See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/2977181

Low-power digital filtering using approximate processing

Article in IEEE Journal of Solid-State Circuits · April 1996

DOI: 10.1109/4.494201 · Source: IEEE Xplore

CITATIONS READS
145 1,588

3 authors, including:

S. Hamid Nawab
Boston University
120 PUBLICATIONS 3,901 CITATIONS

SEE PROFILE

All content following this page was uploaded by S. Hamid Nawab on 02 April 2015.

The user has requested enhancement of the downloaded file.

IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 31, NO. 3, MARCH 1996 395

Low-Power Digital Filtering

Using Approximate Processing
Jeffrey T. Ludwig, S. Hamid Nawab, and Anantha P. Chandrakasan

Abstract—We present an algorithmic approach to the design of computation rate. For such applications, an architecture-driven
low-power frequency-selective digital filters based on the concepts voltage scaling approach has previously been developed in
of adaptive filtering and approximate processing. The proposed
which parallel and pipelined architectures can be used to
approach uses a feedback mechanism in conjunction with well-
known implementation structures for finite impulse response compensate for increased delays at reduced voltages [1]. This
(FIR) and infinite impulse response (IIR) digital filters. Our strategy can result in supply voltages in the 1 to 1.5 V range by
algorithm is designed to reduce the total switched capacitance using conventional CMOS technology. Power supply voltages
by dynamically varying the filter order based on signal statistics. can be further scaled using reduced threshold devices. Circuits
A factor of 10 reduction in power consumption over fixed-order
filters is demonstrated for the filtering of speech signals. operating at power supply voltages as low as 70 mV (at 300 K)
and 27 mV (at 77 K) have been demonstrated [2] [3].
Once the power supply voltage is scaled to the lowest
I. INTRODUCTION
possible level, the goal is to minimize the switched capacitance

T ECHNIQUES for reducing power consumption have be-

come important due to the growing demand for portable
multimedia devices. Since digital signal processing is per-
at all levels of the design abstraction. At the logic level, for
example, modules can be shut down at a very low level based
on signal values [4]. Arithmetic structures (e.g., ripple carry
vasive in such applications, it is useful to consider how versus carry select) can also be optimized to reduce transi-
algorithmic approaches may be exploited in constructing low- tion activity [5]. Architectural techniques include optimizing
power solutions. the sequencing of operations to minimize transition activity,
A significant number of DSP functions involve frequency- avoiding time-multiplexed architectures which destroy sig-
selective digital filtering in which the goal is to reject one or nal correlations, using balanced paths to minimize glitching
more frequency bands while keeping the remaining portions transitions, etc. At the algorithmic level, the computational
of the input spectrum largely unaltered. Examples include complexity or the data representation can be optimized for
lowpass filtering for signal upsampling and downsampling, low power [6].
bandpass filtering for subband coding, and lowpass filtering Another approach to reduce the switched capacitance is to
for frequency-division multiplexing and demultiplexing. The lower . Efforts have been made to minimize by intelli-
exploration of low-power solutions in these areas is therefore gent choice of algorithm, given a particular signal processing
of significant interest.
task [7]. In the case of conventional filter design, the filter
To first order, the average power consumption, , of a
order is fixed based on worst case signal statistics, which is
digital system may be expressed as
inefficient if the worst case seldom occurs. More flexibility
(1) may be incorporated by using adaptive filtering algorithms,
which are characterized by their ability to dynamically adjust
the processing to the data by employing feedback mechanisms.
where is the average capacitance switched per operation
In this paper, we illustrate how adaptive filtering concepts may
of type (corresponding to addition, multiplication, storage,
be exploited to develop low-power implementations for digital
or bus accesses), is the number of operations of type
filtering.
performed per sample, is the operating supply voltage,
Adaptive filtering algorithms have generally been used to
and is the sample frequency.
dynamically change the values of the filter coefficients, while
Real-time digital filtering is an example of a class of appli-
maintaining a fixed filter order [8]. In contrast, our approach
cations in which there is no advantage in exceeding a bounded
involves the dynamic adjustment of the filter order. This
Manuscript received October 23, 1995; revised December 13, 1995. This approach leads to filtering solutions in which the stopband
work was sponsored in part by the Department of the Navy, Office of the
Chief of Naval Research, Contract N00014-93-1-0686 as part of the Advanced energy in the filter output may be kept below a specified
Research Projects Agency’s RASSP program. threshold while using as small a filter order as possible. Since
J. T. Ludwig is with the Digital Signal Processing Group, Massachusetts
Institute of Technology, Cambridge, MA 02139 USA.
power consumption is proportional to filter order, our approach
S. H. Nawab is with the Knowledge-Based Signal Processing Group, Boston achieves power reduction with respect to a fixed-order filter
University, Boston, MA 02215 USA. whose output is similarly guaranteed to have stopband energy
A. P. Chandrakasan is with the Massachusetts Institute of Technology,
Cambridge, MA 02139 USA. below the specified threshold. Power reduction is achieved by
Publisher Item Identifier S 0018-9200(96)02454-7. dynamically minimizing the order of the digital filter.
0018–9200/96$05.00  1996 IEEE
396 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 31, NO. 3, MARCH 1996

The idea of dynamically reducing cost (in our case, power

consumption) while maintaining a desired level of output
quality (in our case, stopband energy in the filter output)
emanates from the concept of approximate processing in
computer science [9]. While approximate processing concepts
may be used to describe a variety of existing techniques in
digital signal processing (DSP), communications, and other
areas, there has recently been progress in formally using these
concepts to develop new DSP techniques [10]–[12]. Since our
adaptive filtering technique falls into this category, we refer
to our approach as adaptive approximate filtering, or simply
approximate filtering.

II. DIGITAL FILTERING TRADE-OFFS

A frequency-selective digital filter may have either a finite
impulse response (FIR) or an infinite impulse response (IIR).
It is well known that IIR filters use fewer taps than FIR
Fig. 1. Frequency response magnitudes for FIR filters of orders N = 20; 80;
filters in order to provide the same amount of attenuation in and 140.
the stopband region. However, IIR filters introduce nonlinear
frequency dispersion in the output signals which is unac-
ceptable in some applications. For such cases, it is desirable
to use symmetric FIR filters because of their linear phase
characteristic.
An important family of symmetric FIR filters corresponds
to the symmetric windowing of the impulse responses of
corresponding ideal filters. For example, a lowpass filter of
this type has an impulse response given by [13]

(2) Fig. 2. Tapped delay line of an FIR filter structure, and the powering down
concept To preserve phase linearity, powering down must be applied at both
ends of the structure.
where is a symmetric -point window. This filter has
cutoff frequency and may be implemented using a tapped
delay line with taps. For the purposes of this paper, we
refer to such a filter as having order . In Fig. 1, we display
the frequency response magnitudes for three different values
of when is a rectangular window and .
It should be observed that the mean attenuation beyond the
cutoff frequency increases with filter order. Furthermore,
with respect to a tapped delay-line implementation (see Fig. 2),
the taps of the shorter Type I filters are subsets of the taps of
the longer Type I filters. This ensures that if the filter order
is to be decreased without changing the cutoff frequency, we
can simply power down portions of the tapped delay line for
the higher order filter. The price paid for such powering down
is that the stopband attenuation of the filter decreases. Fig. 3. Cascade implementation of an IIR filter structure. The detail of one
Butterworth IIR filters are commonly used for performing of the second-order sections is shown.
frequency-selective filtering in applications where frequency
dispersion is tolerable. The frequency response magnitudes of sections in its cascade implementation. An interesting property
such filters do not suffer from the ripples which can be seen in of IIR Butterworth filters is that if the second-order sections
the frequency response magnitudes for FIR filters. These IIR are appropriately ordered, one may sequentially power down
filters are commonly implemented as cascade interconnections the later second-order sections and effectively decrease the net
of second-order sections, each of which consists of five stopban attenuation of the filter.
multiplies and four delays, as shown in Fig. 3. Also in Fig. 3
is an illustration of a cascade structure for an eighth-order
IIR filter as the cascade of four second-order sections. For the III. ADAPTIVE APPROXIMATE FILTERING
purposes of this paper, we consider the order of a Butterworth In this section we present the details of our approximate
IIR filter to be equal to twice the number of second-order processing approach to low-power frequency-selective filter-
LUDWIG et al.: LOW-POWER DIGITAL FILTERING USING APPROXIMATE PROCESSING 397

ing. As discussed earlier, frequency-selective filters are used

in applications where the goal is to extract certain frequency
components from a signal while rejecting others. Suppose a
signal, , consists of a passband component, , and a
stopband component, . That is,
(3)

If it were possible to cost-effectively measure the strength of

the stopband component, , from observation of , we
could determine how much stopband attenuation is needed at
any particular time. When the energy in increases, it is
desirable to increase the stopband attenuation of the filter. This Fig. 4. Overview of approximate filtering strategy.
can be accomplished by using a higher-order filter. Conversely,
the filter order may be lowered when the energy in
decreases. We have developed a practical technique, based where denotes the stopband region. Since for every sample
upon adaptive filtering principles, for dynamically estimating period this approach requires an expensive search over the
the energy fluctuations in the stopband component, , and stored values of , we have designed a more efficient
using them to adjust the order of a frequency-selective FIR or strategy which incrementally updates the most recent filter
IIR filter. As described in the previous section, the decrease order. In this case, we estimate the stopband energy in the
in filter order enables the powering down of various segments output as
of the filter structure. Powering down of the higher order taps (9)
has the effect of reducing the switched capacitance at the cost
of decreasing the attenuation in the stopband. Assuming that The decision rule for choosing is then given by
the FIR delay line is implemented using SRAM, even the data
shifting operation of the higher order taps can be eliminated
through appropriate addressing schemes.
Our overall technique is depicted in Fig. 4. The quantity (10)
, which represents the energy differential between the input
and the output, is obtained as where and are application-specific parameters. It
should be noted that the filter order is changed at most by
(4) during each sample period.
The parameters and in (10) control the sensitivity
where
of the time evolution of the filter order. The choice of the
parameter in (5) and (6) involves a trade-off between
(5) suppression of sensitivity to local fluctuations and preservation
of the possible time-varying nature of the signal energy. For
and the case of FIR filters, we also observe that when the value
of is less than the maximum filter order, there is no extra
(6) storage required to compute beyond that required for
the filter implementation. On the other hand, excess storage is
always required to update .
The filter order for sample period is updated at
The arithmetic cost of the update process can be easily
each sample period. One approach for the update process is
shown to involve five multiplications, five additions, one table
to choose to be the smallest positive integer which
lookup from a small memory module, and simple control.
guarantees that the stopband energy, of the output signal
This cost is roughly equivalent to that of increasing the FIR
will be maintained below a specified threshold . Assuming
filter order by five or the IIR filter order by two. This, for
that the stopband portion of the input spectrum is essentially
example, means that net power savings can be expected in
flat,1 the stopband energy in the output can be estimated as
the FIR case if for significant periods of time the dynamic
(7) FIR filter order decreases by more than five with respect to
the maximum filter order. The overhead of multiplication is
where is a proportionality constant, and represents reduced to one multiplication instead of five per update if
the stopband energy in the frequency response, , of the absolute value operations are used to compute instead
th order filter. That is, of magnitude-squared operations.
(8)
IV. RESULTS
1 In practice we have found that this flatness constraint may be relaxed In the context of FIR filters, we have used simulations of
considerably without detrimental effects. our approximate filtering technique to show that reduction in
398 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 31, NO. 3, MARCH 1996

TABLE I
FILTERING PERFORMANCE FOR DEMODULATING FDM SPEECH

Fig. 5. FIR filter stopband energy, ESB [k ] versus filter order, k, for the
rectangular window family of FIR filters.

power consumption by an order of magnitude is achieved over

fixed-order filter implementations when the stopband energy
of the output signal is stipulated to remain below a given
threshold . The context for most of these simulations is
frequency-division demultiplexing of pairs of speech wave-
forms.
1) The Speech Signals: Each of the speech signals used in
our simulations was sampled at 8 KHz and normalized to have
maximum amplitude of unity. Each signal corresponds to a
complete sentence with negligible silence at its beginning and
end. Fig. 6. Evolution of filter order for an FDM example. Two plots are shown
in the figure. One shows the filter order as a function of time, while the other
2) Frequency-Division Multiplexing: Each digitized speech shows the stopband energy of the input signal as a function of time.
waveform was pre-filtered to have a maximum frequency of
1.5 KHz. A guard band of 1 KHz was used in multiplexing a consumption of the approximate filter with respect to a fixed-
reference speech signal (corresponding to the sentence, “That order filter which is guaranteed to keep the stopband energy in
shirt seems much too long,”) with each of the other speech the output below for all times. We observe that our adaptive
signals. The reference signal always occupied the 0 to 1.5 technique reduces the average power consumption by a factor
KHz band, while the other signals always occupied the 2.5 of 5.9.
KHz to 4 KHz band. To gain further insight into the source for this power
3) The Demultiplexing: Demultiplexing involves lowpass reduction, in Fig. 6 we illustrate the nature of the adaptation
filtering (cutoff frequency 2 KHz) to isolate the reference performed by our technique in the case of one of the FDM
speech signal. The approximate filtering technique was used signals. One of the curves shows the evolution of the filter
to perform this lowpass filtering for each of the 10 frequency- order while the other curve shows the energy profile of
division multiplexed (FDM) signals. The parameter values in the stopband signal. Clearly, the variations in filter order
(10) were chosen to be roughly follow the energy variations of the stopband signal.
(11) In particular, the most power savings is achieved during the
silence regions of the stopband signal.
The family of FIR filters used in these simulations corresponds 5) Speech Communication Implications: Longer periods of
to (2) with rectangular. The values of for this speech communication generally include significantly larger
case are plotted in Fig. 5. fractions of silence periods than an individual sentence. To
4) Performance: In Table I we have listed various mea- factor this into our analysis, we repeated our simulations
sures obtained for the performance of the approximate filter as while inserting additional silence at the end of each speech
it was applied to each FDM signal. The first column contains signal. The average (over all 10 cases) of the relative power
the sentence number for the stopband component of the input consumption is displayed in Fig. 7 as a function of the silence
signal. The second and third columns, respectively, list the duration relative to the duration of the entire signal. As
minimum and maximum filter orders used by the approximate expected, the power reduction improves as the relative amount
filter in each case. The final column shows the relative power of silence is increased.
LUDWIG et al.: LOW-POWER DIGITAL FILTERING USING APPROXIMATE PROCESSING 399

V. CONCLUSIONS
An algorithm-based approach has been presented for ob-
taining low-power implementations of important classes of
IIR and FIR digital filters. In this approach, adaptive filtering
and approximate processing concepts are combined to design
digital filters which have the important property that the
filter order can be dynamically varied in accordance with
the stopband energy of the input signal. Simulations of the
proposed technique using a variety of speech signals have
shown that our approach offers significant power savings over
standard fixed-order implementations. Finally, we note that
while we illustrated our proposed technique in the context of
lowpass filtering applications, it is equally applicable to other
types of frequency-selective filtering.

REFERENCES
Fig. 7. Filter performance versus percentage silence in stopband signal. [1] A. P. Chandrakasan, S. Sheng, and R. W. Brodersen, “Low-power
digital CMOS design,” IEEE J. Solid State Circuits, pp. 473–484, Apr.
1992.
[2] R. M. Swanson and J. D. Meindl, “Ion-implanted complementary MOS
transistors in low-voltage circuits,” IEEE J. Solid-State Circuits, vol.
SC-7, no. 2, pp. 146–152, Apr. 1972.
[3] J. B. Burr, “Cryogenic ultra low power CMOS,” in 1995 IEEE Symp.
on Low Power Electronics, Oct. 1995, pp. 82–83.
[4] M. Alidina, J. Monteiro, S. Devadas, A. Ghosh, and M. Papaefthymiou,
“Precomputation-based sequential logic optimization for low power,”
in 1994 Int. Workshop on Low-Power Design, Apr 1994, pp.
57–62.
[5] T. Callaway and E. Swartzlander, Jr., “Optimizing arithmetic elements
for signal processing,” in VLSI Signal Processing, V, pp. 91–100, IEEE
Special Publications, 1992.
[6] A. Chandrakasan and R. Brodersen, Low Power Digital CMOS Design.
Norwell, MA: Kluwer, July 1995.
[7] B. M. Gordon and T. Meng, “A low power subband video decoder
architecture,” in Proc. Int. Conf. on Acoustics, Speech, and Signal
Processing, Apr. 1994, pp. II-409–412.
[8] S. Haykin, Adaptive Filter Theory. Englewood Cliffs, NJ: Prentice-
Hall, 1991.
[9] V. R. Lesser, J. Pavlin, and E. Durfee, “Approximate processing in
real-time problem solving,” AI Mag., pp. 49–61, Spring, 1988.
[10] S. H. Nawab and E. Dorken, “A framework for quality versus effi-
ciency tradeoffs in STFT Analysis,” IEEE Trans. Signal Processing, pp.
998–1001, Apr. 1995.
[11] S. H. Nawab and J. Winograd, “Approximate signal processing using
incremental refinement and deadline-based algorithms,” in Proc. IEEE
Int. Conf. on Acoustics, Speech, and Signal Processing, Apr. 1995,
Detroit, MI, pp. 2857–2860.
Fig. 8. Filter order evolution for the approximate filtering subband decom- [12] J. T. Ludwig, S. H. Nawab, and A. Chandrakasan, “Low power filtering
position example. The top plot shows the filter order as a function of time, using approximate processing for DSP applications,” in Proc. Custom
which tracks the input’s stopband component xs [n], which is shown in the Integrated Circuits Conf. (CICC), Santa Clara, CA. May, 1995, pp.
bottom plot. 185–188.
[13] A. Oppenheim and R. Schafer, Discrete-Time Signal Processing. En-
glewood Cliffs, NJ: Prentice-Hall, 1989.
6) Subband Coding: Data compression techniques for
voice signals often use a binary tree-structured filterbank of
highpass and lowpass filters, as depicted at the top of Fig. 8.
Each of these filters may be implemented using the proposed
approximate filtering technique. To illustrate the potential for
Jeffrey T. Ludwig received the S.B. degree in aero-
power savings in the first stage of the subband decomposition, nautics and astronautics in 1991 and the S.M. degree
an approximate FIR lowpass filter was applied to a speech in electrical engineering in 1993, both from the
signal, , corresponding to the sentence, “That shirt seems Massachusetts Institute of Technology, Cambridge,
MA.
much too long.” The time-varying FIR filter order used by He is currently a graduate student in the Digital
our technique is shown in the top plot of Fig. 8. The bottom Signal Processing Group of the Research Laboratory
plot in Fig. 8 shows the input’s stopband component, , to of Electronics at MIT, pursuing a Ph.D. in electrical
engineering. His research interests are in digital
demonstrate that the filter order roughly tracks the stopband signal processing and its applications.
energy of the input signal.
400 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 31, NO. 3, MARCH 1996

S. Hamid Nawab received the S.B., S.M., and Anantha P. Chandrakasan received the B.S, M.S.,
Ph.D. degrees in electrical engineering from the and Ph.D. degrees in electrical engineering and
Massachusetts Institute of Technology in 1977, computer sciences from the University of California,
1979, and 1982, respectively. Berkeley, in 1989, 1990, and 1994, respectively.
He is currently an Associate Professor in the Since September 1994, he has been the analog
Department of Electrical, Computer, and Systems devices career development Assistant Professor of
Engineering at Boston University. He has held Electrical Engineering at the Massachusetts Institute
visiting professorships in electrical engineering of Technology, Cambridge. His research interests
at MIT (1994–95) and in computer science at include the ultra low power implementation of cus-
University of Massachusetts at Amherst (1988–89). tom and programmable digital signal processors,
His research primarily involves the exploration of wireless sensors and multimedia devices, emerging
new algorithms and architectures for digital and knowledge-based signal technologies, and CAD tools for VLSI. He is a co-author of the book titled
processing. He is co-editor of the book, Symbolic and Knowledge-based Low Power Digital CMOS Design (Kluwer).
Signal Processing (Prentice-Hall, 1992). He also joins A. V. Oppenheim and Dr. Chandrakasan has received the NSF Career Development Award and
A. S. Willsky on the forthcoming second edition of their Prentice-Hall text the IBM Faculty Development Award. He received the IEEE Communications
on Signals and Systems. Society 1993 Best Tutorial Paper Award for the IEEE Communications
Dr. Nawab is the winner of the 1988 Paper Award from the IEEE Magazine paper titled, “A Portable Multimedia Terminal.”
Signal Processing Society for his paper entitled “Direction Determination
of Wideband Signals.” He is also the recipient of the 1993 Metcalf Award
for Excellence in Teaching at Boston University.

View publication stats