Smart Dust: Communicating with a Cubic- Millimeter Computer

Smart Dust: Communicating 

with a Cubic- Millimeter Computer

Decreasing computing device size, increased

connectivity, and enhanced interaction with

the physical world have characterized com-

puting’s history. Recently, the popularity of

small computing devices, such as handheld

computers and cell phones, burgeoning Internet

growth, and the diminishing size and cost of sensors—

especially transistors—have accelerated these trends.

The emergence of small computing elements, with

sporadic connectivity and increased interaction with

the environment, provides enriched opportunities to

reshape interactions between people and computers

and spur ubiquitous computing research.

1

The Smart Dust project

2

is exploring whether an

autonomous sensing, computing, and communication

system can be packed into a cubic-millimeter mote (a

small particle or speck) to form the basis of integrated,

massively distributed sensor networks. Although we’ve

chosen a somewhat arbitrary size for our sensor sys-

tems, exploring microfabrication technology’s limita-

tions is our fundamental goal. Because of its discrete

size, substantial functionality, connectivity, and antic-

ipated low cost, Smart Dust will facilitate innovative

methods of interacting with the environment, provid-

ing more information from more places less intrusively.

We use Smart Dust to pursue projects such as

• deploying defense networks rapidly by unmanned

aerial vehicles or artillery;

• monitoring rotating-compression-blade high-

cycle fatigue;

• tracking the movements of birds, small animals,

and insects;

• monitoring environmental conditions that affect

crops and livestock;

• building virtual keyboards;

• managing inventory control;

• monitoring product quality;

• constructing smart-office spaces; and

• providing interfaces for the disabled.

SMART DUST REQUIREMENTS

Smart Dust requires both evolutionary and revolu-

tionary advances in miniaturization, integration, and

energy management. Designers can use microelectro-

mechanical systems (MEMS) to build small sensors,

optical communication components, and power sup-

plies, whereas microelectronics provides increasing

functionality in smaller areas, with lower energy con-

sumption. Figure 1 shows the conceptual diagram of

a Smart Dust mote. The power system consists of a

thick-film battery, a solar cell with a charge-integrat-

ing capacitor for periods of darkness, or both.

Depending on its objective, the design integrates var-

ious sensors, including light, temperature, vibration,

magnetic field, acoustic, and wind shear, onto the

mote. An integrated circuit provides sensor-signal pro-

cessing, communication, control, data storage, and

energy management. A photodiode allows optical data

reception. We are presently exploring two transmis-

sion schemes: passive transmission using a corner-cube

retroreflector, and active transmission using a laser

diode and steerable mirrors.

The mote’s minuscule size makes energy manage-

ment a key component. Current battery and capacitor

technology stores approximately 1 joule per cubic mm

The Smart Dust project is probing microfabrication technology’s 

limitations to determine whether an autonomous sensing, computing, 

and communication system can be packed into a cubic-millimeter mote 

to form the basis of integrated, massively distributed sensor networks. 

Brett

Warneke

Matt Last

Brian

Liebowitz

Kristofer S.J.

Pister

University of

California,

Berkeley

COVER FEATURE

and 10 millijoules per cubic mm, respectively, whereas

solar cells provide 1 joule per day per square mm in

sunlight and 1 to 10 millijoules per day per square mm

indoors. Our optical receiver consumes approximately

0.1 nanojoule per bit, and the transmitter uses 1 nano-

joule per bit. We expect our analog-to-digital converter

to require 1 nanojoule per sample and computations to

consume less than 1 picojoule per instruction, in con-

trast to present processors such as the CoolRisc 81

3

core, which uses 22 picojoules per instruction, and the

StrongARM SA1100, which consumes approximately

1 nanojoule per instruction. These estimates demon-

strate that for every sensor sample or transmission, we

can perform about 1,000 8-bit operations, so it is

advantageous to exchange extra calculations for fewer

samples or transmitted bits. Further, given our 1 milli-

joule per day of energy from indoor lighting, each sec-

ond we can sample a sensor, think about the result,

and transmit some data.

To determine our research baseline and quickly

develop hardware for testing networking algorithms,

we used commercial off-the-shelf hardware to build a

series of wireless sensor nodes. We used either optical

or radio-frequency communication models to produce

one-cubic-inch devices. Furthermore, other research

groups have used these motes to develop a tiny oper-

ating system

4

and deploy a 100-node network.

Sensors and motors

The multibillion-dollar MEMS industry has been

growing for several decades, with major markets in

automotive pressure sensors and accelerometers, med-

ical sensors, and process control sensors. Recent

advances in technology have put many of these sensor

processes on exponentially decreasing size/power/cost

curves. In addition, variations of MEMS sensor tech-

nology are used to build micromotors; millions of

these micromotors are used in commercially available

projection display systems, such as the Texas

Instruments Digital Micromirror Device. Micro-

motors, combined with Smart Dust, raise the inter-

esting possibility of making synthetic insects (see the

“Microrobotics” sidebar).

COMPUTING AT THE MILLIMETER SCALE

Traditional computer architecture design has focused

on decreasing a given task’s execution time.

5

To accom-

plish this goal, engineers have improved semiconduc-

tor processing exponentially, increasing the transistors’

speed while decreasing their size, thus allowing more

complex architectures that use increased parallelism on

a single die. In contrast, computing in an autonomous

cubic-millimeter package must focus on minimizing a

given task’s energy consumption. Smaller, faster tran-

sistors have reduced parasitic capacitance, thereby

January 2001 45

Passive transmitter with

corner-cube retroreflector

Active transmitter with

laser diode and beam steering

Receiver with photodetector

Analog I/O, DSP, control

Power capacitor

Solar cell

Thick-film battery

Sensors

1-2 mm

Figure 1. Conceptual diagram showing a Smart Dust mote

’s 

major components: a power system, sensors, an optical transceiver, and an integrated circuit. 

46 Computer

resulting in diminished dynamic power consumption.

Constant electric-field scaling has reduced supply volt-

ages, producing dramatic power reductions for both

high-performance and low-energy computing because

dynamic power has a quadratic dependence on supply

voltage. However, constant electric-field scaling also

calls for a reduction in the threshold voltage. This will

result in larger leakage currents, which are already a

concern in the high-performance processors to be

released in 2001 that will leak amps of current.

Therefore, process engineers need to keep leakage cur-

rents low, which will also benefit low-energy designers.

In millimeter-scale computing, the shrinking transis-

tor’s size lets designers compact significant computing

power into this small area. For example, the Intel 8088

core, originally fabricated in a 3-micron process, would

only require 0.12-square millimeter after shrinking lith-

ographically into a current 0.18-micron process, with

a corresponding 100× decrease in energy/instruction.

Low-energy computation

Besides advanced microfabrication technology

processes, using other techniques at every level

achieves low-energy computation. First, because we

use a high-performance process but operate at low

speeds, we can drop the supply voltage to the mini-

mum level at which the devices still function; theoret-

ically this is 0.1 volt,

6

but for 0.5- to 0.2-micron

processes it is more realistically 0.2 to 0.3 volt. To min-

imize current leakage, which can cause significant

power consumption at the low clock rates and duty

cycles that these low-energy architectures use, we can

increase the channel-to-source junction’s reverse bias,

thus increasing the threshold voltage. Initially, adding

two extra supply voltages in this package may seem

onerous; however, if the mote scavenges solar power,

placing two small photodiodes on the integrated cir-

cuit provides the few atto-amps per device necessary

to bias these junctions. Various low-power layout, cir-

cuit, and logic level techniques have been published.

7

Figure 2 shows a consequence of using these tech-

niques—the worst-case energy consumption of an 8-

bit adder in a 0.25-micron process.

The Smart Dust mote’s tasks closely relate to the

physical realm, where the fastest sampling is 10 to 20

kHz for vibration and acoustic sensors so the amount

Richard Yeh

University of California, Berkeley

Add legs or wings to Smart Dust and

you get microrobots. Like Smart Dust,

these synthetic insects will sense, think,

and communicate. In addition, they will

have the ability to move about and inter-

act physically with their environment. We

can use micromachining to build microac-

tuators and micromechanisms, forming

legs and wings, which are integrated with

other Smart Dust components.

The crawling microrobot shown in

Figure A consumes only tens of micro-

watts of power; the motors can lift more

than 130 times the robot’s own weight.

Figure B shows the flying microrobot,

based on the blowfly Calliphora, which

will have a 10- to 25-mm wingspan and

will sustain autonomous flight. De-

velopers folded 50-micron-thick stainless

steel into the desired shape to create the

wings and exoskeleton. Piezoelectric

motors attached to the exoskeleton actu-

ate the wings. These legged and winged

microrobots will consume a total power

of less than 10 milliwatts, provided by

onboard solar cells. For more informa-

tion, visit http://www-bac.eecs.berkeley.

edu/~yeh/currentbot.html for the crawl-

ing microrobot project, and http://robotics.

eecs.berkeley.edu/~ronf/mfi.html for the

flying microrobot project.

Richard Yeh is a PhD candidate at the

Berkeley Sensor and Actuator Center, Uni-

versity of California, Berkeley. His

research interests include microrobotics,

microactuators, and microfabrication

technology. He received an MS in electri-

cal engineering from the University of Cal-

ifornia, Los Angeles. Contact him at

yeh@eecs.berkeley.edu.

Figure A. Model of a crawling microrobot devel-

oped by University of California researchers.

This device measures less than one cubic 

centimeter. Developers folded 2-micron-thick

silicon sheets to create insect-like legs with

microhinges on the folds and joints. The hollow

structure has lightweight, rigid legs. Silicon 

tendons inside the legs couple each rigid leg seg-

ment to electrostatic motors on the robot’s body.

Figure B. A flying microrobot model capable of

autonomous flight. To create the wings and

exoskeleton, the developers folded 50-micron-

thick stainless steel into the desired shapes.

Piezoelectric motors attached to the exoskele-

ton consume less than 10 milliwatts. 

Microrobotics 

of data is small enough that we can use low data trans-

mission rates. Therefore, we can use clock rates in the

1- to 100-kHz range to decrease dynamic power con-

sumption. Despite these low clock rates, the circuits

perform all their transitions during a small portion of

the cycle; then they remain idle. Thus, powering down

blocks for even a few clock cycles saves energy.

Remote programmability

An autonomous cubic-millimeter platform’s com-

puting requirements depend on the target application

because dedicated hardware solutions usually con-

sume less energy than a software solution. To prevent

extraneous power consumption, we need to determine

the minimum amount of programmability necessary

for a useful platform.

The basic mote periodically samples one or more

sensors, stores the values in memory, listens to an

incoming packet, and transmits current or stored data.

Because transmitting the data and sampling the sen-

sors consume more energy than performing a com-

putation, we can add more computation—such as

thresholding, filtering, spectral analysis,

8

classifica-

tion,

9

Doppler shift determination, and encryption to

improve memory use and determine the significance of

readings—thus providing smarter sampling rates and

reducing the data transmission volume.

Remote programmability plays an important role

in millimeter-scale computing. Given their small size

and large numbers, we prefer to program these devices

en masse, without direct connections. Remote pro-

grammability also avoids the costs of recollecting and

reprogramming devices after we deploy them.

COMMUNICATING FROM A GRAIN OF SAND

Smart Dust’s full potential can only be attained

when the sensor nodes communicate with one another

or with a central base station. Wireless communica-

tion facilitates simultaneous data collection from

thousands of sensors. There are several options for

communicating to and from a cubic-millimeter com-

puter. Radio-frequency and optical communications

each have their strengths and weaknesses.

Radio-frequency communication is well under-

stood, but currently requires minimum power levels in

the multiple milliwatt range due to analog mixers, fil-

ters, and oscillators. If whisker-thin antennas of cen-

timeter length can be accepted as a part of a dust mote,

then reasonably efficient antennas can be made for

radio-frequency communication. While the smallest

complete radios are still on the order of a few hun-

dred cubic millimeters, there is active work in acade-

mia and industry to produce cubic-millimeter radios.

Semiconductor lasers and diode receivers are intrin-

sically small, and the corresponding transmission and

detection circuitry for on/off keyed optical communi-

cation is more amenable to low-power operation than

most radio schema. Perhaps most important, optical

power can be collimated in tight beams even from

small apertures. Diffraction enforces a fundamental

limit on the divergence of a beam, whether it comes

from an antenna or a lens. Laser pointers are cheap

examples of milliradian collimation from a millimeter

aperture. To get similar collimation for a 1-GHz radio-

frequency signal would require an antenna 100 meters

across, due to the difference in wavelength of the two

transmissions. As a result, optical transmitters of mil-

limeter size can get antenna gains of one million or

more, while similarly sized radio-frequency antennas

are doomed by physics to be mostly isotropic.

Collimated optical communication has two major

drawbacks. Line of sight is required for all but the

shortest distances, and narrow beams imply the need

for accurate pointing. Of these, the pointing accuracy

can be solved by MEMS technology and clever algo-

rithms, but an optical transmitter under a leaf or in a

shirt pocket is of little use to anyone. We have chosen

to explore optical communication in some depth due to

the potential for extreme low-power communication.

OPTICAL COMMUNICATIONS

We have explored two approaches to optical com-

munications: passive reflective systems and active-

steered laser systems. In a passive communication

system, the dust mote does not require an onboard

light source. Instead, a special configuration of mir-

rors can either reflect or not reflect light to a remote

source; this procedure resembles how a heliograph

operator bounces sunlight off a mirror to flash a

Morse code message to ships—an idea traced to the

fifth century BC, when the Greeks used reflected sun-

light as a beacon signal. Figure 3 shows the corner-

cube retroreflector (CCR)

10

used to adapt this idea to

Smart Dust. Designers have used this device, but on a

January 2001 47

Eight-bit ripple-carry adder energy

Integration time: 2.6 msecs

A = 10000001 (0 × 81), B = 11111111 (0 × FF)

0

0

100

0.5

V

dd

1

A+B

A

B

0 × 81

0 × FF

Energy (fJ)

V

sb

= 0 V

V

sb

= 0.25 V

V

sb

= 0.5 V

Figure 2. Hspice simulated energy consumption for the worst-case addition in an 8-bit

ripple-carry adder implemented in a 0.25-micron complementary metal oxide semicon-

ductor (CMOS) process, demonstrating the effects of supply voltage (V

dd

) and channel-

to-source junction reverse bias (V

sb

) scaling and low-power circuit techniques. All simu-

lation runs were performed over the same time interval and include leakage current

consumption.

48 Computer

macroscale, for years in laser range-finding applica-

tions. A similar device helped scientists determine the

moon’s distance from Earth.

Passive reflective systems

In its simplest passive configuration, the passive-

reflective device consists of three mutually orthogonal

mirrors. Light enters the CCR, bounces off each of the

three mirrors, and is reflected back parallel to the direc-

tion it entered. In the MEMS version, the device has

one mirror mounted on a spring at an angle slightly

askew from perpendicularity to the other mirrors. 

In this position, because the light entering the CCR

does not return along the same entry path, little light

returns to the source—a digital 0. Applying voltage

between this mirror and an electrode beneath it causes

the mirror to shift to a position perpendicular to other

mirrors, thus causing the light entering the CCR to return

to its source—a digital 1. The mirror’s low mass allows

the CCR to switch between these two states up to a thou-

sand times per second, using less than a nanojoule per 

0 → 1 transition. A 1 → 0 transition, on the other hand,

is practically free because dumping the charge stored on

the electrode to the ground requires almost no energy.

Our latest Smart Dust device is a 63-mm

3

autonomous bidirectional communication mote that

receives an optical signal, generates a pseudorandom

sequence based on this signal to emulate sensor data,

and then optically transmits the result. The system

contains a micromachined corner-cube reflector, a

0.078-mm

3

complementary metal oxide semiconduc-

tor (CMOS) chip that draws 50 microwatts, and a

hearing aid battery. In addition to a battery-based

operation, we have also powered the device using a 

2-mm

2

solar cell. This mote demonstrates Smart

Dust’s essential concepts, such as optical data trans-

mission, data processing, energy management, minia-

turization, and system integration.

A passive communication system suffers several lim-

itations. Unable to communicate with one another,

motes rely on a central station equipped with a light

source to send and receive data from other motes. If a

given mote does not have a clear line of sight to the

central station, that mote will be isolated from the net-

work. Also, because the CCR reflects only a small frac-

tion of the light emitted from the base station, this

system’s range cannot easily extend beyond 1 kilome-

ter. To circumvent these limitations, dust motes must

be active and have their own onboard light source.

Active-steered laser systems

For mote-to-mote communication, an active-steered

laser communication system uses an onboard light

source to send a tightly collimated light beam toward

an intended receiver. Steered laser communication has

the advantage of high power density; for example, a

1-milliwatt laser radiating into 1 milliradian (3.4 arc-

seconds) has a density of approximately 318 kilowatts

per steradian (there are 4 π steradians in a sphere), as

opposed to a 100-watt lightbulb that radiates 8 watts

per steradian isotropically. A Smart Dust mote’s emit-

ted beam would have a divergence of approximately

1 milliradian, permitting communication over enor-

mous distances using milliwatts of power. 

Forming ad hoc multihop networks is the most excit-

ing application of mote-to-mote communication.

Multihop networks present significant challenges to

current network algorithms—routing software must

not only optimize each packet’s latency but also con-

sider both the transmitter’s and receiver’s energy

reserves. Each mote must carefully weigh the needs to

sense, compute, communicate, and evaluate its energy

reserve status before allocating precious nanojoules of

energy to turn on its transmitter or receiver. Because

these motes spend most of their time sleeping, with their

receivers turned off, scheduling a common awake time

across the network is difficult. If motes don’t wake up

in a synchronized manner, a highly dynamic network

topology and large packet latency result. Using burst-

mode communication, in which the laser operates at

up to several tens of megabits per second for a few mil-

liseconds, provides the most energy-efficient way to

schedule this network. This procedure minimizes the

mote’s duty cycle and better utilizes its energy reserves.

The steered agile laser transmitter consists of a semi-

conductor diode laser coupled with a collimating lens

and MEMS beam-steering optics based on a two-

degree-of-freedom silicon micromirror, as Figures 4

and 5 show. This system integrates all optical com-

ponents into an active 8-mm

3 

volume.

Figure 3. Autonomous bidirectional communication mote with

a MEMS optics chip containing a corner-cube retroreflector on

the large die, a CMOS application-specific integrated circuit

(ASIC) for control on the 300 

×

360 micron die, and a hearing

aid battery for power. The total volume is 63 mm

3

.

LISTENING TO A DUST FIELD 

Many Smart Dust applications rely on direct opti-

cal communication from an entire field of dust motes

to one or more base stations. These base stations must

therefore be able to receive a large volume of simul-

taneous optical transmissions. Further, communica-

tion must be possible outdoors in bright sunlight

which has an intensity of approximately 1 kilowatt

per square meter, although the dust motes each trans-

mit information with a few milliwatts of power. Using

a narrow-band optical filter to eliminate all sunlight

except the portion near the light frequency used for

communication can partially solve this second prob-

lem, but the ambient optical power often remains

much stronger than the received signal power.

Advantages of imaging receivers

As with the transmitter, the short wavelength of opti-

cal transmissions compared with radio frequencies

overcomes both challenges. Light from a large field of

view can be focused into an image, as in our eyes or in

a camera. Imaging receivers utilize this to analyze dif-

ferent portions of the image separately to process simul-

taneous transmissions from different angles. This

method of distinguishing transmissions based on their

originating location is referred to as space division mul-

tiple access (SDMA). In contrast, most radio-frequency

antennas receive all incident radio power in a single 

signal, which requires using additional tactics, such as

frequency tuning or code division multiple access

(CDMA), to separate simultaneous transmissions.

Imaging receivers also offer the advantage of dra-

matically decreasing the ratio of ambient optical

power to received signal power.

11

Ideally, the imaging

receiver will focus all of the received power from a

single transmission onto a single photodetector. If the

receiver has an n ×n array of pixels, then the ambient

light that each pixel receives is reduced by a factor n

2

compared with a nonimaging receiver. Typically, using

a value for n between 8 and 32 makes the ambient

light power negligible compared with the electronic

noise in the analog electronics.

Video camera. A video camera is a straightforward

implementation of an imaging receiver. If each mem-

ber in a colony of Smart Dust motes flashes its own

signal at a rate of a few bits per second, then each

transmitter will appear in the video stream at a dif-

ferent location in the image. We have implemented

such a system using a laptop with a frame grabber

that processes a real-time video signal in software. 

We tested this system to transmit weather infor-

mation from Twin Peaks in San Francisco to a video

camera in Berkeley, 21.4 kilometers across San

Francisco Bay. The transmitter consists of a one-

cubic-inch Smart Dust mote mock-up that modulates

an ordinary red laser pointer at a few bits per second,

using only 3.5 milliwatts of peak optical transmis-

sion power in a 2-milliradian cone.

12

Smart Dust’s

improved system will allow similar link distances,

with the added advantage of automated transmitter-

receiver alignment. 

Using a high-speed camera and a dedicated digital

signal processor to process the video signal achieves

higher data rates. With modern cameras and DSPs, pro-

cessing video at about 1,000 frames per second should

be feasible. This would allow communication at a few

January 2001 49

Figure 4. Conceptual diagram of steered agile laser transmit-

ter (side view). A laser emits an infrared beam that is colli-

mated with a lens. The lens directs the narrow laser beam

onto a beam-steering mirror, aiming the beam toward the

intended receiver.

Figure 5. Scanning electron micrograph of the first-genera-

tion steered agile laser transmitter. The chip combines a

laser diode and ball lens with a micromachined two-degree-

of-freedom beam-steering mirror. The optical path runs from

the top of the laser diode’s front facet, through the ball lens,

reflects off the left-hand mirror plate, then finally reflects off

the substrate before leaving the chip.

50 Computer

hundred bits per second, which is acceptable for many

applications. An alternative receiver architecture pro-

vides a more elegant solution at much higher data rates,

avoiding the need for computationally intensive video

processing and very high-speed cameras.

“Smart pixel

”

integrated imaging receivers. We are

developing a fully integrated CMOS imaging device

that receives data at up to a few megabits per second.

As Figure 6 shows, this receiver leverages the power of

shrinking integrated circuits and recent developments

in CMOS “smart pixel” sensors to create a microchip

similar to a digital camera sensor, but with a complete

asynchronous receiver circuit integrated into every

pixel in the imaging array.

During the receiver’s operation, each pixel

autonomously monitors its own signal, looks for a

transmission, decodes it, and transmits the data off

chip when it receives a valid data packet. As Figure 7

shows, to function in this way, each pixel in the

imager requires a photosensor and circuits to perform

analog signal processing and amplification, analog-

to-digital conversion, and an asynchronous serial

receiver. Such a receiver should be able to receive

transmissions of only a few milliwatts in strength at

up to a few megabytes per second over a distance of

several kilometers. Using larger collection lenses that

measure more than 10 centimeters and high-resolu-

tion arrays—64 ×64 or greater—to divide unwanted

background light permits communication with low-

earth-orbit satellites.

We currently have a working pixel front end that

consumes 10 picojoules per bit per pixel. It can detect

100 femtojoules per bit of optical power, but that will

soon be down to 10 femtojoules per bit. This trans-

lates into the need for just 10 picojoules per bit of trans-

mit power for short-range, indoor communication.

Integrating an imaging receiver onto a single micro-

chip imposes severe constraints in silicon area and

power consumption per pixel. Only recently have con-

5 mW, 1 mrad, 1 Mbps

1-10 km

Field of view

of single pixel

Field of view of imager

10 mm

5 mm

Optical

filter

Collection

lens

Imaging

receiver

Figure 6. Pictorial representation of an integrated imaging receiver in action with predicted specifications. The receiver lever-

ages the power of integrated circuits and CMOS imaging sensors to create a microchip with a complete asynchronous receiver

circuit integrated into every pixel in the imaging array. Each pixel autonomously monitors its own signal as it searches for a

transmission, decodes it, and transmits the data off chip when it receives a data packet.

Signal processing,

analog-to-digital

conversion

Off-chip

bus

driver

Photosensor

Serial

decoder

CRC check

local bus driver

Pixel array

Figure 7. Typical components within each pixel of a simple integrated imaging

receiver. Each of the imager’s pixels contains a photosensor and circuits to perform

analog signal processing and amplification, analog-to-digital conversion, and an asyn-

chronous serial receiver. Cyclic redundancy check (CRC) is a data-coding algorithm that

allows a receiver to detect whether a received data packet is valid or corrupt.

tinuing reductions in transistor size allowed for suffi-

cient reductions in circuit area and power consump-

tion to achieve this level of integration. 

R

esearch in the wireless sensor network area is

growing rapidly in both academia and industry.

Most major universities and many companies

now have sensor networking projects, and some prod-

ucts are appearing on the market. Innovative research

includes short-range micropower radio, energy scav-

enging from thermal gradients and vibration, operat-

ing systems, networking and signal-processing algo-

rithms, and applications. While the raw power of

future computing environments will enable more mas-

sive and amazing hardware and software networks, a

growing community will be pushing the limits on the

lower end, building smaller hardware and writing

terser code. We will program the walls and the furni-

ture, and some day even the insects and the dust. ✸

Acknowledgments

Our work is funded by the Defense Advanced

Research Projects Agency’s Microsystems Technology

Office, the Howard Hughes Doctoral Fellowship

Program, and the Fannie and John Hertz Foundation.

References

1. M. Weiser and J.S. Brown, “The Coming Age of Calm

Technology,” http://www.ubiq.com/hypertext/weiser/

acmfuture2endnote.htm, 1996.

2. J. Kahn, R.H. Katz, and K. Pister, “Emerging Challenges:

Mobile Networking for ‘Smart Dust,’” J. Comm. Net-

works, Sept. 2000, pp. 188-196. 

3. C. Piguet et al., “Low-Power Design of 8-b Embedded

CoolRisc Microcontroller Cores,” IEEE J. Solid-State

Circuits, July 1997, pp. 1067-1078.

4. J. Hill et al., “System Architecture Directions for Net-

worked Sensors,” Proc. 9th Int’l Conf. Architectural

Support for Programming Languages and Operating

Systems (ASPLOS-IX), ACM Press, New York, 2000,

pp. 93-104.

5. J.L. Hennessy and D.A. Patterson, Computer Architec-

ture: A Quantitative Approach, 2nd ed., Morgan Kauf-

mann, Menlo Park, Calif., 1996.

6. J.D. Meindl, “Low-Power Microelectronics: Retrospect

and Prospect,” Proc. IEEE, Apr. 1995, pp. 619-635.

7. C. Piguet, “Low-Power and Low-Voltage CMOS Digi-

tal Design,” Microelectronic Eng., Elsevier Science BV,

The Netherlands, 1997, pp. 179-208.

8. G. Asada et al., “Wireless Integrated Network Sensors:

Low-Power Systems on a Chip,” Proc. 24th European

Solid-State Circuits Conf. (ESSCIRC 98), Editions Fron-

tieres, Paris, 1998, pp. 9-16.

9. R. Amirtharajah et al., “A Micropower Programmable DSP

Powered Using a MEMS-Based Vibration-to-Electric Energy

Converter,” 2000 IEEE Int’l Solid-State Circuits Conf., IEEE

Press, Piscataway, N.J., 2000, pp. 362-363, 469.

10. P.B. Chu et al., “Optical Communication Using Micro

Corner Cube Reflectors,” Proc. 10th IEEE Int’l Work-

shop on Micro Electro Mechanical Systems, IEEE Press,

Piscataway, N.J., 1997, pp. 350-355.

11. J.M. Kahn et al., “Imaging Diversity Receivers for High-

Speed Infrared Wireless Communication,” IEEE

Comm., Dec. 1998, pp. 88-94.

12. M. Last and K.S.J. Pister, “2-DOF Actuated Micromir-

ror Designed for Large DC Deflection,” Proc. 3rd Int’l

Conf. Micro Opto Electro Mechanical Systems, IEEE

Press, Piscataway, N.J., 1999, pp. 239-243.

Brett Warneke is a PhD candidate at the Berkeley Sen-

sor and Actuator Center, University of California,

Berkeley. His research interests include ultra-low-

power integrated circuits, CMOS micromachining,

distributed sensor networks, and RF MEMS. He

received an MS in electrical engineering from the Uni-

versity of California, Los Angeles. Contact him at

warneke@eecs.berkeley.edu.

Matt Last is a graduate student researcher at the

Berkeley Sensor and Actuator Center, University of

California, Berkeley. His research interests include

optical MEMS. He has an MS in electrical engineer-

ing from the University of California, Berkeley. Con-

tact him at mattlast@eecs.berkeley.edu.

Brian Liebowitz is a graduate student researcher at

the Berkeley Sensor and Actuator Center, University

of California, Berkeley. His research interests include

low-power analog integrated circuits, optical com-

munication, and MEMS. He received a BS in electri-

cal engineering from Columbia University. Contact

him at leibowitz@ieee.org.

Kristofer S.J. Pister is an associate professor of elec-

trical engineering and computer sciences at the Berke-

ley Sensor and Actuator Center, University of Cali-

fornia, Berkeley. His research interests include the

development and use of standard MEMS fabrication

technologies, microrobotics, and CAD for MEMS. He

received a PhD in electrical engineering from the 

University of California, Berkeley. Contact him at

pister@eecs.berkeley.edu