation more nearly consistent with standard derivations
V
p
V
s
and V
p
V
d
should be replaced by max(V
p
V
s
, 0) and max(V
p
V
d
, 0). This
CMOS Device Modeling for Subthreshold Circuits 5
change does not conict with the subsequent use of the equation in the EVO model,
and yields conventional square-law behavior in the above-threshold regime. However,
this is not a valid model in the region near subthreshold where gate voltages are
relatively low, and therefore either V
s
or V
d
will approach V
p
. With this modication,
equation 6 becomes:
I
d
=
nS
2
(max(V
p
V
s
, 0))
2
(max(V
p
V
d
, 0))
2
. (7)
This expression covers conduction and strong inversion with both forward and reverse
saturation. This leaves the problem of nding an approximation that connects the
regions of strong and weak inversion.
4.3 The Enz/Oguey Approximation
The key result in Enzs thesis is an approximation* which covers both weak and
strong inversion without discontinuity. The approximation is based on the function
F(x) such that:
F(x) =
(
x
2
)
2
x >> 0
e
x
x << 0.
(8)
The function that approximates this behavior is:
F(x) = ln
2
(1 + e
x/2
). (9)
It can be seen that this function behaves like equation (8) by considering the two
cases, x >> 0 and x << 0. For x >> 0,
F(x) ln
2
(e
x/2
) = (x/2)
2
. (10)
For the case x << 0, note that if << 0, e
max(V
p
V
s
, 0)
2U
T
max(V
p
V
d
, 0)
2U
T
2
(11)
* The references for the original suggestion of this approximation given in the thesis
are:
[1] H.J. Oguey and S. Cserveny, MOS Modelling at low current density, Summer
Course on Process and Device Modelling, ESAT Leuven-Heverlee, Belgium, June
1983.
[2] H. Oguey et S. Cserveny, Mod`ele du transistor MOS valable dans un grand
domaine de courants, Bull. SEV/VSE, Feb. 1982.
6 CMOS Device Modeling for Subthreshold Circuits
then from the weak and strong inversion equations (equations 5 and 11) we see that
I
d
F
V
p
V
s
U
T
V
p
V
d
U
T
(12)
where:
F(x) = ln
2
(1 + e
x
2
).
It is the assignment of K
W
= 2n that forces the pre-exponential constant in both
regimes to have the value derived from the strong inversion model. Figure 1 shows
the following three functions:
F
1
(V
g
) = ln
2
1 + e
(V
p
V
s
)/(2U
T
)
ln
2
1 + e
(V
p
V
d
)/(2U
T
)
F
2
(V
g
) =
max(V
p
V
s
, 0)
max(V
p
V
d
, 0)
2
F
3
(V
g
) = e
(V
p
V
s
)/U
T
e
(V
p
V
d
)/U
T
(13)
10
-11
10
-8
10
-5
10
-2
10
1
10
4
10
7
10
10
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
I
s
a
t
(
u
n
n
o
r
m
a
l
i
z
e
d
)
Vgs (v)
kappa = 0.676, Vt = 0.852.
F3
F2
F1
Figure 1: Behavior of Approximation
The solid curve, F
1
, shows the approximation, while the curve labeled F
2
shows the
behavior of the square law equation and F
3
shows the exponential. At this scale the
gap between F
2
and F
3
looks small, but as was discussed in Section 3 an accurate
model of device behavior in this transition region is essential to useful circuit modeling.
In order for equation 12 to be made an equality, the pre-exponential terms must
be equated. However, these equations require that 2 = / or that some additional
parameter be introduced into either the above or below threshold equations. Enz [1]
introduced the factor K
W
for this purpose, but did not provide a justication for
K
W
.
4.4 The Enz/Oguey Approximation in Computational Form
One of the very nice features of the Enz/Oguey approximation is its computational
simplicity. Using the approximation V
p
= (V
g
V
t
) and including a term for the
CMOS Device Modeling for Subthreshold Circuits 7
linear Early eect we can write the equation for F
1
(V
g
) from equations (13) as*:
K
1
= 2nSU
2
t
/
I
f
= ln
2
1 + exp
(V
g
V
t
) V
s
2U
t
I
r
= ln
2
1 + exp
(V
g
V
t
) V
d
2U
t
I
ds
= K
1
(I
f
I
r
)(1 + V
d
)
(14)
where is the slope term from the linearized Early eect in the square-law regime.
However, note that this form has carried over the constant term, 2nSU
2
t
/, from
the strong inversion regime. This may lead to large errors if the equations are used
for subthreshold. For V
g
< V
t
equations 14 take the form:
I
ds
= K
1
e
((VgV
t
)Vs)
2U
t
e
((VgV
t
)V
d
)
2U
t
= K
1
e
(VgV
t
)
2U
T
(e
Vs
2U
t
e
V
d
2U
t
)
(15)
The occurrence of V
t
in the exponent of the pre-exponential constant is the source of
the claim that lack of process-control of V
t
leads to instability of I
ds
for subthreshold
devices. For subthreshold analysis it is not appropriate to derive I
0
from the denition
of K
1
above. This point is explored in Section 7.1 below.
5. Experimental Results
With the model in this form, and available in MatLab, it is a simple matter to use it
to investigate reported or measured data from transistors operated in subthreshold.
Equations (14), written in MatLab, are shown in the Appendix.
5.1 Results reported by Enz
Enzs Thesis contains plots of experimental measurements with the model curves
tted to the data points. His procedure was to adjust the model parameter values
to get the best t. This shows that the model has a suitable functional form, and
sucient parameters, so that it can be made to t experimental data very closely.
It does not however indicate whether the model can be used to predict results given
measured values of the parameters.
5.2 Failure to Predict I
d
and I
sat
The attempt to apply the EVO model in the form presented by Enz does not yield
satisfactory results. We took the EVO model and applied it to data reported in [3]
with the following results.
* This is the form implemented by John Lazzaro in the analog simulator, anaLOG.
8 CMOS Device Modeling for Subthreshold Circuits
IV curves for subthreshold devices are shown in Figures 3.6 and 3.7 on pages 37
and 38 of [3]. The device and process parameters for the measured devices are given in
the captions and accompanying text. The following gures show the same plots (of I
d
vs. V
ds
and I
sat
vs. V
gs
) as in Figures 3.6 and 3.7. In Figure 2 we adjusted the process
parameters and an additional scale parameter to produce a close approximation to
the Figures in [3]. In order to get an approximate t it was necessary to use a scaling
factor (I
f
) of 80. Figure 3 shows the result of using the measured process parameter
values directly and setting the scaling factor to 1.0. Comparing Figure 3 with either
Figure 2 or the Figures from [3] shows a dierence in current values of a factor of
about 80.
Discrepancies of this magnitude can have harmful eects on analog designs.
0
0.5
1
1.5
2
2.5
x10
-9
0 1 2 3 4 5
I
d
(
A
)
Vds (v)
beta = 50, kappa = 0.7, Vt = 0.9. If = 80
Vgs = 0.60
Vgs = 0.575
Vgs = 0.55
Vgs = 0.525
10
-12
10
-11
10
-10
10
-9
10
-8
10
-7
10
-6
10
-5
10
-4
0.4 0.5 0.6 0.7 0.8 0.9 1
I
s
a
t
(
A
)
Vgs (v)
beta = 50, kappa = 0.7, Vt = 0.9. If = 80
Figure 2: Approximate Reproduction of Figures 3.6 and 3.7 of [3]
CMOS Device Modeling for Subthreshold Circuits 9
0
0.5
1
1.5
2
2.5
x10
-11
0 1 2 3 4 5
I
d
(
A
)
Vds (v)
beta = 50, kappa = 0.7, Vt = 0.9. If = 1
Vgs = 0.60
Vgs = 0.575
Vgs = 0.55
Vgs = 0.525
10
-14
10
-13
10
-12
10
-11
10
-10
10
-9
10
-8
10
-7
10
-6
0.4 0.5 0.6 0.7 0.8 0.9 1
I
s
a
t
(
A
)
Vgs (v)
beta = 50, kappa = 0.7, Vt = 0.9. Io = 1
Figure 3: Enz/Vittoz Model Using Device Parameters from [3]
6. Improvement to the Model
The mathematical attractions of the EVO model provide an incentive to nd the
source of the problems with its use for accurate description and prediction. Investi-
gation shows that a major source of problems is the denition and measurement of
V
th
, and the relation of V
th
to I
0
and .
6.1 Measurement of V
th
The standard method of measuring V
th
is:
10 CMOS Device Modeling for Subthreshold Circuits
1. Set V
ds
to some value which is much less than twice the Fermi potential, such as
0.05v.
2. Measure I
d
vs. V
g
for values of V
g
from 0 to about 2v.
3. Extrapolate the linear part of the resulting curve (above V
th
) back to the I
d
origin. This value of V
g
is the measured V
th
.
0
1
2
3
4
5
6
x10
-6
0 0.5 1 1.5 2
I
d
(
A
)
Vg (v)
beta = 50, kappa = 0.676, Vt = 0.9. Io = 1
Vds = 0.1 Vto = 0.9806
Figure 4: Measurement of V
th
from EVO Model
If this measurement procedure is applied to the EVO model (pretending the model is
a transistor) a result like that shown in Figure 4 is obtained. In the Figure, V
t
is the
threshold voltage used by the model and V
t0
is the threshold voltage value measured
from the model output. This shows immediately that the threshold voltage used by
the model and the measured value are substantially dierent. The magnitude of the
dierence depends on the value of V
th
and on the value of V
ds
used in the measure-
ment. However, this measurement procedure for V
th
measures device performance in
the linear regime. The EVO model is not valid in this regime.
7. Sensitivity of I
ds
to V
th
It is often stated (for example in [2]) that the source of variation in I
ds
is lack of
process control over V
th
. As has been shown above, part of the problem stems from
inconsistent denitions of V
th
. However, it is also clear that variation in V
th
can be a
source of unexpected variation in I
ds
. Because of the exponential dependence of I
ds
on V
th
. Very small variation in V
th
produces a large variation in I
ds
. Figure 5 shows
this eect.
CMOS Device Modeling for Subthreshold Circuits 11
10
-11
10
-10
10
-9
10
-8
10
-7
0.8 0.82 0.84 0.86 0.88 0.9 0.92
I
d
(
A
)
Vth (v)
beta = 50, kappa = 0.676, Vg = 0.6. Io = 1
Figure 5: Dependence of I
ds
on V
th
.
A one percent change in V
th
yields a nearly 20 percent change in I
ds
. It is unrealistic
to expect control of V
th
below about a percent, but this it what would be required
if values of I
ds
with accuracy better than 20% were to be obtained. Pavasovic,
Andreou and Westgate [7] and Pavasovic [8] have investigated variation in drain
current in subthreshold within individual die. The results that they report are from
modern standard fabrication processes, such as those provided through MOSIS. They
constructed arrays of transistors so that they could measure variation in behavior as
a function of position on the die. An important conclusion in their work is that
the spatial variation (which also, curiously, has a distinct spatial periodicity) was, in
terms of measured current, about 2% for 16 by 16m transistors and about 7.2% for
4 by 4m devices. This variation, if it is really due to variation in V
th
, represents
less than 0.5% variation in V
th
. Some fabricators have recently claimed to be able to
hold V
th
to within 1%. These results and claims suggest that the minimum device
to device variation in I
ds
that is likely to be achievable will be above 20%. This,
however, is a variation which is comparable to other device variation. It is much less
than seems to have been suggested in previous literature.
7.1 Measurement of I
0
, , and V
th
The most direct way to achieve good results in the subthreshold regime, using this
model, is to measure I
0
for the process being used. It may, in addition, prove to
be possible to develop an accurate model of I
0
as a function of physical and process
parameters. We have made a number of measurements of I
0
from the MOSIS-provided
2 p-well process. At present we only have data from one fabrication run. (We have
a more complete set of test structures in fabrication at present. We therefore hope to
be able to extend these results in the near future.)
The fabrication through MOSIS used a standard CMOS p-well process. The
12 CMOS Device Modeling for Subthreshold Circuits
measured gate oxide thickness was 408