Handout #20: EE214 Fall 2002

Voltage References and Biasing 1993 Thomas H. Lee; rev. November 27, 2002; All rights reserved Page 1 of 15
Voltage References and Biasing
1.0 Introduction
In this set of notes, we take up the study of an important topic: How to generate voltages
and currents that are relatively independent of supply voltage and/or temperature. Because
CMOS offers relatively limited options for realizing bias circuits, well see that some of
the most useful biasing idioms are actually those based on bipolar circuits. A parasitic
bipolar device exists in every CMOS technology, and may be used in a bandgap voltage
reference, for example. Even though the characteristics of parasitic transistors are far from
ideal, the performance of bias circuits made with such devices is frequently vastly superior
to that of pure CMOS bias circuits.
In what follows, it is worthwhile to keep in mind that any voltage we produce must depend
on some collection of parameters that ultimately have the dimensions of a voltage (such as
kT/q, for example). Similarly, any current we produce must depend on parameters that
ultimately have the dimensions of current (such as V/R). Although seemingly obvious and
trivial statements, well see that they are extremely useful guides for the design of stable
references.
2.0 Review of Behavior of Diodes
While the voltage across a forward-biased diode is relatively insensitive to current
because of the logarithmic dependence of diode current on diode voltage, its variation
with temperature is significant. To understand the precise nature of the temperature depen-
dence, recall that the diode voltage may be expressed as:
(1)
where V
T
is the thermal voltage, kT/q, and n, the ideality factor, is typically between 1 and
1.5 in diodes. A transistors V
BE
conforms more closely to the ideal diode law than do
ordinary diodes, so we will assign n a unity value in all that follows.
It is frequently inferred incorrectly from Eqn.1 that V
D
has a positive temperature coeffi-
cient (TC) because of its proportionality to V
T
. The fly in the ointment is that I
D
itself has
an exponential temperature dependence, and this alters the situation considerably. To clar-
ify matters, consider the following quasi-empirical expression for I
S
:
(2)
V
D
nV
T
l n
I
D
I
S
-----


=
I
S
I
0
exp
V
G0
V
T
---------


=
Handout #20: EE214 Fall 2002
Voltage References and Biasing 1993 Thomas H. Lee; rev. November 27, 2002; All rights reserved Page 2 of 15
where I
O
is some process- and geometry-dependent current
1
(I
0
is typically around 20
orders of magnitude larger than I
S
at room temperature, so I
0
is much larger than typical
values of I
D
), and V
G0
is the bandgap voltage (about 1.2 volts) extrapolated to absolute
zero.
Using this detailed expression for I
S
, we can expand the equation for V
D
as follows:
2
(3)
Thus, we see that the junction voltage decreases (linearly if I
0
were constant) from a value
of V
G0
, as seen in the following plot of V
D
vs. temperature at constant diode current:
FIGURE 1. Approximate behavior of V
D
vs. temperature
Note that this equation tells us that V
D
always equals V
G0
at absolute zero.
3
Furthermore
its easy to see that, if I
0
were constant, the temperature coefficient at any temperature is
simply
. (4)
1. It also depends weakly on temperature, but well defer a detailed discussion about the behavior of I
0
until
the section on bandgap voltage references.
2. The minus sign is not an error. Just remember that the argument of the log here is typically much larger
than unity.
3. Again, this value is an extrapolated one. It must be stressed that the behavior of real junctions at both
extremes of temperature will differ from that shown; the equations presented lose validity at extremely cold
temperatures (say, <100K) because of carrier freeze-out (i.e., failure of dopants to ionize) and because of
bandgap variation with temperature, and at high temperatures (>450-500K) because the silicon goes intrin-
sic.
V
D
V
GO
V
T
l n
I
0
I
D
-----


=
T (kelvins)
V
D
600K, typ.
higher I
D
lower I
D
V
G0
~ 1.2V
about 2mV/K
T d
dV
D
V
G0
V
D

T
---------------------- =
Handout #20: EE214 Fall 2002
Voltage References and Biasing 1993 Thomas H. Lee; rev. November 27, 2002; All rights reserved Page 3 of 15
With the assumption of constant I
0
, the temperature coefficient is independent of tempera-
ture and equal to about 2mV/K. This linearly decreasing behavior is known as CTAT, for
complementary to absolute temperature. Note that the voltage does depend (logarithmi-
cally) on diode current, so the temperature coefficient also depends somewhat on the
diode current, with lower currents associated with higher temperature coefficients.
Although a V
D
-based reference can provide an output that depends very little on supply
voltage, the CTAT behavior may or may not be acceptable, depending on the application.
However, we shall see that the CTAT behavior of a V
D
is particularly valuable for use in a
class of references based on the bandgap voltage V
G0
. Well take up the detailed study of
bandgap references in Section5.0.
3.0 Diodes and Bipolar Transistors in CMOS Technology
The most flexible option for realizing diodes and bipolar transistors in standard CMOS
technology derives from the parasitic substrate pnp transistor available in n-well pro-
cesses. The p+ source/drain diffusions serve as the emitter, the n-well as the base, and the
substrate as the collector:
FIGURE 2. Parasitic substrate PNP in n-well CMOS (not drawn to scale)
In applications where it is important to reduce series base resistance, it is advisable to sur-
round completely the emitter with n+ diffusions placed as close to the emitter as the
design rules allow, as suggested by Fig.2.
Just as its counterpart in inexpensive bipolar processes, the substrate pnp in CMOS tech-
nology can only be used in circuits that allow the collector to be at substrate potential. For-
tunately, there are numerous circuits that satisfy this condition. For example, a simple
voltage reference can be constructed with this device connected as a grounded diode, in
which the emitter is the anode, and the cathode is the base and collector (substrate) tied
together.
p-substrate
n-well
p+ n+ n+
emitter base collector
Handout #20: EE214 Fall 2002
Voltage References and Biasing 1993 Thomas H. Lee; rev. November 27, 2002; All rights reserved Page 4 of 15
4.0 Supply-Independent Bias Circuits
To minimize sensitivity to power supply variations, it is desirable to derive the bias cur-
rents for reference voltages from the reference voltages themselves, rather than directly
from the power supply. Although it may seem a violation of some fundamental law (the
no free lunch principle), it is possible to arrange for this condition. To illustrate how one
may accomplish this feat, consider the following circuit:
FIGURE 3. Self-biased reference
As you can see, the current through the diode depends on the diode voltage itself, rather
than on the supply voltage. This technique thus provides excellent supply-voltage inde-
pendence.
An important practical note is that a start-up network is always necessary in self-biased
circuits because there are two states, one which is stable in the conventional sense, and
another in which all currents are zero.
4
The start-up network guarantees that the circuit
gets out of the undesired metastable state.
Most practical implementations of the self-biased circuit dispense with the op-amp:
FIGURE 4. Self-biased reference
4. Even though the all-zero state is metastable, practical circuits are found in this state a maddeningly large
portion of the time. A start-up network is therefore mandatory for reliable operation.
R
R
1
R
2
I
D
V
D
R
2
R
1
------


R
------------------- =
I
D
V
D
+
-
R
V
DD
I
D
V
D
R
------- =
Start-up
network
V
D
+
-
Handout #20: EE214 Fall 2002
Voltage References and Biasing 1993 Thomas H. Lee; rev. November 27, 2002; All rights reserved Page 5 of 15
The PMOS mirror
5
enforces equality of the NMOS drain currents, and hence that of the
NMOS V
GS
s. Thus, the diode voltage appears across R; the corresponding current is the
same in both halves of the mirror and is therefore the bias current of the diode itself. Thus,
as in the op-amp version of this circuit, the diode provides its own bias current.
The self-biased circuit of Fig.4 is quite versatile. It should be clear that the diode may be
replaced by a variety of elements. For example, a diode-connected MOSFET would pro-
duce a bias current of V
GS
/R, or a zener diode (if available) could be used instead. As well
see in the next section, the self-biased circuit is particularly useful in realizing bandgap
voltage references in CMOS technology.
5.0 The Bandgap Voltage Reference
Since IC technology directly offers no reference voltages that are inherently constant, the
only practical option is two combine two voltages with precisely complementary tempera-
ture behavior. Thus, the general recipe for making temperature-independent references is
to add a voltage that goes up with temperature to one that goes down with temperature. If
the two slopes are equal in magnitude but opposite in sign, the sum will be independent of
temperature.
Without question, the most elegant realization of this idea is the bandgap voltage refer-
ence, for it produces an output voltage that is traceable to fundamental constants, and
therefore is relatively insensitive to process, temperature and supply variations.
The first widely used bandgap voltage reference was designed by Bob Widlar in the
1970s, in the hugely popular and revolutionary LM309 5-volt regulator IC from National
Semiconductor. It was the first reference whose initial accuracy was good enough to elim-
inate the requirement for adjustment by the end user. Thus, only three terminals were
needed (allowing use of inexpensive transistor packages), making this part as easy to use
as one could hope.
To understand quantitatively how bandgap references work, we need to re-examine the
detailed behavior of junction voltage with temperature. Since transistor junctions exhibit
more nearly ideal characteristics than ordinary diodes, we will assume bandgap imple-
mentations that use transistors. A plot of V
BE
vs. temperature is repeated here for conve-
nience:
6
5. Better mirrors would almost always be used in practice; simple ones are shown to reduce schematic clut-
ter.
6. Again, keep in mind that this plot of V
BE
is a slight fiction because we have neglected the small curvature
caused by the weak temperature dependence of I
0
. The correction is second-order, and we will take care of
this little detail shortly.
Handout #20: EE214 Fall 2002
Voltage References and Biasing 1993 Thomas H. Lee; rev. November 27, 2002; All rights reserved Page 6 of 15
FIGURE 5. V
BE
vs. temperature
Recall that V
BE
is nearly perfectly CTAT (i.e., it goes down almost linearly with tempera-
ture). Now suppose we add to this CTAT V
BE
a voltage that is perfectly proportional to
absolute temperature (PTAT). If we choose the slope of the PTAT term equal in magni-
tude to that of the CTAT term, the sum will be independent of temperature:
FIGURE 6. Illustration of bandgap reference principle
We see that something funny happens above about 600K, but the fact that the principle
fails at temperatures high enough to melt lead is rarely a practical concern.
Note that the addition of a PTAT and CTAT voltage in the proper ratio yields an output
equal to the bandgap voltage (extrapolated to 0K), independent of temperature. Stated
another way, if we adjust the PTAT component to make the output voltage equal to V
G0
at
any temperature, the output voltage will equal V
G0
at all temperatures, at least in this
slightly simplified picture.
T (kelvins)
V
BE
600K, typ.
V
G0
~ 1.2
almost perfectly CTAT
(but slope is current dependent)
T (kelvins)
600K, typ.
V
G0
~ 1.2
V
BE
(CTAT)
PTAT
V
CTAT + PTAT term = constant ~ V
G0
Handout #20: EE214 Fall 2002
Voltage References and Biasing 1993 Thomas H. Lee; rev. November 27, 2002; All rights reserved Page 7 of 15
At this point, its natural to consider how one obtains a PTAT voltage, since this whole
concept relies on having one around. Lets start with the familiar equation for V
BE
:
(5)
Using this expression, we can readily compute the difference in two base-emitter voltages
for identical transistors operating at two different values of collector current (or, more
generally, for transistors made in the same process, operating at two different values of
collector current density):
(6)
The misleading I
S
term drops out, so we can conclude confidently that V
BE
truly is PTAT
if the collector current densities are in a fixed ratio. Thus, while each V
BE
is nearly CTAT,
the difference between two V
BE
s is perfectly PTAT.
5.1 Classic Bandgap Reference
Now that weve got all the ingredients, all that remains is to sum the CTAT V
BE
term with
the right amount of PTAT V
BE
. While one could imagine a large number of methods for
doing so, the Brokaw cell is a particularly elegant (and accurate) implementation of the
bandgap reference. The classic bipolar implementation is shown in the following figure
(again, basic mirrors are shown for simplicitys sake); well modify this circuit shortly for
implementation in CMOS technology:
FIGURE 7. Classic Brokaw bandgap reference circuit
As we shall see, the output voltage is the sum of a PTAT voltage and a V
BE
. Here, Q
1
and
Q
2
operate at a fixed current density ratio of m (typically 8, because it lays out nicely) set,
for example, by ratioing the effective emitter areas. Now, by KVL, the voltage across R
2
is
V
BE
V
T
ln
I
C
I
S
-----


=
V
BE
V
BE2
V
BE1
= V
T
= ln
J
C2
J
C1
--------


R
2
V
CC
R
1
1:m
Start-up
network
V
OUT
Q
1
Q
2
V
BE
+

= V
BE
/R2
Handout #20: EE214 Fall 2002
Voltage References and Biasing 1993 Thomas H. Lee; rev. November 27, 2002; All rights reserved Page 8 of 15
the difference in V
BE
s of Q
1
and Q
2
, and is therefore PTAT and equal to V
T
ln m. Assum-
ing that R
2
s TC is negligibly small, the current through it will also be PTAT. Further-
more, the current through R
1
is simply twice that through R
2
since the two collector
currents are equal.
7
Therefore, the voltage drop across the entire resistor string is purely
PTAT. Finally, the output voltage is just this PTAT voltage plus the V
BE
of Q
2
, as adver-
tised. With proper choice of R
1
and R
2
, the output voltage will have zero TC. As a free
bonus, a PTAT voltage is available at the emitters of Q
1
and Q
2
, providing thermometer
outputs.
Design Example:
To carry out a detailed design, we need some process-specific device data. As a particular
example, suppose we go to the lab and find that V
BE
= 0.65V @ 300K and 100A for a
transistor of Q
2
s size. Furthermore, let m = 8. This choice of m
8
sets V
BE
=53.8mV (a
number comfortably larger than any offsets that we expect) @ 300K. Since we definitely
know the value of V
BE
at 100A, a prudent choice for the collector currents would be this
value of 100A, and this choice then fixes the value of R
2
= V
BE
/100A = 538. Now,
since we want the output voltage to be 1.2 volts, the drop across R
1
must be 1.2 V
BE

V
BE
= 0.496V. Finally noting that the current through R
1
is twice that through R
2
, we
conclude that we should choose R
1
= 0.496V/200A = 2.48k, completing the design.
You may have noticed that the collector currents in the Brokaw cell are not constant (in
fact, they are PTAT if we assume that the resistors have zero TC). To see why this does
not invalidate all weve done so far (in fact, it is beneficial), it is now time to take care of a
few details, namely, those involving the temperature dependency of I
0
.
A quasi-empirical expression for I
0
is
(7)
where A
e
is the emitter area, B is a process-dependent constant, T is the absolute tempera-
ture and r is a process-dependent quantity well call the curvature coefficient. For the rela-
tively deep, diffused emitters of older bipolar processes, r typically has a value between 2
and 3, while for the shallow, implanted (and very heavily doped
9
) diffusions that are more
common in modern CMOS and high-speed bipolar processes, r can range from 4 to 6.
Your mileage may vary, so dont undertake a detailed design without first characterizing
the junctions in your process.
With this equation for I
0
, we can express V
BE
as follows:
7. We are neglecting errors due to device mismatch, nonzero base currents and finite Early voltage.
8. It is important that Q2 be laid out as eight instances of Q1 to guarantee that Q2 behaves as eight paral-
leled devices of Q1s size. If Q1 is placed at the center of a common-centroid arrangement, errors due to pro-
cess variation will be minimized.
9. Bandgap narrowing and nonlinearity in the heavily doped emitters are probably responsible for the high
values of r.
I
0
A
e
BT
r
=
Handout #20: EE214 Fall 2002
Voltage References and Biasing 1993 Thomas H. Lee; rev. November 27, 2002; All rights reserved Page 9 of 15
(8)
Plotting as before, we can see why it is reasonable to call the parameter r the curvature
coefficient (aside from the euphonious alliteration):
FIGURE 8. V
BE
vs. temperature
Because the argument of the log is not quite independent of T, the temperature coefficient
of V
BE
is not quite constant, leading to a small departure from CTAT behavior for V
BE
.
Additionally, weve seen at least one implementation of a bandgap reference in which the
collector current is also not constant. So, lets compute the actual TC that results if, in
addition to the temperature dependence of I
0
, we also consider a collector current that var-
ies as the nth power of T:
(9)
so that
(10)
which we may re-write in a somewhat lower entropy form, as in Section2.0:
(11)
V
BE
V
G0
V
T
ln
A
e
BT
r
I
C
---------------



=
T (kelvins)
V
BE
600K, typ.
V
G0
~ 1.2
GREATLY exaggerated
curvature
Ideal (shown for
reference)
T d
dV
BE
T d
d
V
T
ln
CT
n
A
e
BT
r
---------------



T d
d
V
T
l n
C
A
e
BT
r n
----------------------



= =
T d
dV
BE k
q
--- l n
C
A
e
BT
r n
----------------------



r n ( ) =
T d
dV
BE
V
G0
V
BE
r n ( )V
T
+ [ ]
T
----------------------------------------------------------- =
Handout #20: EE214 Fall 2002
Voltage References and Biasing 1993 Thomas H. Lee; rev. November 27, 2002; All rights reserved Page 10 of 15
Note that the curvature term disappears if r = n, and were left with the same expression
for the temperature coefficient as derived earlier. In the Brokaw cell, n = 1 to a good
approximation (depending on the temperature coefficient of the resistors), which reduces
the effect of, but does not cancel, r (remember, r is typically a minimum of 2, and can
range up to about 6). Graphically, think of the increasing collector current with tempera-
ture as straightening out the V
BE
curve.
The next question is, how does the curvature term affect the bandgap reference itself? The
most expedient answer comes from deriving the condition for net zero TC. Suppose we
call GV
T
the PTAT component that we add to V
BE
, where G is a constant to be found.
Then, the TC of the PTAT component may be written as GV
T
/T and the condition for zero
TC is therefore
(12)
which corresponds to an output voltage of
(13)
This last equation depends on V
T
and therefore implies that the output voltage cannot have
zero TC at all temperatures; the best we can do is achieve zero TC at one temperature. Fur-
thermore, to achieve this zero TC condition at that one temperature, we need to adjust the
output to a voltage higher than V
G0
by an amount equal to (r - n)V
T
. Fortunately, this cor-
rection term is relatively small, typically amounting to just tens of mV out of a total that is
greater than a volt. Hence, the output only has to be trimmed to a value a few percent
greater than V
G0
at the temperature where zero TC is desired (generally the center of the
operating temperature range).
At this point, wed like to quantify the errors that are caused by the curvature. Unfortu-
nately, while the equations weve developed so far are valuable for design, they arent
quite suitable for analysis.To derive one that is, let us choose the factor m so that the out-
put voltage has zero TC at some temperature well call T
R
(for the reference temperature).
Throughout, well use the subscript R to denote a variables value at this reference temper-
ature. With this notational convention, we may express V
OUT
as follows:
(14)
or, as some prefer:
(15)
Note that this equation has the right limiting behavior: when T = T
R
, it yields the output
voltage corresponding to the zero TC condition. Also note that, if we were able to arrange
T d
dV
BE
GV
T
T
----------- + 0 G
V
G0
V
BE
r n ( )V
T
+ [ ]
V
T
----------------------------------------------------------- = =
V
OUT
TC 0 =
V
BE
GV
T
+ V
G0
r n ( ) V
T
+ = =
V
OUT
T ( ) V
G0
T
T
R
------ r n ( )V
TR
T
T
R
------ r n ( )V
TR
ln
T
T
R
------


+ =
V
OUT
T ( ) V
G0
T
T
R
------ r n ( )V
TR
1 l n
T
T
R
------


+ =
Handout #20: EE214 Fall 2002
Voltage References and Biasing 1993 Thomas H. Lee; rev. November 27, 2002; All rights reserved Page 11 of 15
for the collector currents to vary as T
n
with n = r, the output voltage would have zero TC
at all temperatures if V
OUT
were adjusted to a value V
G0
at any temperature. This last
observation is at the core of many efforts to synthesize curvature-corrected bandgap refer-
ences.
Even without elaborate curvature correction methods, though, the Brokaw cell (where
n =1 in the classic implementation), provides outstanding performance, because the cur-
vature inherent in bipolars is simply not all that bad, and the Brokaw cell contributes little
error of its own.
To illustrate this point, lets compute the actual error one could expect over a temperature
range of 55 to +150C, if T
R
is chosen as +50C, and if the quantity (r n) ranges from 1
to 5.
As is evident, the total change in output voltage is ideally less than a percent over the
entire temperature range, even with relatively large values of (r n). Furthermore, the out-
put is a maximum at the reference temperature, and drops off above and below this tem-
perature in a quasi-parabolic manner. As a consequence, setting T
R
equal to the center of
the desired operating temperature range nearly minimizes the maximum deviation from
the value at T
R
.
As a final note, the levels of performance in Table1 assume an ideal scenario in which
second-order errors due to device mismatch, nonzero contact and interconnect resistance,
nonzero resistor TC, drift, etc., are ignored. Actual performance will be worse in prac-
tice owing to the combined effect of these sources. Careful layout of all devices is manda-
tory to minimize errors, but even so, there can be a considerable spread in output voltages
(e.g., 50-100mV) for CMOS bandgap implementations.
5.2 Bandgap References in CMOS Technology
The classic Brokaw cell uses bipolar transistors in which all device terminals float, so it
cannot be implemented directly in this form in CMOS technology. Re-arranging the cir-
cuit to accommodate the restrictions placed on the parasitic substrate pnp yields the fol-
lowing:
TABLE 1. Ideal output voltage as function of T and (r n)
r n
V
OUT
@ T
= 55C
V
OUT
@ T
= 50C
V
OUT
@ T
= 150C
Max. Error
1 1.226 1.228 1.227 2mV
2 1.252 1.256 1.253 4mV
3 1.279 1.284 1.280 5mV
4 1.305 1.312 1.307 7mV
5 1.331 1.339 1.333 8mV
Handout #20: EE214 Fall 2002
Voltage References and Biasing 1993 Thomas H. Lee; rev. November 27, 2002; All rights reserved Page 12 of 15
FIGURE 9. CMOS bandgap reference (simplified)
Transistor Q
2
is designed to have m times the emitter area of Q
1
. The quad of CMOS tran-
sistors enforces equal emitter currents, so that the collector current density ratio is only
approximately m. Implicit in the last statement is that this circuit has a greater sensitivity
to than the original Brokaw cell. This unfortunate consequence of being forced to use
the substrate pnps leads to larger errors than the classic bandgap cell, particularly because
is rarely large enough to be ignored (values of 5-10 are typical, and values of unity are
not unheard of). Nevertheless, even a poorly performing bandgap reference is consider-
ably superior to anything that can be built out of pure CMOS components.
Choosing component values for this circuit proceeds in a manner quite similar to that for
the classic cell. Begin by specifying a reference temperature T
R
at which the TC is to be
zero. For illustrative purposes, assume that this temperature is to be 350 kelvins.
The next step is to calculate the target output voltage at this reference temperature. As
mentioned previously, the shallow, heavily-doped p+ diffusions used to make the emitters
lead to relatively large curvature coefficients, with r typically 4 or 5. If no device data or
models are available, a reasonable starting point is to assume a value of 3 for the quantity
(r n). Hence, the target output voltage should be:
(16)
Now assume that we have selected 100A for the individual emitter currents, and that the
larger transistor, Q
2
, has a V
BE
of 0.65V at this current at T
R
. Then R2 is simply:
(17)
R
2
V
DD
R
1
1:m
Start-up
network
V
OUT
M
1
M
2
Q
1
Q
2
I
OUT
V
OUT
V
G0
r n ( )V
T
+ 1.3V =
R2
V
OUT
V
BE2

I
E
------------------------------- 6.5k = =
Handout #20: EE214 Fall 2002
Voltage References and Biasing 1993 Thomas H. Lee; rev. November 27, 2002; All rights reserved Page 13 of 15
Assuming an m equal to 8, then V
BE1
is about 63mV larger than V
BE2
at the reference tem-
perature. Hence,
(18)
thus completing the design.
As a final comment on this circuit, one usually finds that the bias current is relatively con-
stant with temperature because resistors typically have a positive TC which offsets the
PTAT tendency of the core design. Thus, currents from a mirror slaved to the PMOS mir-
ror may be roughly constant. The precise TC obtained may be adjusted through a suitable
choice of resistor type and values if the ultimate goal is to generate a bias current rather
than a reference voltage.
6.0 Constant-g
m
Bias
While a constant current or constant voltage is often desirable, there are important exam-
ples in which it is the transconductance that must be held constant. In those cases, the bias
current must be allowed to vary with temperature (and process) in a way that enforces that
constancy.
A circuit whose bias current corresponds to a g
m
that is inversely proportional to a refer-
ence resistance is a modification of the self-biased CMOS quad of transistors weve
already seen (again, a crude PMOS mirror is shown for simplicity; better mirrors would
generally be used in practice):
FIGURE 10. Basic constant g
m
reference
Here transistor M
2
is a factor m times wider than M
1
, so its V
GS
is the smaller of the two.
Because the voltage across R equals the difference in gate-source voltages, we may write:
R1
V
OUT
V
BE1

I
E
------------------------------- 5.87k = =
R
V
DD
Start-up
network
M
1
M
2
I
REFN
1:m
1:1
Handout #20: EE214 Fall 2002
Voltage References and Biasing 1993 Thomas H. Lee; rev. November 27, 2002; All rights reserved Page 14 of 15
, (19)
where we have neglected body effect. Thus, the two threshold voltages are equal, allowing
us to say that the difference in gate-source voltages equals the difference in overdrive volt-
ages.
Now recall that the transconductance of a long-channel MOSFET may be expressed as
. (20)
The transconductance of M
1
is therefore
. (21)
The ratio of overdrive voltages of two long-channel devices operated at equal currents, but
differing widths is just:
. (22)
Therefore,
. (23)
Note that, miraculously, the transconductance of M
1
depends only on a resistance and a
width ratio; all process parameters have magically disappeared. As long as the resistors
parameters are more stable than those of the transistors, this circuit provides a net advan-
tage over other methods of setting transconductance.
10
For the special case of m = 4, the transconductance of M
1
becomes precisely equal to 1/R.
For this reason, this particular width ratio is frequently used.
A potentially significant source of error in implementations of this circuit is the body
effect, since the two transistors are not operated at equal source-bulk potentials. The
resulting difference in threshold voltages introduces a device-dependent term into the
expression for transconductance. In n-well technologies, this error term can be eliminated
for PMOS transistors by connecting the well of each transistor to its source. To take care
of this problem for the NMOS transistors, one may employ the following modification:
10. If on-chip resistor options are unsatisfactory, one may always employ an external resistor with superior
characteristics.
I
2
V
GS
R
-------------
V
OD
R
-------------- I
1
I = = = =
g
m
2I
V
OD
---------- =
g
m1
2I
V
OD1
-------------
2V
OD
RV
OD1
------------------
2
R
--- 1
V
OD2
V
OD1
-------------


= = =
V
OD2
V
OD1
-------------
W
1
W
2
-------
1
m
-------- = =
g
m1
2
R
--- 1
1
m
--------


=
Handout #20: EE214 Fall 2002
Voltage References and Biasing 1993 Thomas H. Lee; rev. November 27, 2002; All rights reserved Page 15 of 15
FIGURE 11. Modification of constant g
m
reference to eliminate error due to body effect
Here, as before, M
1
and M
2
operate at equal drain currents. And, as before, the difference
in their gate-source voltages appears across a resistor R, because the op-amp forces equal
source voltages of M
1
and M
2
. Because the same KVL equation applies to this M
1
-M
2
-R
loop as to that of the original circuit, all of the equations for the preceding case apply to
this one. Whats different is that M
1
and M
2
now operate at equal source-bulk voltages,
preventing body effect from causing an error. Finally, equal-sized M
1
and M
3
operate at
the same gate overdrive because they operate at the same current. They therefore possess
the same transconductance. Mirroring via M
5
thus enables delivery of the magic current to
other NMOS devices to set their transconductances.
7.0 Summary
Weve seen that the self-biased cell is quite versatile, permitting the generation of currents
proportional to the ratio of a voltage in one branch to the resistance in the other. The volt-
age may be provided by a variety of elements, such as a forward-biased junction. While a
V
BE
by itself has limited utility as a voltage reference because of its negative TC, its
CTAT behavior is valuable in compensating the PTAT V
BE
in a bandgap reference cir-
cuit to yield an output roughly equal to V
G0
with extremely small temperature variation.
Even when parasitic bipolars are used in an otherwise CMOS circuit, the bandgap princi-
ple allows the synthesis of more accurate and stable voltages or currents than possible
with ordinary CMOS circuits.
Finally, constant-g
m
bias circuits were presented, allowing the stable biasing of transcon-
ductance-sensitive circuits, such as some types of active filters and amplifiers.
R
V
DD
M
5
M
4
I
OUT

+
M
2
M
1
M
3
1:m
1:1