IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS, VOL. 58, NO. 4, APRIL 2011 205
A High-IIP3 Third-Order Elliptic Filter With
Current-Efficient Feedforward-Compensated Opamps
Nagendra Krishnapura, Abhishek Agrawal, and Sameer Singh
Abstract—A low-distortion active filter is realized using current-
efficient feedforward-compensated operational amplifiers in the
integrators and feedforward current injection in the summing
amplifier. A third-order elliptic low-pass filter with two possible
bandwidth settings of 17 and 8.5 MHz consumes 1.8 mW from a
1.8-V supply and occupies 0.17 mm
2
in a 0.18-μm CMOS process.
The measured maximum signal-to-noise and distortion ratios at
the two bandwidth settings are 50.5 and 52.5 dB, respectively. The
corresponding third-order intermodulation intercept points (IIP3)
are +28.2 and +30.8 dBm. Automatic tuning is used at the startup
to counter process variations and set the bandwidth accurately.
Index Terms—Automatic tuning, continuous-time filters, elliptic
filters, feedforward compensation.
I. INTRODUCTION
I
N zero-intermediate-frequency integrated-circuit radio re-
ceivers, low-pass filters [1]–[3] are used before analog/
digital conversion. Their bandwidths range from a few mega-
hertz to a few tens of megahertz. A switchable bandwidth is re-
quired for multistandard radios. For example, 802.11 [wireless
local-area network (WLAN)] has twice the signal bandwidth
in the “n” mode as in the “a, b, and g” modes [4]. Another
important requirement for these filters is a low distortion per-
formance, usually specified using the third-order intermodula-
tion intercept point (IIP3). Among integrated continuous-time
filters, active RC filters provide the best distortion performance
at medium frequencies because they consist of integrators and
amplifiers using operational amplifiers (opamps) in negative
feedback loops, which suppress the distortion arising from
active elements. The distortion in these filters can be reduced
by reducing the virtual-ground voltage swings, which are the
“error” voltages of feedback loops around opamps. This can
be achieved by increasing the gain of the opamp or reducing
the output current driven by the opamp. For a given unity
gain frequency and bias current, higher gain at low frequencies
can be obtained using feedforward-compensated opamps [5]
instead of their Miller-compensated counterparts. The output
current driven from the opamp can be reduced by injecting
a replica of the current to the output through a separate path
[6]. Both these techniques lower the voltage variations at the
virtual-ground nodes of opamps, making the operation of the
feedback loop more ideal. This brief presents a low-distortion
third-order elliptic filter that uses power-efficient feedforward-
Manuscript received July 8, 2010; revised October 19, 2010 and December 2,
2010; accepted February 11, 2011. Date of current version April 20, 2011. This
paper was recommended by Associate Editor R. Martins.
The authors are with the Department of Electrical Engineering, Indian
Institute of Technology, Madras, Chennai 600036, India (e-mail: nagendra@
iitm.ac.in).
Digital Object Identifier 10.1109/TCSII.2011.2124571
Fig. 1. (a) Filter block diagram. (b) Integrator realization (single-ended
equivalent shown for clarity). (c) Automatic tuning scheme.
compensated opamps in the integrator and partial injection of
the opamp output current in the summing amplifier. The filter
incorporates automatic tuning to correct for process variations,
and its bandwidth is programmable to 17 and 8.5 MHz to cater
to the two modes mentioned above.
This brief is organized as follows. Section II describes the fil-
ter architecture and automatic tuning and illustrates the advan-
tages of using feedforward-compensated opamps in active-RC
filters. A current-efficient feedforward-compensated opamp is
described in Section IV. The summing amplifier with current
injection for the distortion reduction is described in Section V.
The measured results are discussed in Section VI. Section VII
concludes the brief.
II. T
HIRD-ORDER ELLIPTIC FILTER
Fig. 1(a) shows the block diagram of the third-order elliptic
filter. The integrator outputs are scaled for equal maximum
transfer function magnitude. Outputs of a cascade of first- and
second-order sections are summed to realize the transmission
zero. Fig. 1(b) shows the realization of each integrator. A 5-bit
control word b4:0 switches the resistors and the capacitors
and varies the RC product from 55% to 175% of the mid-
code value to compensate for process variations. Doubling the
feedback capacitor array, as shown in Fig. 1(b), halves the
bandwidth to 8.5 MHz while maintaining the same passband-
noise spectral density as required in WLAN receivers.
1549-7747/$26.00 © 2011 IEEE
206 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS, VOL. 58, NO. 4, APRIL 2011
Fig. 2. Macromodel of (a) feedforward- and (b) Miller-compensated opamps
(single-ended equivalents shown for clarity).
The automatic tuning scheme is shown in Fig. 1(c). It uses
an additional integrator, which is a replica of Fig. 1(b). With
nominal resistor and capacitor values, the replica integrator’s
time constant 1/ω
0
is 100 ns when the digital bits are set to
midcode. Across process variations, the time constant can be
tuned to 100 ns by setting the five control bits. The appropriate
control bits are generated as follows. The replica integrator is
initially reset, and an internally generated reference voltage V
ref
(approximately 400 mV) is applied to it for a period of 100 ns.
The 100-ns interval is generated using an accurate external
clock. At the end of the integration period, the output of the
integrator is compared with V
ref
. The tuning bits are updated
based on the result of the comparison [7]. The integrator is reset,
and the cycle repeats. After five cycles of successive approxi-
mation, the bits converge to a value that sets the time constant
of the replica integrator to 100 ns. The same bits are applied
to all the integrators, and by the virtue of matching, their unity
gain frequencies are set to their correct values. Due to the digital
nature of tuning, there can be up to 5% error in the bandwidth.
The tuning circuitry is disabled after the process is complete.
III. A
DVANTAGES OF FEEDFORWARD OPAMPS
Fig. 2 shows the macromodels of the feedforward- and
Miller-compensated opamps. The first stage consists of g
m1
loaded by R
L1
and C
1
, and the second stage consists of g
m2
loaded by R
L2
and C
L
. In Fig. 2(a), the feedforward stage g
mf
provides a high-speed path from the input to the output for
frequency compensation [5]. In Fig. 2(b), C
c
(with zero can-
celing resistor 1/g
m2
in series) is used across the second stage
for frequency compensation [8]. The cascade of two stages
using g
m1
and g
m2
provides a high gain in both cases. Table I
compares the parameters of the two opamps. For stability, the
nondominant pole p
2
of the Miller-compensated opamp has to
be higher than its unity gain frequency ω
u
. Therefore, for a
given unity gain frequency ω
u
and a given load capacitance C
L
,
g
m2
in the Miller-compensated opamp has to be significantly
more than g
mf
in the feedforward opamp. The input-referred
noise voltage at low frequencies is determined only by the first
stage g
m1
in both cases.
TAB LE I
C
OMPARISON BETWEEN FEEDFORWARD AND MILLER OPA MPS
Fig. 3. (a) Magnitude response of the opamp models in Fig. 2. (b) S/IM3 of
the elliptic filter using the opamp models in Fig. 2. The input consists of two
tones 1 MHz apart and with a 225-mV peak each. Their center frequency is
swept from 2.5 to 19.5 MHz.
Fig. 3(a) shows the gain magnitude response of Miller- and
feedforward-compensated opamps designed for a dc gain of
66 dB and a unity gain frequency of 1.8 GHz while driving
a load of 45 fF. The numerical values of all components are
shown in Fig. 2. The first stage has a gain of about 40 dB.
The feedforward-compensated opamp has a significantly higher
gain in a certain range of frequencies. This results in a lesser
distortion when operated in a closed loop.
To verify this, the filter in Fig. 1(a) is realized using
the opamps in Fig. 2 with weak nonlinearities added to
transconductors and resistors. Transconductors are modeled
as i
out,gm
= I
0
tanh(v
i,gm
/V
0
), and resistors are modeled as
i
R
= I
0
tanh(v
R
/V
0
). The small signal (trans)conductance is
given by I
0
/V
0
, and the value of V
0
determines the nonlinearity.
In our models, V
0
is 100 mV for g
m1
, 225 mV for g
m2,f
, and
600 mV for R
L1,L2
. The filter is simulated with the opamp
macromodels in a circuit simulator, and the resulting ratio
of the signal (S) to the third-order intermodulation distortion
product (IM3) is shown in Fig. 3(b). The distortion of the filter
with feedforward-compensated opamps is significantly lower
than that with Miller-compensated opamps. The input-referred
noise voltage of the two opamps is the same because the same
value of g
m1
is used for both. The total transconductance used
in the Miller-compensated opamp is 2 mS, and that used in
the feedforward-compensated opamp is 1.2 mS, implying pro-
portionately smaller power consumption in the latter. We can
therefore conclude that filters using feedforward-compensated
opamps will have a better dynamic range for a given power
dissipation.
IV. C
URRENT-EFFICIENT FEEDFORWARD OPAMP
Fig. 4 shows the feedforward-compensated opamp proposed
in [5]. A cascade of two n-channel MOS (nMOS) differential
KRISHNAPURA et al.: HIGH-IIP3 THIRD-ORDER ELLIPTIC FILTER WITH FEEDFORWARD-COMPENSATED OPAMPS 207
Fig. 4. Feedforward-compensated opamp in [5].
Fig. 5. Feedforward-compensated opamp with shared bias currents.
pairs (M
1
,M
2
) provides a high gain. Feedforward compensa-
tion is provided by using another nMOS differential pair M
3
.
A more current-efficient structure for feedforward compen-
sation is to use a p-channel MOS differential pair M
3
, which
shares its bias current with the second-stage differential pair
M
2
, as shown in Fig. 5. Separate common-mode feedback
circuit (CMFB) stages are used to drive current sources M
4
and
M
5
and to stabilize the common-mode output of each stage.
Fig. 6 shows the CMFB circuits used for the two stages. The
common-mode voltage is fed back through g
m,cm
(a differential
pair with a current mirror load). C
c
compensates the CMFB
loop by providing a fast path to the gates of M
4
or M
5
.For
the first-stage CMFB loop [see Fig. 6(a)], source followers are
used to drive the common-mode detector so that a high dc gain
is maintained. The opamp consumes about 0.2 mA and is used
for all integrators and the summing amplifier in Fig. 1.
V. S
UMMING AMPLIFIER
Fig. 7(a) shows the conventional summing amplifier. Because
multiple inputs are summed, the loop gain tends to be lower, and
the distortion tends to be higher than in an amplifier with a sin-
gle input. With an ideal opamp, the virtual ground of the opamp
is at zero. With a real opamp, the virtual ground experiences
a signal swing proportional to (a
1
x
1
+ a
2
x
2
+ a
3
x
3
)/R, i.e.,
the current driven from the opamp. If the swing at the virtual-
ground node is reduced, the weak nonlinearities of the opamp
Fig. 6. CMFB for the (a) first and (b) second stages of the opamp in Fig. 5.
Fig. 7. (a) Conventional summing amplifier. (b) Summing amplifier with
feedforward injection of the output current. (c) Our implementation with partial
injection from only the largest contributor. (d) Transconductor implementation.
are exercised to a lesser extent, leading to lower distortion.
Conventionally, this is done by increasing the gain of the opamp
(at the relevant signal frequencies). Alternatively, the same can
be accomplished by lowering the current driven from the opamp
[6]. Fig. 7(b) shows the summing amplifier with a current
(a
1
x
1
+ a
2
x
2
+ a
3
x
3
)/R injected at the output. As the opamp
has to deliver zero current, the virtual-ground voltage is zero,
and the summing operation is ideal. In our implementation [see
Fig. 7(c)], only a
3
x
3
/R, which is the largest component of the
output current, is injected to the output using a transconductor.
Fig. 7(d) shows the implementation of the transconductor used
in Fig. 7(c). Its transconductance, which depends on R
gm
and
the transconductance of M
a
, is adjusted to be equal to a
3
/R
in the nominal process corner. For the same total bias current
in Fig. 7(a) and (c), the latter has lesser distortion of 4 to 6 dB
over process corners.
VI. M
EASURED RESULTS
The third-order filter with automatic tuning is fabricated in
a 0.18-μm CMOS process. Fig. 8 shows the layout, the test
schematic, and the test board. The buffers in Fig. 8(b) can be
switched to have gains of +A, −A, or 0 (off). By determining
the transfer function of the filter path and the buffer path
208 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS, VOL. 58, NO. 4, APRIL 2011
Fig. 8. (a) Chip layout. (b) Test schematic (the buffers are on chip). (c) Test
board.
Fig. 9. Measured magnitude response for 17- and 8.5-MHz settings.
for both positive and negative settings of the buffer gain, the
external feedthrough can be cancelled, and the transfer function
of the filter can be accurately measured [9]. The measured
magnitude response at 17- and 8.5-MHz settings are shown
in Fig. 9. The attenuation is more than 30 dB for frequencies
more than 50% of the bandwidth. The maximum deviation of
the inband group delay from the average value is 62 ns for the
8.5-MHz bandwidth and 45 ns for the 17-MHz bandwidth. In
the small number of characterized samples, the 3-dB bandwidth
after automatic tuning is 16.3–16.6 MHz in the high-bandwidth
mode and 7.5–8.4 MHz in the low-bandwidth mode. The noise
voltage integrated from 0.1 to 20 MHz are 0.79 mV for the
17-MHz bandwidth and 0.68 mV for the 8.5-MHz bandwidth.
The inband S/IM3 for two tones spaced 1 MHz apart and with a
combined strength of +2 dBm is shown in Fig. 10. As expected,
the worst cases occur near the band edges due to peaking in
the magnitude response at the internal nodes. The variation of
S/IM3 with the amplitude is characterized by feeding two tones
at the worst case frequencies and sweeping their amplitude.
Fig. 11 shows the ratios of signal to noise, signal to IM3, and
Fig. 10. S/IM3 versus frequency in the two bandwidth settings. The input
consists of two sinusoids 1 MHz apart and −1 dBm each.
Fig. 11. Noise and distortion. (a) 17- and (b) 8.5-MHz bandwidth. The input
consists of two tones at the worst case frequencies in Fig. 10.
TAB LE II
P
ERFORMANCE SUMMARY
signal to the sum of noise and IM3 [signal-to-noise and dis-
tortion ratios (SNDRs)] versus the input level. The worst case
inband IIP3 values at the high- and low-bandwidth settings are
+28.2 and +30.8 dBm, respectively. The maximum SNDRs
are 50.5 dB for the 17-MHz bandwidth and 52.5 dB for the
8.5-MHz bandwidth. The filter consumes 1 mA, and the au-
tomatic tuning circuit consumes 270 μA from a 1.8-V supply.
Table II summarizes the performance of the chip.
KRISHNAPURA et al.: HIGH-IIP3 THIRD-ORDER ELLIPTIC FILTER WITH FEEDFORWARD-COMPENSATED OPAMPS 209
TABLE III
C
OMPARISON WITH PUBLISHED FILTERS
VII. CONCLUSION
Power-efficient feedforward-compensated opamps in the in-
tegrators and feedforward current injection in the summing
amplifier enable filters with an inherently high IIP3. To com-
pare this filter to other continuous-time low-pass filters in the
literature, we have used the fundamental relationship between
the power and the bandwidth and the signal-to-noise ratio
(SNR). In a first-order passive RC filter, the ratio of the power
dissipated P
d
to the product of the bandwidth f
B
and the SNR
is a fundamental constant [13], i.e., P
d
/(f
B
× 10
SNR/10
)=
2πkT, where the SNR is in decibels. There is no distortion in a
passive filter and, hence, no upper limit to the signal amplitude.
The power dissipation depends on the signal level. In an active
filter, distortion places an upper limit on the input amplitude.
The power dissipation is usually independent of the signal level.
A figure of merit (FOM) inspired by the above relationship
is the ratio of the power dissipated to the product of order n,
bandwidth f
B
, and the SNR at an input amplitude for which
IM3 is equal to the noise level,
1
and it is given by
FOM = P
d
/
n × f
B
× 10
[2/3(IIP3−N )]/10
. (1)
Table III compares
2
the filter described here to other pub-
lished filters designed for similar bandwidths. The FOM and
2/3(IIP3 − N) are shown for each filter. The table also lists the
attenuation at twice the filter’s bandwidth as a measure of the
width of the transition band. The IIP3 of the filter described
here is at least 8 dB more than that of the others. For 17- and
8.5-MHz bandwidth settings, the FOM of the filters described
here are 0.2 and 0.3 fJ. The FOM of the filters in [2] and [11]
are lower, but these filters have a much broader transition band.
The FOM of the filters presented here is better than those that
have a comparably narrow transition band.
1
At an input signal level equal to the IIP3, the IM3 (extrapolated from small-
signal levels) is also equal to the IIP3. Since the third-order distortion increases
at 3 dB for a 1-dB increase in the input signal, it can be worked out that the
SNR is equal to 2/3(IIP3 − N) when the distortion is equal to noise (both the
IIP3 and N are in decibels).
2
Where the integrated noise is not reported, it is calculated by multiplying
the inband spectral density by the square root of the bandwidth. Because of the
noise peaking near the band edge, this underestimates the noise to some extent.
The IIP3 shown is the worst case inband.
REFERENCES
[1] S. Kousai, M. Hamada, R. Ito, and T. Itakura, “A 19.7 MHz,
fifth-order active-RC Chebyshev LPF for draft IEEE802.11n with auto-
matic quality-factor tuning scheme,” IEEE J. Solid-State Circuits, vol. 42,
no. 11, pp. 2326–2337, Nov. 2007.
[2] S. D’Amico, V. Giannini, and A. Baschirotto, “A 4th-order active-Gm-
RC reconfigurable (UMTS/WLAN) filter,” IEEE J. Solid-State Circuits,
vol. 41, no. 7, pp. 1630–1637, Jul. 2006.
[3] B. Shi and W. Shan, “A Gm-C baseband filter with automatic frequency
tuning for a direct conversion IEEE802.11a wireless LAN receiver,” in
Proc. 30th Eur. Solid-State Circuits Conf., Sep. 2004, pp. 103–106.
[4] F. Horlin and A. Bourdoux, Digital Compensation for Analog Front Ends.
New York: Wiley, 2008, p. 186.
[5] J. Harrison and N. Weste, “A 500 MHz CMOS anti-alias filter using feed-
forward op-amps with local common-mode feedback,” in Proc. IEEE Int.
Solid-State Circuits Conf., Feb. 2003, vol. 1, pp. 132–483.
[6] S. Pavan and P. Sankar, “A 110 μW single bit audio continuous-time
oversampled converter with 92.5dB dynamic range,” in Proc. Eur. Solid-
State Circuits Conf., Sep. 2009, pp. 320–323.
[7] J. Lim, Y. Cho, K. Jung, J. Park, J. Choi, and J. Kim, “A wide-band
active-RC filter with a fast tuning scheme for wireless communication
receivers,” in Proc. IEEE Custom Integr. Circuits Conf., Sep. 2005,
pp. 637–640.
[8] B. Razavi, Design of Analog CMOS Integrated Circuits.NewYork:
McGraw-Hill, 2000.
[9] S. Pavan and T. Laxminidhi, “Accurate characterization of integrated
continuous-time filters,” IEEE J. Solid-State Circuits, vol. 42, no. 8,
pp. 1758–1766, Aug. 2007.
[10] T. Lo and C. Hung, “Low-voltage multi-mode Gm-C channel selection
filter for mobile applications,” in Proc. IEEE Custom Integr. Circuits
Conf., Sep. 2007, pp. 635–638.
[11] S. D’Amico, M. Conta, and A. Baschirotto, “A 4.1-mW 10-MHz
fourth-order source-follower-based continuous-time filter with 79-dB
DR,” IEEE J. Solid-State Circuits, vol. 41, no. 12, pp. 2713–2719,
Dec. 2006.
[12] A. Vasilopoulos, G. Vitzilaios, G. Theodoratos, and Y. Papananos, “A low-
power wideband reconfigurable integrated active-RC filter with 73 dB
SFDR,” IEEE J. Solid-State Circuits, vol. 41, no. 9, pp. 1997–2008,
Sep. 2006.
[13] E. Vittoz, “Low power low-voltage limitations and prospects in
analog design,” in Analog Circuit Design, Low-Power, Low-Voltage, In-
tegrated Filters and Smart-Power,R.J.V.D.Plassche,W.Sansen,and
J. Huijsing, Eds. Boston, MA: Kluwer, 1995.
![Fig. 4. Feedforward-compensated opamp in [5].](/figures/fig-4-feedforward-compensated-opamp-in-5-20vn9jhu.webp)
