, MatthiasLippunerAndres Gomez, Luca Benini, Lothar Thiele

Designing Reliable Transient Applica�ons

Dept. Electrical Engineering and Information Technology, ETH Zürich

Institut für Integrierte SystemeIntegrated Systems Laboratory

Energy harvesting is a clean, long-term supply of relatively low and volatile power.

Burst-based application design can operate efficiently over a wide input power range.

Burst-based design must satisfy certain constraints to provide application reliability:

Transient camera applicationsEnergy bursts

Experimental evaluation

Intermittent power and reliable execution can co-exist

t

Vcap

Vload,min

Vmax

cold-start energy build-up

taskexecution

tsleep ton

Power ranges

Energy Conservation:

Eactive = ηsystem * EinPin ∝

Pactive

ton

tsleep1) 2)

Quantized energy transfer [1]

Maximizes task-level energy efficiency.

Power-point tracking for both source and load.

Guarantees atomic task execution.

Optimizes capacitor size for reduced wake-up times.

Transducers(area = 1 cm2)

MCU's(Factive=1MHz)

Sensors(active)

Radios(PTX,avg: 0/10 dBm)

100nW

1µW

1mW

100mW

10µW

100µW

10mW

1W

Energy scalability

Motion estimation using MSP432

[1] A. Gomez et al. "Dynamic Energy Burst Scaling for Transiently Powered Systems," Proc. DATE Conf. 2016.

[2] A. Gomez et al. "Efficient Long-Term Logging of Rich Data Sensors using Transient Sensors Nodes," ACM Trans. Embed. Comput. Syst. 2017.

[3] A. Gomez et al. "Wearable, Energy-Opportunistic Vision Sensing for Walking Speed Estimation," Proc. SAS. 2017.

Minimum energy availability - necessary to guarantee application progress.

Temporal independence - necessary for functional correctness.

Non-volatility - necessary to guarantee data dependencies between bursts.

2)Walking speed estimation [3]

Track pixel displacement to estimateuser's walking speed.

1) Long-term logging [2]

Non-Volatile Memory Hierarchy (NVMH)reduces costs of logging images to SD Card.

Burst-based application using NVMH:

0 2 4 6 8 10 120

1

2

3

4

5

add DEBS

add NVMH

← better Energy per Image Stored [mJ]

←be

tter

Cap

acitor

Size[m

F]

Single BurstDEBS onlyNVMH onlyDEBS+NVMH

MCU

ADC

FlashDMA SPI

Application Circuit

digital controlanalog data

Energy Flow

Control Signals

StonymanCameraSolar

CellEMU

FRAM

Data Flow

GPIO

Transient Camera

Camera Schematic:

Wearable Prototype:

500 1000 1500 2000 2500 3000Frame Number

-101234567

Posit

ion in

x-Di

recti

on (m

)

block size = 96block size = 48block size = 32block size = 16Reference Points

Controlled displacement to measure accuracy:

Camera mounted on trolley without power constraints.

Accumulated error: 57 cm over 42 m.

0 20 40 60 80 100 120 140 160 180Estimation Number

0

0.5

1

1.5

2

Estim

ated

Vel

ocity

(m/s

)

raw datalow-pass filtered data

standing walking slowly standing walkingfast

st.

Real-world experiment to estimate walking speed.

Burst-based design guarantees functionalityindependently of harvester's properties.

Harvester selection relates only to application's performance (execution rate).

Efficient EMU design can minimize therequired harvester area for a desired performance.

0 500 1000 1500 2000 2500

101

102

103

Input Power [µW]

Start-UpTim

e[s]

Single Burst (2 000µF)DEBS Only (1 470µF)NVMH only (4 300µF)DEBS+NVMH (3300µF)

0 500 1'000 1'500 2'0000

5

10

15

20

25

Illuminance [lx]

Imag

esSt

ored

/Tim

e[m

in−

1] Apanel = 8 cm

2

Apanel = 28.5 cm2

Apanel = 42 cm2

Cost vs performance trade-offs

Dynamic Energy Burst Scaling (DEBS) canreduce the capacitance and the cost per image.

NVMH can reduce the cost per image further,but requires a larger capacitance.

Long-term logging using MSP430FR

Evaluate cost/performance of different EMUconfigurations for a long-term logging app.For each, there is an associated:

- Minimum energy guarantee (capacitance).

- Cost for storing an image.

- NVMH was configured with Nimg=10.

...FRAMAcquisition

Burst 1

SD init SD flush

Burst Nimg

FRAMAcquisition

Burst Nimg+1

weara

ble

TEG

ind

ust.

pie

zo

ou

tdoor

sola

rin

door

sola

r

kin

eti

c

acti

ve

deep

sle

ep

sle

ep

ult

rason

icra

ng

ete

mp

.h

um

idit

y

vis

ion

cam

era

therm

al

cam

era

accel.

gas

LoR

a

Zig

bee

BLE