ekhonet: high speed ultra low-power backscatter for next generation sensors · 2014-09-11 ·...
TRANSCRIPT
EkhoNet: High Speed Ultra Low-power Backscatter for Next Generation Sensors
Pengyu Zhang, Pan Hu, Vijay Pasikanti, Deepak Ganesan !!
School of Computer Science University of Massachusetts Amherst
Frontier of Wireless Sensing
Google Glass EEG sensorThermostat Humidity
richer content
higher bandwidth
Frontier of Wireless Sensing
Google Glass EEG sensorThermostat Humidity
- BLE / ZigBee - Heavy duty-cycling
richer content
higher bandwidth
Frontier of Wireless Sensing
Google Glass EEG sensorThermostat Humidity
richer content
higher bandwidth
- Duty-cycling? - Sacrifice lifetime - Local processing
- BLE / ZigBee - Heavy duty-cycling
Frontier of Wireless Sensing
Google Glass EEG sensorThermostat Humidity
richer content
higher bandwidth
Face Social context Gesture Speech
Raw data!
- Duty-cycling? - Sacrifice lifetime - Local processing
- BLE / ZigBee - Heavy duty-cycling
Can we achieve ~1 Mbps wireless transmission while consuming 10~100 micro-watts?
Backscatter - Ultra Low-power Communication Primitive
Backscatter reader Sensor
Backscatter communication
Backscatter is extremely efficient because the device is reflecting the signal rather than generating a signal.
In principle, backscatter can operate at micro-watts while transmitting at megabits/second. What is the reality?
Transistor
Existing Backscatter System - NOT Low Power
microphone - 704kbps
~ 30uW
> 3mW
Backscatter is NOT low power! Why?
accelerometer - 1kbps
34uW
3uW Intel WISP UMass Moo
Canonical Wireless Sensing Pipeline
Sensor Interface Processing Network
StackRadio
Control
Computational blocks
Sensor Interface Processing Network
StackRadio
Control
Computational blocks
Sensing Pipeline of Backscatter Systems
A"
B"
Cin"
S"
Cout"
(a) 1 bit adder circuits for computation.
D"
E"
(a) 1 bit adder circuits for computation.
(b) 1 bit shift register circuits forbackscatter.
Figure 2: 1 bit adder and 1 bit shift register circuit.
source prior to communication. Indeed, this tradeo↵ hasbeen reinforced by performance/power trends over the pastdecade — power consumption of embedded processors havedropped dramatically, while power reduction in active radioshas been relatively slower.
QQHowever, backscatter communication challenges this long-
held view. Backscatter is inherently extraordinarily e�cientsince the carrier wave is generated by the reader, and the tagonly backscatters the signal without any additional ampli-fication. Thus, each bit of backscatter is extremely simple,and only requires a handful of gates (Figure 2). This impliesthat for computation to be cheaper energy-wise, the compu-tational operations on each bit would have to use fewer gatesthan that required to communicate the bit. This is often atall order due to the simplicity of backscatter.
Consider, for example, a simple aggregation operationthat sums ten sensor readings before transmitting the aggre-gate value over the radio. On traditional sensor platforms,such data reduction would have direct and significant powerbenefits since communication dominates power, and our ag-gregation scheme cuts this cost by a factor of ten. The sameoperation on a backscatter-based platform has dubious ben-efits. Figure 2 shows that the number of gates required forsumming two bits is roughly nine, but only four gates areneeded to transmit the same data via the shift register cir-cuit of backscatter! As power consumption is proportionalto number of gates, a nine gate adder consumes 2.2⇥ morepower than the a four gate backscatter circuit.
It is necessary to add a few caveats to our simplified com-parison of computation and communication. The clock ratesof communication subsystems are limited by signal to noiseratio considerations, whereas the clock rates of processorscan be higher, and thereby reduce power. In addition, low-power processors use many tricks to reduce power consump-tion including optimized signal processing circuits, di↵er-ent power domains, extremely tight duty-cycling, and so on.Despite these optimizations, the cards are stacked againstcomputation. Backscatter is so incredibly simple in termsof circuitry that even matching the e�ciency of backscatterbecomes a challenging architectural design problem.
Thus, the crux of our argument is the following: backscat-
ter drives down the optimal cross-over point between compu-
tation vs communication, such that communication of raw
data may be preferable to computation in a wider spectrum
of real-world scenarios.
Implications on architecture design: This observa-tion has an immediate implication on the architecture of abackscatter-based sensor platform. Traditional sensing plat-forms add a lot of computational modules between the sensorand the radio for sensor data acquisition, processing, filter-ing, bu↵ering, etc. The contribution of these components tooverall power consumption of an active radio-based sensorsystem is minimal and can largely be ignored. However, onbackscatter-based platforms, these components become thebottleneck.
This raises an intriguing question — with the power con-sumption of backscatter being so low, would it in fact bemore e�cient to eliminate all of these modules en-bloc, andjust connect the sensor directly to the radio? In other words,would it be better to just stream every bit of data that issensed directly through the radio?
We take a measurement-driven approach towards answer-ing these questions. First, we look at the computationalblocks between sensing and the RF interface on existingbackscatter-based sensing platforms to understand how muchpower they consume, as well as why they su↵er in termsof throughput. Second, we build on our empirical studyand design a radically new backscatter-based sensor plat-form that addresses these limitations.
3. INVESTIGATING EXISTING WIRELESSSENSING ARCHITECTURES
In this section, we investigate why current backscatter-based platforms are unable to achieve end-to-end power con-sumption of µWs for high-rate sensing and transfer. We alsoinvestigate why they are unable to achieve high-data ratecommunication, particularly while operating at low power.To empirically understand these factors, we look at the UMassMoo/UW WISP class platforms that are equipped with sen-sors, a low-power MCU (MSP 430 family) and a backscatterradio.
3.1 Poor energy efficiencyWe start with a break down of the power consumed by
three key computational modules on a UMass Moo (Fig-ure 3): 1) the sensor data acquisition subsystem which han-dles the protocols for operating sensors, 2) the data handlingsubsystem on a micro-controller where sensor data is stored,processed (if needed), formatted into packet, and sent to thenetwork stack, and 3) the network stack implemented in acombination of hardware and software.
3.1.1 Sensor data acquisitionSensor data acquisition is a relatively simple operation —
some sensors have an on-board ADC, hence data acquisitionis via a protocol such as SPI or I2C, whereas other sensorsjust provide an analog signal which is digitized using themicro-controller’s ADC. Despite its simplicity, even theseoperations are not as cheap as one might expect. For exam-ple, sampling an accelerometer via the SPI bus would requireperiodic wakeup of the MCU to fill the SPI bu↵er, sendingthe read command and read address to the sensor, as well as
(a) 1 bit adder circuits for computation.
(b) 1 bit shift register circuits forbackscatter.
Figure 2: 1 bit adder and 1 bit shift register circuit.
source prior to communication. Indeed, this tradeo↵ hasbeen reinforced by performance/power trends over the pastdecade — power consumption of embedded processors havedropped dramatically, while power reduction in active radioshas been relatively slower.
QQHowever, backscatter communication challenges this long-
held view. Backscatter is inherently extraordinarily e�cientsince the carrier wave is generated by the reader, and the tagonly backscatters the signal without any additional ampli-fication. Thus, each bit of backscatter is extremely simple,and only requires a handful of gates (Figure 2). This impliesthat for computation to be cheaper energy-wise, the compu-tational operations on each bit would have to use fewer gatesthan that required to communicate the bit. This is often atall order due to the simplicity of backscatter.
Consider, for example, a simple aggregation operationthat sums ten sensor readings before transmitting the aggre-gate value over the radio. On traditional sensor platforms,such data reduction would have direct and significant powerbenefits since communication dominates power, and our ag-gregation scheme cuts this cost by a factor of ten. The sameoperation on a backscatter-based platform has dubious ben-efits. Figure 2 shows that the number of gates required forsumming two bits is roughly nine, but only four gates areneeded to transmit the same data via the shift register cir-cuit of backscatter! As power consumption is proportionalto number of gates, a nine gate adder consumes 2.2⇥ morepower than the a four gate backscatter circuit.
It is necessary to add a few caveats to our simplified com-parison of computation and communication. The clock ratesof communication subsystems are limited by signal to noiseratio considerations, whereas the clock rates of processorscan be higher, and thereby reduce power. In addition, low-power processors use many tricks to reduce power consump-tion including optimized signal processing circuits, di↵er-ent power domains, extremely tight duty-cycling, and so on.Despite these optimizations, the cards are stacked againstcomputation. Backscatter is so incredibly simple in termsof circuitry that even matching the e�ciency of backscatterbecomes a challenging architectural design problem.
Thus, the crux of our argument is the following: backscat-
ter drives down the optimal cross-over point between compu-
tation vs communication, such that communication of raw
data may be preferable to computation in a wider spectrum
of real-world scenarios.
Implications on architecture design: This observa-tion has an immediate implication on the architecture of abackscatter-based sensor platform. Traditional sensing plat-forms add a lot of computational modules between the sensorand the radio for sensor data acquisition, processing, filter-ing, bu↵ering, etc. The contribution of these components tooverall power consumption of an active radio-based sensorsystem is minimal and can largely be ignored. However, onbackscatter-based platforms, these components become thebottleneck.
This raises an intriguing question — with the power con-sumption of backscatter being so low, would it in fact bemore e�cient to eliminate all of these modules en-bloc, andjust connect the sensor directly to the radio? In other words,would it be better to just stream every bit of data that issensed directly through the radio?
We take a measurement-driven approach towards answer-ing these questions. First, we look at the computationalblocks between sensing and the RF interface on existingbackscatter-based sensing platforms to understand how muchpower they consume, as well as why they su↵er in termsof throughput. Second, we build on our empirical studyand design a radically new backscatter-based sensor plat-form that addresses these limitations.
3. INVESTIGATING EXISTING WIRELESSSENSING ARCHITECTURES
In this section, we investigate why current backscatter-based platforms are unable to achieve end-to-end power con-sumption of µWs for high-rate sensing and transfer. We alsoinvestigate why they are unable to achieve high-data ratecommunication, particularly while operating at low power.To empirically understand these factors, we look at the UMassMoo/UW WISP class platforms that are equipped with sen-sors, a low-power MCU (MSP 430 family) and a backscatterradio.
3.1 Poor energy efficiencyWe start with a break down of the power consumed by
three key computational modules on a UMass Moo (Fig-ure 3): 1) the sensor data acquisition subsystem which han-dles the protocols for operating sensors, 2) the data handlingsubsystem on a micro-controller where sensor data is stored,processed (if needed), formatted into packet, and sent to thenetwork stack, and 3) the network stack implemented in acombination of hardware and software.
3.1.1 Sensor data acquisitionSensor data acquisition is a relatively simple operation —
some sensors have an on-board ADC, hence data acquisitionis via a protocol such as SPI or I2C, whereas other sensorsjust provide an analog signal which is digitized using themicro-controller’s ADC. Despite its simplicity, even theseoperations are not as cheap as one might expect. For exam-ple, sampling an accelerometer via the SPI bus would requireperiodic wakeup of the MCU to fill the SPI bu↵er, sendingthe read command and read address to the sensor, as well as
1"bit"shi*"register" Backsca2er"RF"
(b) 1 bit shift register controls a backscatter tran-sistor.
Figure 2: 1 bit adder and 1 bit shift register circuit.
dropped dramatically, while power reduction in active radioshas been relatively slower.
However, backscatter communication challenges this long-held view. Backscatter is inherently extraordinarily e�cientsince the carrier wave is generated by the reader, and the tagonly backscatters the signal without any additional ampli-fication. Thus, each bit of backscatter is extremely simple,and only requires a handful of gates (Figure 2). This impliesthat for computation to be cheaper energy-wise, the compu-tational operations on each bit would have to use fewer gatesthan that required to communicate the bit. This is often atall order due to the simplicity of backscatter.
Consider, for example, a simple aggregation operationthat sums ten sensor readings before transmitting the aggre-gate value over the radio. On traditional sensor platforms,such data reduction would have direct and significant powerbenefits since communication dominates power, and our ag-gregation scheme cuts this cost by a factor of ten. Thesame operation on a backscatter-based platform has dubi-ous benefits. Figure 2 shows that the number of NANDgates required for summing two bits is roughly nine (thirtysix transistors), but only four NAND gates (sixteen tran-sistors) and an additional transistor for backscattering thesignal are needed to transmit the same data via the shift-register controlled backscatter RF! As power consumptionis proportional to number of transistors, a nine gate adderconsumes 2.1⇥ more power than the shift-register controlledbackscatter RF.
It is necessary to add a few caveats to our simplified com-parison of computation and communication. The clock ratesof communication subsystems are limited by signal to noiseratio considerations, whereas the clock rates of processorscan be higher, and thereby reduce power. In addition, low-power processors use many tricks to reduce power consump-tion including optimized signal processing circuits, di↵er-ent power domains, extremely tight duty-cycling, and so on.Despite these optimizations, the cards are stacked againstcomputation. Backscatter is so incredibly simple in terms
of circuitry that even matching the e�ciency of backscatterbecomes a challenging architectural design problem.
Thus, the crux of our argument is the following: backscat-
ter drives down the optimal cross-over point between compu-
tation vs communication, such that communication of raw
data may be preferable to computation in a wider spectrum
of real-world scenarios.
Implications on architecture design: This observa-tion has an immediate implication on the architecture of abackscatter-based sensor platform. Traditional sensing plat-forms add a lot of computational modules between the sensorand the radio for sensor data acquisition, processing, filter-ing, bu↵ering, etc. The contribution of these components tooverall power consumption of an active radio-based sensorsystem is minimal and can largely be ignored. However, onbackscatter-based platforms, these components become thebottleneck.
This raises an intriguing question — with the power con-sumption of backscatter being so low, would it in fact bemore e�cient to eliminate all of these modules en-bloc, andjust connect the sensor directly to the radio? In other words,would it be better to just stream every bit of data that issensed directly through the radio?
We take a measurement-driven approach towards answer-ing these questions. First, we look at the computationalblocks between sensing and the RF interface on existingbackscatter-based sensing platforms to understand how muchpower they consume, as well as why they su↵er in termsof throughput. Second, we build on our empirical studyand design a radically new backscatter-based sensor plat-form that addresses these limitations.
3. INVESTIGATING EXISTING WIRELESSSENSING ARCHITECTURES
In this section, we investigate why current backscatter-based platforms are unable to achieve end-to-end power con-sumption of µWs for high-rate sensing and transfer. We alsoinvestigate why they are unable to achieve high-data ratecommunication, particularly while operating at low power.To empirically understand these factors, we look at the UMassMoo/UW WISP class platforms that are equipped with sen-sors, a low-power MCU (MSP 430 family) and a backscatterradio.
3.1 Poor energy efficiencyWe start with a break down of the power consumed by
three key computational modules on a UMass Moo (Fig-ure 3): 1) the sensor data acquisition subsystem which han-dles the protocols for operating sensors, 2) the data handlingsubsystem on a micro-controller where sensor data is stored,processed (if needed), formatted into packet, and sent to thenetwork stack, and 3) the network stack implemented in acombination of hardware and software.
3.1.1 Sensor data acquisitionSensor data acquisition is a relatively simple operation —
some sensors have an on-board ADC, hence data acquisitionis via a protocol such as SPI or I2C, whereas other sensorsjust provide an analog signal which is digitized using themicro-controller’s ADC. Despite its simplicity, even theseoperations are not as cheap as one might expect. For exam-ple, sampling an accelerometer via the SPI bus would require
~3.7uW
~34uW
Sensor Interface Processing Network
StackRadio
Control
Computational blocks
Sensing Pipeline of Backscatter Systems
A"
B"
Cin"
S"
Cout"
(a) 1 bit adder circuits for computation.
D"
E"
(a) 1 bit adder circuits for computation.
(b) 1 bit shift register circuits forbackscatter.
Figure 2: 1 bit adder and 1 bit shift register circuit.
source prior to communication. Indeed, this tradeo↵ hasbeen reinforced by performance/power trends over the pastdecade — power consumption of embedded processors havedropped dramatically, while power reduction in active radioshas been relatively slower.
QQHowever, backscatter communication challenges this long-
held view. Backscatter is inherently extraordinarily e�cientsince the carrier wave is generated by the reader, and the tagonly backscatters the signal without any additional ampli-fication. Thus, each bit of backscatter is extremely simple,and only requires a handful of gates (Figure 2). This impliesthat for computation to be cheaper energy-wise, the compu-tational operations on each bit would have to use fewer gatesthan that required to communicate the bit. This is often atall order due to the simplicity of backscatter.
Consider, for example, a simple aggregation operationthat sums ten sensor readings before transmitting the aggre-gate value over the radio. On traditional sensor platforms,such data reduction would have direct and significant powerbenefits since communication dominates power, and our ag-gregation scheme cuts this cost by a factor of ten. The sameoperation on a backscatter-based platform has dubious ben-efits. Figure 2 shows that the number of gates required forsumming two bits is roughly nine, but only four gates areneeded to transmit the same data via the shift register cir-cuit of backscatter! As power consumption is proportionalto number of gates, a nine gate adder consumes 2.2⇥ morepower than the a four gate backscatter circuit.
It is necessary to add a few caveats to our simplified com-parison of computation and communication. The clock ratesof communication subsystems are limited by signal to noiseratio considerations, whereas the clock rates of processorscan be higher, and thereby reduce power. In addition, low-power processors use many tricks to reduce power consump-tion including optimized signal processing circuits, di↵er-ent power domains, extremely tight duty-cycling, and so on.Despite these optimizations, the cards are stacked againstcomputation. Backscatter is so incredibly simple in termsof circuitry that even matching the e�ciency of backscatterbecomes a challenging architectural design problem.
Thus, the crux of our argument is the following: backscat-
ter drives down the optimal cross-over point between compu-
tation vs communication, such that communication of raw
data may be preferable to computation in a wider spectrum
of real-world scenarios.
Implications on architecture design: This observa-tion has an immediate implication on the architecture of abackscatter-based sensor platform. Traditional sensing plat-forms add a lot of computational modules between the sensorand the radio for sensor data acquisition, processing, filter-ing, bu↵ering, etc. The contribution of these components tooverall power consumption of an active radio-based sensorsystem is minimal and can largely be ignored. However, onbackscatter-based platforms, these components become thebottleneck.
This raises an intriguing question — with the power con-sumption of backscatter being so low, would it in fact bemore e�cient to eliminate all of these modules en-bloc, andjust connect the sensor directly to the radio? In other words,would it be better to just stream every bit of data that issensed directly through the radio?
We take a measurement-driven approach towards answer-ing these questions. First, we look at the computationalblocks between sensing and the RF interface on existingbackscatter-based sensing platforms to understand how muchpower they consume, as well as why they su↵er in termsof throughput. Second, we build on our empirical studyand design a radically new backscatter-based sensor plat-form that addresses these limitations.
3. INVESTIGATING EXISTING WIRELESSSENSING ARCHITECTURES
In this section, we investigate why current backscatter-based platforms are unable to achieve end-to-end power con-sumption of µWs for high-rate sensing and transfer. We alsoinvestigate why they are unable to achieve high-data ratecommunication, particularly while operating at low power.To empirically understand these factors, we look at the UMassMoo/UW WISP class platforms that are equipped with sen-sors, a low-power MCU (MSP 430 family) and a backscatterradio.
3.1 Poor energy efficiencyWe start with a break down of the power consumed by
three key computational modules on a UMass Moo (Fig-ure 3): 1) the sensor data acquisition subsystem which han-dles the protocols for operating sensors, 2) the data handlingsubsystem on a micro-controller where sensor data is stored,processed (if needed), formatted into packet, and sent to thenetwork stack, and 3) the network stack implemented in acombination of hardware and software.
3.1.1 Sensor data acquisitionSensor data acquisition is a relatively simple operation —
some sensors have an on-board ADC, hence data acquisitionis via a protocol such as SPI or I2C, whereas other sensorsjust provide an analog signal which is digitized using themicro-controller’s ADC. Despite its simplicity, even theseoperations are not as cheap as one might expect. For exam-ple, sampling an accelerometer via the SPI bus would requireperiodic wakeup of the MCU to fill the SPI bu↵er, sendingthe read command and read address to the sensor, as well as
(a) 1 bit adder circuits for computation.
(b) 1 bit shift register circuits forbackscatter.
Figure 2: 1 bit adder and 1 bit shift register circuit.
source prior to communication. Indeed, this tradeo↵ hasbeen reinforced by performance/power trends over the pastdecade — power consumption of embedded processors havedropped dramatically, while power reduction in active radioshas been relatively slower.
QQHowever, backscatter communication challenges this long-
held view. Backscatter is inherently extraordinarily e�cientsince the carrier wave is generated by the reader, and the tagonly backscatters the signal without any additional ampli-fication. Thus, each bit of backscatter is extremely simple,and only requires a handful of gates (Figure 2). This impliesthat for computation to be cheaper energy-wise, the compu-tational operations on each bit would have to use fewer gatesthan that required to communicate the bit. This is often atall order due to the simplicity of backscatter.
Consider, for example, a simple aggregation operationthat sums ten sensor readings before transmitting the aggre-gate value over the radio. On traditional sensor platforms,such data reduction would have direct and significant powerbenefits since communication dominates power, and our ag-gregation scheme cuts this cost by a factor of ten. The sameoperation on a backscatter-based platform has dubious ben-efits. Figure 2 shows that the number of gates required forsumming two bits is roughly nine, but only four gates areneeded to transmit the same data via the shift register cir-cuit of backscatter! As power consumption is proportionalto number of gates, a nine gate adder consumes 2.2⇥ morepower than the a four gate backscatter circuit.
It is necessary to add a few caveats to our simplified com-parison of computation and communication. The clock ratesof communication subsystems are limited by signal to noiseratio considerations, whereas the clock rates of processorscan be higher, and thereby reduce power. In addition, low-power processors use many tricks to reduce power consump-tion including optimized signal processing circuits, di↵er-ent power domains, extremely tight duty-cycling, and so on.Despite these optimizations, the cards are stacked againstcomputation. Backscatter is so incredibly simple in termsof circuitry that even matching the e�ciency of backscatterbecomes a challenging architectural design problem.
Thus, the crux of our argument is the following: backscat-
ter drives down the optimal cross-over point between compu-
tation vs communication, such that communication of raw
data may be preferable to computation in a wider spectrum
of real-world scenarios.
Implications on architecture design: This observa-tion has an immediate implication on the architecture of abackscatter-based sensor platform. Traditional sensing plat-forms add a lot of computational modules between the sensorand the radio for sensor data acquisition, processing, filter-ing, bu↵ering, etc. The contribution of these components tooverall power consumption of an active radio-based sensorsystem is minimal and can largely be ignored. However, onbackscatter-based platforms, these components become thebottleneck.
This raises an intriguing question — with the power con-sumption of backscatter being so low, would it in fact bemore e�cient to eliminate all of these modules en-bloc, andjust connect the sensor directly to the radio? In other words,would it be better to just stream every bit of data that issensed directly through the radio?
We take a measurement-driven approach towards answer-ing these questions. First, we look at the computationalblocks between sensing and the RF interface on existingbackscatter-based sensing platforms to understand how muchpower they consume, as well as why they su↵er in termsof throughput. Second, we build on our empirical studyand design a radically new backscatter-based sensor plat-form that addresses these limitations.
3. INVESTIGATING EXISTING WIRELESSSENSING ARCHITECTURES
In this section, we investigate why current backscatter-based platforms are unable to achieve end-to-end power con-sumption of µWs for high-rate sensing and transfer. We alsoinvestigate why they are unable to achieve high-data ratecommunication, particularly while operating at low power.To empirically understand these factors, we look at the UMassMoo/UW WISP class platforms that are equipped with sen-sors, a low-power MCU (MSP 430 family) and a backscatterradio.
3.1 Poor energy efficiencyWe start with a break down of the power consumed by
three key computational modules on a UMass Moo (Fig-ure 3): 1) the sensor data acquisition subsystem which han-dles the protocols for operating sensors, 2) the data handlingsubsystem on a micro-controller where sensor data is stored,processed (if needed), formatted into packet, and sent to thenetwork stack, and 3) the network stack implemented in acombination of hardware and software.
3.1.1 Sensor data acquisitionSensor data acquisition is a relatively simple operation —
some sensors have an on-board ADC, hence data acquisitionis via a protocol such as SPI or I2C, whereas other sensorsjust provide an analog signal which is digitized using themicro-controller’s ADC. Despite its simplicity, even theseoperations are not as cheap as one might expect. For exam-ple, sampling an accelerometer via the SPI bus would requireperiodic wakeup of the MCU to fill the SPI bu↵er, sendingthe read command and read address to the sensor, as well as
1"bit"shi*"register" Backsca2er"RF"
(b) 1 bit shift register controls a backscatter tran-sistor.
Figure 2: 1 bit adder and 1 bit shift register circuit.
dropped dramatically, while power reduction in active radioshas been relatively slower.
However, backscatter communication challenges this long-held view. Backscatter is inherently extraordinarily e�cientsince the carrier wave is generated by the reader, and the tagonly backscatters the signal without any additional ampli-fication. Thus, each bit of backscatter is extremely simple,and only requires a handful of gates (Figure 2). This impliesthat for computation to be cheaper energy-wise, the compu-tational operations on each bit would have to use fewer gatesthan that required to communicate the bit. This is often atall order due to the simplicity of backscatter.
Consider, for example, a simple aggregation operationthat sums ten sensor readings before transmitting the aggre-gate value over the radio. On traditional sensor platforms,such data reduction would have direct and significant powerbenefits since communication dominates power, and our ag-gregation scheme cuts this cost by a factor of ten. Thesame operation on a backscatter-based platform has dubi-ous benefits. Figure 2 shows that the number of NANDgates required for summing two bits is roughly nine (thirtysix transistors), but only four NAND gates (sixteen tran-sistors) and an additional transistor for backscattering thesignal are needed to transmit the same data via the shift-register controlled backscatter RF! As power consumptionis proportional to number of transistors, a nine gate adderconsumes 2.1⇥ more power than the shift-register controlledbackscatter RF.
It is necessary to add a few caveats to our simplified com-parison of computation and communication. The clock ratesof communication subsystems are limited by signal to noiseratio considerations, whereas the clock rates of processorscan be higher, and thereby reduce power. In addition, low-power processors use many tricks to reduce power consump-tion including optimized signal processing circuits, di↵er-ent power domains, extremely tight duty-cycling, and so on.Despite these optimizations, the cards are stacked againstcomputation. Backscatter is so incredibly simple in terms
of circuitry that even matching the e�ciency of backscatterbecomes a challenging architectural design problem.
Thus, the crux of our argument is the following: backscat-
ter drives down the optimal cross-over point between compu-
tation vs communication, such that communication of raw
data may be preferable to computation in a wider spectrum
of real-world scenarios.
Implications on architecture design: This observa-tion has an immediate implication on the architecture of abackscatter-based sensor platform. Traditional sensing plat-forms add a lot of computational modules between the sensorand the radio for sensor data acquisition, processing, filter-ing, bu↵ering, etc. The contribution of these components tooverall power consumption of an active radio-based sensorsystem is minimal and can largely be ignored. However, onbackscatter-based platforms, these components become thebottleneck.
This raises an intriguing question — with the power con-sumption of backscatter being so low, would it in fact bemore e�cient to eliminate all of these modules en-bloc, andjust connect the sensor directly to the radio? In other words,would it be better to just stream every bit of data that issensed directly through the radio?
We take a measurement-driven approach towards answer-ing these questions. First, we look at the computationalblocks between sensing and the RF interface on existingbackscatter-based sensing platforms to understand how muchpower they consume, as well as why they su↵er in termsof throughput. Second, we build on our empirical studyand design a radically new backscatter-based sensor plat-form that addresses these limitations.
3. INVESTIGATING EXISTING WIRELESSSENSING ARCHITECTURES
In this section, we investigate why current backscatter-based platforms are unable to achieve end-to-end power con-sumption of µWs for high-rate sensing and transfer. We alsoinvestigate why they are unable to achieve high-data ratecommunication, particularly while operating at low power.To empirically understand these factors, we look at the UMassMoo/UW WISP class platforms that are equipped with sen-sors, a low-power MCU (MSP 430 family) and a backscatterradio.
3.1 Poor energy efficiencyWe start with a break down of the power consumed by
three key computational modules on a UMass Moo (Fig-ure 3): 1) the sensor data acquisition subsystem which han-dles the protocols for operating sensors, 2) the data handlingsubsystem on a micro-controller where sensor data is stored,processed (if needed), formatted into packet, and sent to thenetwork stack, and 3) the network stack implemented in acombination of hardware and software.
3.1.1 Sensor data acquisitionSensor data acquisition is a relatively simple operation —
some sensors have an on-board ADC, hence data acquisitionis via a protocol such as SPI or I2C, whereas other sensorsjust provide an analog signal which is digitized using themicro-controller’s ADC. Despite its simplicity, even theseoperations are not as cheap as one might expect. For exam-ple, sampling an accelerometer via the SPI bus would require
~3.7uW
~34uW
3mW power hog
Computation IS more Expensive Than Communication
A"
B"
Cin"
S"
Cout"
(a) 1 bit adder circuits for computation.
D"
E"
(a) 1 bit adder circuits for computation.
(b) 1 bit shift register circuits forbackscatter.
Figure 2: 1 bit adder and 1 bit shift register circuit.
source prior to communication. Indeed, this tradeo↵ hasbeen reinforced by performance/power trends over the pastdecade — power consumption of embedded processors havedropped dramatically, while power reduction in active radioshas been relatively slower.
QQHowever, backscatter communication challenges this long-
held view. Backscatter is inherently extraordinarily e�cientsince the carrier wave is generated by the reader, and the tagonly backscatters the signal without any additional ampli-fication. Thus, each bit of backscatter is extremely simple,and only requires a handful of gates (Figure 2). This impliesthat for computation to be cheaper energy-wise, the compu-tational operations on each bit would have to use fewer gatesthan that required to communicate the bit. This is often atall order due to the simplicity of backscatter.
Consider, for example, a simple aggregation operationthat sums ten sensor readings before transmitting the aggre-gate value over the radio. On traditional sensor platforms,such data reduction would have direct and significant powerbenefits since communication dominates power, and our ag-gregation scheme cuts this cost by a factor of ten. The sameoperation on a backscatter-based platform has dubious ben-efits. Figure 2 shows that the number of gates required forsumming two bits is roughly nine, but only four gates areneeded to transmit the same data via the shift register cir-cuit of backscatter! As power consumption is proportionalto number of gates, a nine gate adder consumes 2.2⇥ morepower than the a four gate backscatter circuit.
It is necessary to add a few caveats to our simplified com-parison of computation and communication. The clock ratesof communication subsystems are limited by signal to noiseratio considerations, whereas the clock rates of processorscan be higher, and thereby reduce power. In addition, low-power processors use many tricks to reduce power consump-tion including optimized signal processing circuits, di↵er-ent power domains, extremely tight duty-cycling, and so on.Despite these optimizations, the cards are stacked againstcomputation. Backscatter is so incredibly simple in termsof circuitry that even matching the e�ciency of backscatterbecomes a challenging architectural design problem.
Thus, the crux of our argument is the following: backscat-
ter drives down the optimal cross-over point between compu-
tation vs communication, such that communication of raw
data may be preferable to computation in a wider spectrum
of real-world scenarios.
Implications on architecture design: This observa-tion has an immediate implication on the architecture of abackscatter-based sensor platform. Traditional sensing plat-forms add a lot of computational modules between the sensorand the radio for sensor data acquisition, processing, filter-ing, bu↵ering, etc. The contribution of these components tooverall power consumption of an active radio-based sensorsystem is minimal and can largely be ignored. However, onbackscatter-based platforms, these components become thebottleneck.
This raises an intriguing question — with the power con-sumption of backscatter being so low, would it in fact bemore e�cient to eliminate all of these modules en-bloc, andjust connect the sensor directly to the radio? In other words,would it be better to just stream every bit of data that issensed directly through the radio?
We take a measurement-driven approach towards answer-ing these questions. First, we look at the computationalblocks between sensing and the RF interface on existingbackscatter-based sensing platforms to understand how muchpower they consume, as well as why they su↵er in termsof throughput. Second, we build on our empirical studyand design a radically new backscatter-based sensor plat-form that addresses these limitations.
3. INVESTIGATING EXISTING WIRELESSSENSING ARCHITECTURES
In this section, we investigate why current backscatter-based platforms are unable to achieve end-to-end power con-sumption of µWs for high-rate sensing and transfer. We alsoinvestigate why they are unable to achieve high-data ratecommunication, particularly while operating at low power.To empirically understand these factors, we look at the UMassMoo/UW WISP class platforms that are equipped with sen-sors, a low-power MCU (MSP 430 family) and a backscatterradio.
3.1 Poor energy efficiencyWe start with a break down of the power consumed by
three key computational modules on a UMass Moo (Fig-ure 3): 1) the sensor data acquisition subsystem which han-dles the protocols for operating sensors, 2) the data handlingsubsystem on a micro-controller where sensor data is stored,processed (if needed), formatted into packet, and sent to thenetwork stack, and 3) the network stack implemented in acombination of hardware and software.
3.1.1 Sensor data acquisitionSensor data acquisition is a relatively simple operation —
some sensors have an on-board ADC, hence data acquisitionis via a protocol such as SPI or I2C, whereas other sensorsjust provide an analog signal which is digitized using themicro-controller’s ADC. Despite its simplicity, even theseoperations are not as cheap as one might expect. For exam-ple, sampling an accelerometer via the SPI bus would requireperiodic wakeup of the MCU to fill the SPI bu↵er, sendingthe read command and read address to the sensor, as well as
(a) 1 bit adder circuits for computation.
(b) 1 bit shift register circuits forbackscatter.
Figure 2: 1 bit adder and 1 bit shift register circuit.
source prior to communication. Indeed, this tradeo↵ hasbeen reinforced by performance/power trends over the pastdecade — power consumption of embedded processors havedropped dramatically, while power reduction in active radioshas been relatively slower.
QQHowever, backscatter communication challenges this long-
held view. Backscatter is inherently extraordinarily e�cientsince the carrier wave is generated by the reader, and the tagonly backscatters the signal without any additional ampli-fication. Thus, each bit of backscatter is extremely simple,and only requires a handful of gates (Figure 2). This impliesthat for computation to be cheaper energy-wise, the compu-tational operations on each bit would have to use fewer gatesthan that required to communicate the bit. This is often atall order due to the simplicity of backscatter.
Consider, for example, a simple aggregation operationthat sums ten sensor readings before transmitting the aggre-gate value over the radio. On traditional sensor platforms,such data reduction would have direct and significant powerbenefits since communication dominates power, and our ag-gregation scheme cuts this cost by a factor of ten. The sameoperation on a backscatter-based platform has dubious ben-efits. Figure 2 shows that the number of gates required forsumming two bits is roughly nine, but only four gates areneeded to transmit the same data via the shift register cir-cuit of backscatter! As power consumption is proportionalto number of gates, a nine gate adder consumes 2.2⇥ morepower than the a four gate backscatter circuit.
It is necessary to add a few caveats to our simplified com-parison of computation and communication. The clock ratesof communication subsystems are limited by signal to noiseratio considerations, whereas the clock rates of processorscan be higher, and thereby reduce power. In addition, low-power processors use many tricks to reduce power consump-tion including optimized signal processing circuits, di↵er-ent power domains, extremely tight duty-cycling, and so on.Despite these optimizations, the cards are stacked againstcomputation. Backscatter is so incredibly simple in termsof circuitry that even matching the e�ciency of backscatterbecomes a challenging architectural design problem.
Thus, the crux of our argument is the following: backscat-
ter drives down the optimal cross-over point between compu-
tation vs communication, such that communication of raw
data may be preferable to computation in a wider spectrum
of real-world scenarios.
Implications on architecture design: This observa-tion has an immediate implication on the architecture of abackscatter-based sensor platform. Traditional sensing plat-forms add a lot of computational modules between the sensorand the radio for sensor data acquisition, processing, filter-ing, bu↵ering, etc. The contribution of these components tooverall power consumption of an active radio-based sensorsystem is minimal and can largely be ignored. However, onbackscatter-based platforms, these components become thebottleneck.
This raises an intriguing question — with the power con-sumption of backscatter being so low, would it in fact bemore e�cient to eliminate all of these modules en-bloc, andjust connect the sensor directly to the radio? In other words,would it be better to just stream every bit of data that issensed directly through the radio?
We take a measurement-driven approach towards answer-ing these questions. First, we look at the computationalblocks between sensing and the RF interface on existingbackscatter-based sensing platforms to understand how muchpower they consume, as well as why they su↵er in termsof throughput. Second, we build on our empirical studyand design a radically new backscatter-based sensor plat-form that addresses these limitations.
3. INVESTIGATING EXISTING WIRELESSSENSING ARCHITECTURES
In this section, we investigate why current backscatter-based platforms are unable to achieve end-to-end power con-sumption of µWs for high-rate sensing and transfer. We alsoinvestigate why they are unable to achieve high-data ratecommunication, particularly while operating at low power.To empirically understand these factors, we look at the UMassMoo/UW WISP class platforms that are equipped with sen-sors, a low-power MCU (MSP 430 family) and a backscatterradio.
3.1 Poor energy efficiencyWe start with a break down of the power consumed by
three key computational modules on a UMass Moo (Fig-ure 3): 1) the sensor data acquisition subsystem which han-dles the protocols for operating sensors, 2) the data handlingsubsystem on a micro-controller where sensor data is stored,processed (if needed), formatted into packet, and sent to thenetwork stack, and 3) the network stack implemented in acombination of hardware and software.
3.1.1 Sensor data acquisitionSensor data acquisition is a relatively simple operation —
some sensors have an on-board ADC, hence data acquisitionis via a protocol such as SPI or I2C, whereas other sensorsjust provide an analog signal which is digitized using themicro-controller’s ADC. Despite its simplicity, even theseoperations are not as cheap as one might expect. For exam-ple, sampling an accelerometer via the SPI bus would requireperiodic wakeup of the MCU to fill the SPI bu↵er, sendingthe read command and read address to the sensor, as well as
1"bit"shi*"register" Backsca2er"RF"
(b) 1 bit shift register controls a backscatter tran-sistor.
Figure 2: 1 bit adder and 1 bit shift register circuit.
dropped dramatically, while power reduction in active radioshas been relatively slower.
However, backscatter communication challenges this long-held view. Backscatter is inherently extraordinarily e�cientsince the carrier wave is generated by the reader, and the tagonly backscatters the signal without any additional ampli-fication. Thus, each bit of backscatter is extremely simple,and only requires a handful of gates (Figure 2). This impliesthat for computation to be cheaper energy-wise, the compu-tational operations on each bit would have to use fewer gatesthan that required to communicate the bit. This is often atall order due to the simplicity of backscatter.
Consider, for example, a simple aggregation operationthat sums ten sensor readings before transmitting the aggre-gate value over the radio. On traditional sensor platforms,such data reduction would have direct and significant powerbenefits since communication dominates power, and our ag-gregation scheme cuts this cost by a factor of ten. Thesame operation on a backscatter-based platform has dubi-ous benefits. Figure 2 shows that the number of NANDgates required for summing two bits is roughly nine (thirtysix transistors), but only four NAND gates (sixteen tran-sistors) and an additional transistor for backscattering thesignal are needed to transmit the same data via the shift-register controlled backscatter RF! As power consumptionis proportional to number of transistors, a nine gate adderconsumes 2.1⇥ more power than the shift-register controlledbackscatter RF.
It is necessary to add a few caveats to our simplified com-parison of computation and communication. The clock ratesof communication subsystems are limited by signal to noiseratio considerations, whereas the clock rates of processorscan be higher, and thereby reduce power. In addition, low-power processors use many tricks to reduce power consump-tion including optimized signal processing circuits, di↵er-ent power domains, extremely tight duty-cycling, and so on.Despite these optimizations, the cards are stacked againstcomputation. Backscatter is so incredibly simple in terms
of circuitry that even matching the e�ciency of backscatterbecomes a challenging architectural design problem.
Thus, the crux of our argument is the following: backscat-
ter drives down the optimal cross-over point between compu-
tation vs communication, such that communication of raw
data may be preferable to computation in a wider spectrum
of real-world scenarios.
Implications on architecture design: This observa-tion has an immediate implication on the architecture of abackscatter-based sensor platform. Traditional sensing plat-forms add a lot of computational modules between the sensorand the radio for sensor data acquisition, processing, filter-ing, bu↵ering, etc. The contribution of these components tooverall power consumption of an active radio-based sensorsystem is minimal and can largely be ignored. However, onbackscatter-based platforms, these components become thebottleneck.
This raises an intriguing question — with the power con-sumption of backscatter being so low, would it in fact bemore e�cient to eliminate all of these modules en-bloc, andjust connect the sensor directly to the radio? In other words,would it be better to just stream every bit of data that issensed directly through the radio?
We take a measurement-driven approach towards answer-ing these questions. First, we look at the computationalblocks between sensing and the RF interface on existingbackscatter-based sensing platforms to understand how muchpower they consume, as well as why they su↵er in termsof throughput. Second, we build on our empirical studyand design a radically new backscatter-based sensor plat-form that addresses these limitations.
3. INVESTIGATING EXISTING WIRELESSSENSING ARCHITECTURES
In this section, we investigate why current backscatter-based platforms are unable to achieve end-to-end power con-sumption of µWs for high-rate sensing and transfer. We alsoinvestigate why they are unable to achieve high-data ratecommunication, particularly while operating at low power.To empirically understand these factors, we look at the UMassMoo/UW WISP class platforms that are equipped with sen-sors, a low-power MCU (MSP 430 family) and a backscatterradio.
3.1 Poor energy efficiencyWe start with a break down of the power consumed by
three key computational modules on a UMass Moo (Fig-ure 3): 1) the sensor data acquisition subsystem which han-dles the protocols for operating sensors, 2) the data handlingsubsystem on a micro-controller where sensor data is stored,processed (if needed), formatted into packet, and sent to thenetwork stack, and 3) the network stack implemented in acombination of hardware and software.
3.1.1 Sensor data acquisitionSensor data acquisition is a relatively simple operation —
some sensors have an on-board ADC, hence data acquisitionis via a protocol such as SPI or I2C, whereas other sensorsjust provide an analog signal which is digitized using themicro-controller’s ADC. Despite its simplicity, even theseoperations are not as cheap as one might expect. For exam-ple, sampling an accelerometer via the SPI bus would require
Compute - add, sub, etc Transmit - via backscatter
A"
B"
Cin"
S"
Cout"
(a) 1 bit adder circuits for computation.
D"
E"
(a) 1 bit adder circuits for computation.
(b) 1 bit shift register circuits forbackscatter.
Figure 2: 1 bit adder and 1 bit shift register circuit.
source prior to communication. Indeed, this tradeo↵ hasbeen reinforced by performance/power trends over the pastdecade — power consumption of embedded processors havedropped dramatically, while power reduction in active radioshas been relatively slower.
QQHowever, backscatter communication challenges this long-
held view. Backscatter is inherently extraordinarily e�cientsince the carrier wave is generated by the reader, and the tagonly backscatters the signal without any additional ampli-fication. Thus, each bit of backscatter is extremely simple,and only requires a handful of gates (Figure 2). This impliesthat for computation to be cheaper energy-wise, the compu-tational operations on each bit would have to use fewer gatesthan that required to communicate the bit. This is often atall order due to the simplicity of backscatter.
Consider, for example, a simple aggregation operationthat sums ten sensor readings before transmitting the aggre-gate value over the radio. On traditional sensor platforms,such data reduction would have direct and significant powerbenefits since communication dominates power, and our ag-gregation scheme cuts this cost by a factor of ten. The sameoperation on a backscatter-based platform has dubious ben-efits. Figure 2 shows that the number of gates required forsumming two bits is roughly nine, but only four gates areneeded to transmit the same data via the shift register cir-cuit of backscatter! As power consumption is proportionalto number of gates, a nine gate adder consumes 2.2⇥ morepower than the a four gate backscatter circuit.
It is necessary to add a few caveats to our simplified com-parison of computation and communication. The clock ratesof communication subsystems are limited by signal to noiseratio considerations, whereas the clock rates of processorscan be higher, and thereby reduce power. In addition, low-power processors use many tricks to reduce power consump-tion including optimized signal processing circuits, di↵er-ent power domains, extremely tight duty-cycling, and so on.Despite these optimizations, the cards are stacked againstcomputation. Backscatter is so incredibly simple in termsof circuitry that even matching the e�ciency of backscatterbecomes a challenging architectural design problem.
Thus, the crux of our argument is the following: backscat-
ter drives down the optimal cross-over point between compu-
tation vs communication, such that communication of raw
data may be preferable to computation in a wider spectrum
of real-world scenarios.
Implications on architecture design: This observa-tion has an immediate implication on the architecture of abackscatter-based sensor platform. Traditional sensing plat-forms add a lot of computational modules between the sensorand the radio for sensor data acquisition, processing, filter-ing, bu↵ering, etc. The contribution of these components tooverall power consumption of an active radio-based sensorsystem is minimal and can largely be ignored. However, onbackscatter-based platforms, these components become thebottleneck.
This raises an intriguing question — with the power con-sumption of backscatter being so low, would it in fact bemore e�cient to eliminate all of these modules en-bloc, andjust connect the sensor directly to the radio? In other words,would it be better to just stream every bit of data that issensed directly through the radio?
We take a measurement-driven approach towards answer-ing these questions. First, we look at the computationalblocks between sensing and the RF interface on existingbackscatter-based sensing platforms to understand how muchpower they consume, as well as why they su↵er in termsof throughput. Second, we build on our empirical studyand design a radically new backscatter-based sensor plat-form that addresses these limitations.
3. INVESTIGATING EXISTING WIRELESSSENSING ARCHITECTURES
In this section, we investigate why current backscatter-based platforms are unable to achieve end-to-end power con-sumption of µWs for high-rate sensing and transfer. We alsoinvestigate why they are unable to achieve high-data ratecommunication, particularly while operating at low power.To empirically understand these factors, we look at the UMassMoo/UW WISP class platforms that are equipped with sen-sors, a low-power MCU (MSP 430 family) and a backscatterradio.
3.1 Poor energy efficiencyWe start with a break down of the power consumed by
three key computational modules on a UMass Moo (Fig-ure 3): 1) the sensor data acquisition subsystem which han-dles the protocols for operating sensors, 2) the data handlingsubsystem on a micro-controller where sensor data is stored,processed (if needed), formatted into packet, and sent to thenetwork stack, and 3) the network stack implemented in acombination of hardware and software.
3.1.1 Sensor data acquisitionSensor data acquisition is a relatively simple operation —
some sensors have an on-board ADC, hence data acquisitionis via a protocol such as SPI or I2C, whereas other sensorsjust provide an analog signal which is digitized using themicro-controller’s ADC. Despite its simplicity, even theseoperations are not as cheap as one might expect. For exam-ple, sampling an accelerometer via the SPI bus would requireperiodic wakeup of the MCU to fill the SPI bu↵er, sendingthe read command and read address to the sensor, as well as
(a) 1 bit adder circuits for computation.
(b) 1 bit shift register circuits forbackscatter.
Figure 2: 1 bit adder and 1 bit shift register circuit.
source prior to communication. Indeed, this tradeo↵ hasbeen reinforced by performance/power trends over the pastdecade — power consumption of embedded processors havedropped dramatically, while power reduction in active radioshas been relatively slower.
QQHowever, backscatter communication challenges this long-
held view. Backscatter is inherently extraordinarily e�cientsince the carrier wave is generated by the reader, and the tagonly backscatters the signal without any additional ampli-fication. Thus, each bit of backscatter is extremely simple,and only requires a handful of gates (Figure 2). This impliesthat for computation to be cheaper energy-wise, the compu-tational operations on each bit would have to use fewer gatesthan that required to communicate the bit. This is often atall order due to the simplicity of backscatter.
Consider, for example, a simple aggregation operationthat sums ten sensor readings before transmitting the aggre-gate value over the radio. On traditional sensor platforms,such data reduction would have direct and significant powerbenefits since communication dominates power, and our ag-gregation scheme cuts this cost by a factor of ten. The sameoperation on a backscatter-based platform has dubious ben-efits. Figure 2 shows that the number of gates required forsumming two bits is roughly nine, but only four gates areneeded to transmit the same data via the shift register cir-cuit of backscatter! As power consumption is proportionalto number of gates, a nine gate adder consumes 2.2⇥ morepower than the a four gate backscatter circuit.
It is necessary to add a few caveats to our simplified com-parison of computation and communication. The clock ratesof communication subsystems are limited by signal to noiseratio considerations, whereas the clock rates of processorscan be higher, and thereby reduce power. In addition, low-power processors use many tricks to reduce power consump-tion including optimized signal processing circuits, di↵er-ent power domains, extremely tight duty-cycling, and so on.Despite these optimizations, the cards are stacked againstcomputation. Backscatter is so incredibly simple in termsof circuitry that even matching the e�ciency of backscatterbecomes a challenging architectural design problem.
Thus, the crux of our argument is the following: backscat-
ter drives down the optimal cross-over point between compu-
tation vs communication, such that communication of raw
data may be preferable to computation in a wider spectrum
of real-world scenarios.
Implications on architecture design: This observa-tion has an immediate implication on the architecture of abackscatter-based sensor platform. Traditional sensing plat-forms add a lot of computational modules between the sensorand the radio for sensor data acquisition, processing, filter-ing, bu↵ering, etc. The contribution of these components tooverall power consumption of an active radio-based sensorsystem is minimal and can largely be ignored. However, onbackscatter-based platforms, these components become thebottleneck.
This raises an intriguing question — with the power con-sumption of backscatter being so low, would it in fact bemore e�cient to eliminate all of these modules en-bloc, andjust connect the sensor directly to the radio? In other words,would it be better to just stream every bit of data that issensed directly through the radio?
We take a measurement-driven approach towards answer-ing these questions. First, we look at the computationalblocks between sensing and the RF interface on existingbackscatter-based sensing platforms to understand how muchpower they consume, as well as why they su↵er in termsof throughput. Second, we build on our empirical studyand design a radically new backscatter-based sensor plat-form that addresses these limitations.
3. INVESTIGATING EXISTING WIRELESSSENSING ARCHITECTURES
In this section, we investigate why current backscatter-based platforms are unable to achieve end-to-end power con-sumption of µWs for high-rate sensing and transfer. We alsoinvestigate why they are unable to achieve high-data ratecommunication, particularly while operating at low power.To empirically understand these factors, we look at the UMassMoo/UW WISP class platforms that are equipped with sen-sors, a low-power MCU (MSP 430 family) and a backscatterradio.
3.1 Poor energy efficiencyWe start with a break down of the power consumed by
three key computational modules on a UMass Moo (Fig-ure 3): 1) the sensor data acquisition subsystem which han-dles the protocols for operating sensors, 2) the data handlingsubsystem on a micro-controller where sensor data is stored,processed (if needed), formatted into packet, and sent to thenetwork stack, and 3) the network stack implemented in acombination of hardware and software.
3.1.1 Sensor data acquisitionSensor data acquisition is a relatively simple operation —
some sensors have an on-board ADC, hence data acquisitionis via a protocol such as SPI or I2C, whereas other sensorsjust provide an analog signal which is digitized using themicro-controller’s ADC. Despite its simplicity, even theseoperations are not as cheap as one might expect. For exam-ple, sampling an accelerometer via the SPI bus would requireperiodic wakeup of the MCU to fill the SPI bu↵er, sendingthe read command and read address to the sensor, as well as
1"bit"shi*"register" Backsca2er"RF"
(b) 1 bit shift register controls a backscatter tran-sistor.
Figure 2: 1 bit adder and 1 bit shift register circuit.
dropped dramatically, while power reduction in active radioshas been relatively slower.
However, backscatter communication challenges this long-held view. Backscatter is inherently extraordinarily e�cientsince the carrier wave is generated by the reader, and the tagonly backscatters the signal without any additional ampli-fication. Thus, each bit of backscatter is extremely simple,and only requires a handful of gates (Figure 2). This impliesthat for computation to be cheaper energy-wise, the compu-tational operations on each bit would have to use fewer gatesthan that required to communicate the bit. This is often atall order due to the simplicity of backscatter.
Consider, for example, a simple aggregation operationthat sums ten sensor readings before transmitting the aggre-gate value over the radio. On traditional sensor platforms,such data reduction would have direct and significant powerbenefits since communication dominates power, and our ag-gregation scheme cuts this cost by a factor of ten. Thesame operation on a backscatter-based platform has dubi-ous benefits. Figure 2 shows that the number of NANDgates required for summing two bits is roughly nine (thirtysix transistors), but only four NAND gates (sixteen tran-sistors) and an additional transistor for backscattering thesignal are needed to transmit the same data via the shift-register controlled backscatter RF! As power consumptionis proportional to number of transistors, a nine gate adderconsumes 2.1⇥ more power than the shift-register controlledbackscatter RF.
It is necessary to add a few caveats to our simplified com-parison of computation and communication. The clock ratesof communication subsystems are limited by signal to noiseratio considerations, whereas the clock rates of processorscan be higher, and thereby reduce power. In addition, low-power processors use many tricks to reduce power consump-tion including optimized signal processing circuits, di↵er-ent power domains, extremely tight duty-cycling, and so on.Despite these optimizations, the cards are stacked againstcomputation. Backscatter is so incredibly simple in terms
of circuitry that even matching the e�ciency of backscatterbecomes a challenging architectural design problem.
Thus, the crux of our argument is the following: backscat-
ter drives down the optimal cross-over point between compu-
tation vs communication, such that communication of raw
data may be preferable to computation in a wider spectrum
of real-world scenarios.
Implications on architecture design: This observa-tion has an immediate implication on the architecture of abackscatter-based sensor platform. Traditional sensing plat-forms add a lot of computational modules between the sensorand the radio for sensor data acquisition, processing, filter-ing, bu↵ering, etc. The contribution of these components tooverall power consumption of an active radio-based sensorsystem is minimal and can largely be ignored. However, onbackscatter-based platforms, these components become thebottleneck.
This raises an intriguing question — with the power con-sumption of backscatter being so low, would it in fact bemore e�cient to eliminate all of these modules en-bloc, andjust connect the sensor directly to the radio? In other words,would it be better to just stream every bit of data that issensed directly through the radio?
We take a measurement-driven approach towards answer-ing these questions. First, we look at the computationalblocks between sensing and the RF interface on existingbackscatter-based sensing platforms to understand how muchpower they consume, as well as why they su↵er in termsof throughput. Second, we build on our empirical studyand design a radically new backscatter-based sensor plat-form that addresses these limitations.
3. INVESTIGATING EXISTING WIRELESSSENSING ARCHITECTURES
In this section, we investigate why current backscatter-based platforms are unable to achieve end-to-end power con-sumption of µWs for high-rate sensing and transfer. We alsoinvestigate why they are unable to achieve high-data ratecommunication, particularly while operating at low power.To empirically understand these factors, we look at the UMassMoo/UW WISP class platforms that are equipped with sen-sors, a low-power MCU (MSP 430 family) and a backscatterradio.
3.1 Poor energy efficiencyWe start with a break down of the power consumed by
three key computational modules on a UMass Moo (Fig-ure 3): 1) the sensor data acquisition subsystem which han-dles the protocols for operating sensors, 2) the data handlingsubsystem on a micro-controller where sensor data is stored,processed (if needed), formatted into packet, and sent to thenetwork stack, and 3) the network stack implemented in acombination of hardware and software.
3.1.1 Sensor data acquisitionSensor data acquisition is a relatively simple operation —
some sensors have an on-board ADC, hence data acquisitionis via a protocol such as SPI or I2C, whereas other sensorsjust provide an analog signal which is digitized using themicro-controller’s ADC. Despite its simplicity, even theseoperations are not as cheap as one might expect. For exam-ple, sampling an accelerometer via the SPI bus would require
1 bit adder consumes 36 transistors 1 bit TX consumes 17 transistors
Sensor Interface Processing Network
StackRadio
Control
Computation is typically more expensive than backscatter!
Intuition - Cut Down Computational Overhead
Sensor Interface Processing Network
StackRadio
Control
Intuition - Cut Down Computational Overhead
Minimalist System Design
Which components contribute to computational overhead?
Acquiring Sensor Data
SensorSPI/I2C
ADC
MCU MCU on time
Sampling rate
Sensor Interface Processing Network
StackRadio
Control
PowerAccel + SPI 84uW @ 400 HzMIC + ADC 492uW @ 44kHz
Acquiring Sensor Data
SensorSPI/I2C
ADC
MCU MCU on time
Sampling rate
Sensor Interface Processing Network
StackRadio
Control
PowerAccel + SPI 84uW @ 400 HzMIC + ADC 492uW @ 44kHz
High rate data acquisition does contribute to significant power increase.
DMA Driven Data Migration
Accelerometer)
Microphone)
Temperature)
Sensors)
A/D)Converter)
SPI)
I2C)
Sensing))subsystem)
Timer)ISR)
DMA)
RAM)
Data)handling)subsystem)
BackscaBer)Radio)
Encoding)
Network)Stack)
CommunicaFon)subsystem)
Figure 3: Computational blocks on existingbackscatter-based sensors.
0
50
100
150
200
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Pow
er (u
W)
Time (ms)
Figure 4: Power consumed for handling timer inter-rupts at 4kHz. The MCU is unable to switch to sleepmode due to frequent interrupts.
0 20 40 60 80
100 120 140 160
0.001 0.01 0.1 1 10 100
Pow
er (u
W)
Frequency (kHz)
DMALPM3 SLeep
Figure 5: The power consumption of DMA transferat di↵erent frequencies.
Figure 6: Power consumption of DMA transfer at100Hz. DMA is slow to return to sleep mode.
periodic wakeup of the MCU to fill the SPI bu↵er, sendingthe read command and read address to the sensor, as well asproviding the clock for the SPI bus. The overall result is thatthe MCU is active for about 40% of the time when acquiringdata from an accelerometer sampling at 400 Hz. This ac-quisition operation, in itself, consumes 84µW of power, 14⇥
higher than the accelerometer (6µW). The cost of acquiringaudio data is equally high — when sampling an audio sen-sor (ADMP803) at 44KHz, acquisition consumes 492µW ofpower, 14.5⇥ higher than the audio sensor (34µW).
3.1.2 Data handling subsystemThe data handling subsystem is the block that processes
the acquired sensor data, formats and packetizes it, andsends it to the network stack. To minimize this overhead,sensor systems typically operate in a duty-cycled mode wherethe MCU is turned on for a minimal amount of time neededto handle the data, before switching back into sleep modeto conserve energy.
However, this optimization is no longer e↵ective when thissubsystem handles high-rate sensors. Figure 4 shows thepower consumption for executing the timer interrupt ser-vice routine to handle each acquired audio sample. At highrate, the MCU is rarely able to switch completely back intothe ultra-low power sleep mode due to frequent interrupts.Thus, the overall power consumption of the data handlingmodule is roughly the ballpark of active mode power con-sumption of the MCU (a few mW), which is several orders ofmagnitude higher than the power consumed by the sensor.
One method to reduce power of the data handling subsys-tem is to use Direct Memory Access (DMA), which allowstransfer of data from the sensor to memory without wak-ing up the MCU. This raises the possibility that waking upthe MCU can, perhaps, be avoided altogether if the data istransferred directly from the sensor to the network queuewithout any processing.
Surprisingly, DMA does not reduce power consumption.Figure 5 shows empirically measured power consumption forDMA transfer on an MSP 430, which moves the sensor datafrom a sensor to a local memory at di↵erent frequencies. Weobserve that while DMA is e�cient at low rates (e..g below100Hz), it has high power consumption at high transfer rates— for example, DMA transfer consumes 149.2 µW of powerat 44 kHz, 60⇥ higher than the 2.5 µW of LPM3 sleep modeof the MCU. This is surprising since one would expect thatthe MCU is in sleep mode while DMA operates.
The culprit for high power consumption of DMA turns outto be its tail energy consumption. Figure 6 shows the powerconsumption of repeated DMA transfer at 100 Hz. This ex-periment is done with an MSP 430 set to LPM3 sleep modeand a timer that periodically triggers DMA transfer. Whena DMA transfer is initiated, its power consumption increasesto 40µW within 10us, and starts decreasing once the DMAtransfer is done. However, the power consumption decays ata relatively slow rate compared to the sharp increase, result-ing in a long tail of roughly 3.5ms. When the DMA transferfrequency is high, such as 5kHz shown in Figure 5, the longtail leads to high power consumption. While we are notcertain about the cause of this behavior, one hypothesis is
DMA60x
CPU Sleep
Sensor Interface
Network Stack
Sensor Interface Processing Network
StackRadio
Control
Emulating Network Stack in Software
Query
RN16
Reader Sensor
Sensor Interface Processing Network
StackRadio
Control
ACKEPC
EPC Gen 2 Timing Diagram
MCU is on for 67% of time. 2mW
Control Backscatter via Hardware Peripheral
Sensor Interface Processing Network
StackRadio
Control
Reader Sensor
EPC Gen 2
Software to toggle the backscatter transistor
3mW
Control Backscatter via Hardware Peripheral
Sensor Interface Processing Network
StackRadio
Control
Reader Sensor
EPC Gen 2
Buffer: 0101
ASK signal
UART Hardware Peripheral
2mW
How to design an architecture that leverages the ultra low-power characteristic of backscatter and achieve ~1 Mbps
wireless transmission while consuming 10~100 micro-watts?
Ekho Architecture
Minimalist Design
data ASK signal
Trigger TX
WISP architecture
Ekho architecture
Software ADC DMA Network
StackRadio
Control
Hardware ADC FIFO Timer Shift
Register
Implementation
Processor
RFanalog
Sensors
Figure 11: Ekho is implemented as a low-profileprinted circuit board with small form factor.
Our backscatter circuit directly provide a small bias currentto the transistor and retains a sharp edge for the generatedOOK signal.
A critical element of our hardware design is the clock sys-tem which drives the FPGA logic. The core of our clocksystem is a 1MHz ultra low power crystal oscillator that di-rectly feeds into the FPGA. The 1MHz clock is divided todrive di↵erent components of the architecture because sens-ing, data handling, and communication subsystems operateat di↵erent speeds. Our clock system is di↵erent from theMoo and WISP platforms, where a digital generated clock(DCO) is used. Although the DCO can also be dividedfor driving di↵erent components, it couples the operationalmodes of the system and its clock speed, as a result of whichthe high speed clock is only available when the system op-erates as a whole in a high power mode.
5.2 Software defined backscatter readerWe used the USRP N210 mother board and the SBX
RF daughterboard to build our software defined backscat-ter reader for receiving high speed backscatter signals fromEkho. We construct a signal processing pipeline that is ableto track the amplitude of the carrier wave that is used as thereference for decoding the OOK signal generated by Ekho.Our decoding is di↵erent from Moo and WISP platformswhere Miller-4 encoding is used on top of the OOK signaland a decoding template can be used for correlating the re-ceived signal and output a bit when the template matchesthe received signal. In Ekho, the data is sent directly viaOOK and encoding is not used. Therefore, we need to trackthe amplitude of carrier wave to determine whether the re-ceived signal is a high or low pulse.
5.3 MAC layer protocolFigure 12 shows the timing diagram of the Ekho MAC
layer. The first stage is to inventory the nodes in the net-work, and obtain information about their SNR and othersensor-related information. This phase executes very sim-ilar to an EPC Gen 2 singulation phase, where nodes canselect a slot to transfer in, and send a short sequence ofbits with the appropriate information. After the singulationphase, the reader executes the optimization algorithm de-scribed in §4 and determines the time period and bit ratefor each sensor, which is then relayed to the sensor. Thereader initiates the singulation phase under several circum-stances: a) when significant changes are observed in SNR,which might signify changes in position or orientation, andb) when collisions are detected, which might signify that anew node is attempting to join the network.
Once the reader informs each sensor of its bit rate andperiod, it initializes slots by sending a synchronization signalduring which it shuts down the carrier for a short 10 µs
Backsca&er)Reader) Sensor)
Sync)
Sensor)1)
Sensor)2)
Sensor)n)
Sensor)1)
Sensor)2)
Sync)
Sensor)1)
Sensor)2)
Timer)driven)TX)
Overlap))detected)
Reset))Timer)
Backsca&er)Reader) Sensor)
Query)
RN16)
ACK)
Sensor1)
Query)
RN16)
ACK)
Sensor2)
Query)
EPC)Gen)2))Eming)diagram)
Ekho)MAC))Eming)diagram)
Figure 12: Timeline of Ekho MAC.
window. This pulse informs all nodes simultaneously thatthey should start their timers, thereby initiating the TDMAschedule. The length of the sync message needs to be chosensmall enough to amortize overhead, but large enough to bedetectable at the sensor, hence our choice of 10µs.
When the reader detects that data transmitted during ad-jacent slots are overlapping into each other (due to clockdrift), it re-issues a synchronization pulse to restart thetimers on all nodes. Overlap between sensors can be de-tected by looking at the constellation map of the receivedsignal — if two clusters are present, it indicates that acollision-free signal is received and if more clusters are present,it indicates that a collided signal is observed [22]. If multiplesynchronization pulses fail to eliminate collisions, the readerswitches back into inventory mode.
6. EVALUATIONWe now evaluate the overall performance of EkhoNet in-
cluding 1) demonstrating the power benefit of the Ekho ar-chitecture, 2) benchmarking the performance of the EkhoNetMAC, and 3) evaluating EkhoNet’s ability to support high-rate streams from many sensors while operating at extremelylow power consumption.
6.1 Experimental setupWe deploy 10 Ekho nodes 1 feet to 9 feet from a backscat-
ter reader. Our experiments do not cover distances largerthan 9 feet because of the poor signal quality beyond 9 feet.This is a result of the 100mW maximum power issued by theSBX RF daughterboard, which is 10⇥ smaller than commer-cial RFID readers.
To understand the power benefits of Ekho, we compareagainst the UMass Moo (equivalent of Intel WISP 4.0) andthe WISP5.0 platforms. Since the WISP5.0 platform is notcurrently available, we evaluate its power consumption witha prototype that uses the same MCU (MSP430FR5969).Since the MCU is the main power hog in the system, thisprovides a good proxy for measuring power consumption.
FPGABackscatter reader
- USRP N210
Hardware ADC FIFO Timer Shift
Register
Experiments - WISP5.0 (MSP430 FR5969) - UMass Moo
Power Consumption of Operating a Microphone
10 15 20 25 30 35 40
1 10 100 1000 10000
Pow
er (u
W)
Sampling Rate (Hz)
EkhoEkho consumes 37uW @44kHz
Hardware ADC FIFO Timer Shift
Register
Power Consumption of Operating a Microphone
Hardware ADC FIFO Timer Shift
Register
0 10 20 30 40 50 60 70 80 90
1 10 100 1000
Pow
er (u
W)
Sampling Rate (Hz)
Software SPIHardware SPI
(a) Accelerometer.
0 50
100 150 200 250 300 350 400 450 500
0.001 0.01 0.1 1 10 100
Pow
er (u
W)
Sampling Rate (kHz)
Software ADCHardware ADC
(b) Microphone.
Figure 13: Power reduction for sensing subsystem: a) sampling anaccelerometer, b) sampling a microphone.
1
10
100
1000
10000
1 10 100 1000
Pow
er (u
W)
Write Rate (kHz)
Timer ISRDMAFIFO
Figure 14: Power reduction fordata transfer to network queue.
0 500
1000 1500 2000 2500 3000 3500
0.001 0.01 0.1 1 10 100
Pow
er (u
W)
Bit Rate (Mbps)
SoftwareUART
Shift Reg
Figure 15: The power consumptionof operating a backscatter radio.
0 50
100 150 200 250 300
1 10 100 1000
Pow
er (u
W)
Sampling Rate (Hz)
MooWISP5.0
Ekho
Figure 16: Whole-system powerconsumption for operating an ac-celerometer sensor.
0 500
1000 1500 2000 2500 3000
1 10 100 1000 10000
Pow
er (u
W)
Sampling Rate (Hz)
MooWISP5.0
Ekho
Figure 17: Whole-system powerconsumption for operating an au-dio sensor.
6.2 Ekho power benchmarksWe begin our evaluation by validating the claim that the
power optimizations on Ekho can substantially reduce theoverheads incurred by existing platforms. We follow theorganization in §4, and show benchmarks for each module —sensor data acquisition, sensor data handling, and networkstack.
Figure 13 measure the power of the sensing subsystemwhen Ekho interacts with two types of sensors — an ac-celerometer with on-board ADC that connects to the MCUvia a SPI interface, and an audio sensor where the MCU’sADC is used to sample the sensor. We compare Ekho versusa WISP/Mote-class sensor device (i.e. a device where thesensor connects to an MCU that acquires data). In bothcases, we can see that Ekho reduces power consumptionsubstantially — for sampling the accelerometer, Ekho re-duces power by 1.5⇥ at 400Hz by eliminating the overheadof software-controlled SPI, and for sampling the audio sen-sor, Ekho reduces power by 22⇥ by trimming the overheadof running the software-controlled ADC at 44kHz.
Figure 14 measures the power consumption of the datahandling subsystem of Ekho, which is composed by a 2kBFIFO bu↵er for connecting sensors to the RF analog frontend. The 2kB FIFO bu↵er only consumes 26.5µW of powerwhen data is written into the FIFO at 500kHz, 14.4⇥ lowerthan the 384µW consumed by DMA driven data migrationand 92⇥ lower than the 1.5mW consumed by timer drivendata migration.
Figure 15 shows the power consumption of the communi-cation subsystem which is composed of a shift register andbackscatter radio. At 1Mbps, Ekho’s communication sub-
system consumes only 77µW of power, 13.4⇥ lower thana UART controlled backscatter radio implemented on theWISP and 44⇥ lower than a software controlled backscatterradio implemented on the WISP. For software and UARTcontrolled backscatter radios, we do not measure power atbit rates higher than 6Mbps because the maximum clock onrate on WISP platform is 24MHz, which limits the maxi-mum achievable bit rate.
6.3 Whole-system power consumptionHaving looked at power benchmarks for individual com-
ponents of Ekho, we turn to a whole-system power measure-ment from sensing to transmission. We look at the overallpower consumed by Ekho when operating the same two sen-sors as earlier — accelerometer and microphone.
We start with a measurement of Ekho with an accelerom-eter. The sensor has a built-in ADC and talks via SPI tothe sensor platform. Figure 16 shows that at 1Hz, the powerconsumption of Ekho is higher than Moo and WISP5.0 plat-forms. This is because the static current consumption of theFPGA at the core of Ekho is 8.9µA, much higher than the0.1µA static current draw of Moo and WISP5.0. However,when the frequency of operating the accelerometer increases,the power consumed by Moo and WISP5.0 platforms in-creases significantly while the Ekho system still consumesonly tens of µW. At 400Hz, the Ekho system consumes35µW of power, 7.6⇥ lower than the 266µW of Moo and3.3⇥ lower than the 118µW of WISP5.0.
We now turn to power measurements when Ekho is con-nected to a microphone. An external ADC is used to samplethe audio sensor, and send a digital signal to the core plat-
76x over Moo 13.5x over WISP5.0 @44kHz
Power Consumption of Whole System
Hardware ADC FIFO Timer Shift
Register
Power Frequency Rate
Accelerometer 41uW 400Hz 1 Mbps
Microphone 71uW 44kHz 1 Mbps
We achieve less 100 uW for high rate sensors.
How to coordinate multiple Ekhos?
Coordination - Leveraging the Powerful Reader
Powerful backscatter reader
Ekho
Insight - Intelligent decision making logic, such as scheduling, should be placed at the reader!
Coordination - Balance Three Factors
Powerful backscatter reader
EkhoTX energy efficiency!- Bit rate - Bits per Joule
Channel quality!- Bit rate - SNR
Data quality!- Sampling rate - MOS score - Bit rate
Optimization Solver
Scheduling
Coordinating 10 Ekhos Connected Microphones
USRP N210
EkhoSNR
Energy efficiency
Data quality
Aggregated throughput ~780kbps.
Conclusion
Ekho achieves ~1 Mbps wireless transmission while consuming ~100 micro-watts.
Raw data streaming!
Hardware ADC FIFO Timer Shift
Register