millimeter-scale computing - tti/vanguard · millimeter-scale computing . ... 1950 1960 1970 1980...
TRANSCRIPT
1 1
1
University of Michigan 1
Millimeter-Scale Computing
Dennis Sylvester
University of Michigan
Joint work with David Blaauw and many excellent PhD students
2 2
2
University of Michigan 2
Modern Computing Landscape Cloud Social networking,
Google Docs Mobile Netbooks, tablets,
smart phones Sea of sensors Ubiquitous
computing The next explosion
3 3
3
University of Michigan 3
The Challenge & Opportunity Tiny nearly invisible computing devices will change
health care, security, infrastructure and environmental monitoring Hey – that’s Smart Dust, right?!
Very hard to achieve Size vs. lifetime
We are close to delivering on this vision What are the killer apps? How small can we make them?
4 4
4
University of Michigan 4
Retina
Lens
Implant Location
Optic Nerve
AnteriorChamber
High Intraocular Pressure
Iris
Cornea
Ex App: Continuous IOP Monitoring
Glaucoma Second leading cause of blindness; affects 60M Progress checked by measuring intraocular pressure
Continuous monitoring with an implanted microsystem Gives doctors a more complete view of disease Faster response time for tailoring treatments
How many more applications are out there??
5 5
5
University of Michigan 5
1950 1960 1970 1980 1990 2000 2010 20200.01
0.1
1
10
100
1000
10000
100000
Infla
tion
Adju
sted
Pric
e (1
000s
of U
SD)
Bell’s Law Mainframe
Mini Computer
Personal Computer
Workstation
Smartphone
Laptop
[Bell et al. Computer, 1971, Bell, ACM, 2008]
New class of computing
systems every decade
6 6
6
University of Michigan 6
Bell’s Law – Corollary Mainframe
Mini Computer
Personal Computer
Workstation
Smartphone
100x smaller every decade [Nakagawa08]
Laptop
1950 1960 1970 1980 1990 2000 2010 2020100m
110
1001k
10k100k
1M10M
100M1G
10G100G
1T10T
Size
(mm
3 )
7 7
7
University of Michigan 7 1950 1960 1970 1980 1990 2000 2010 2020
100m1
101001k
10k100k
1M10M
100M1G
10G100G
1T10T
Size
(mm
3 )Bell’s Law – Production Volume
Mainframe
Mini Computer
Personal Computer
Workstation
Smartphone
100x smaller every decade [Nakagawa08]
1 per Enterprise
1 per Company
1 per Professional
1 per person 1 per Family
1 per Engineer
Laptop
8 8
8
University of Michigan 8 1950 1960 1970 1980 1990 2000 2010 2020
100m1
101001k
10k100k
1M10M
100M1G
10G100G
1T10T
Size
(mm
3 )Bell’s Law – Production Volume
Mainframe
Mini Computer
Personal Computer
Workstation
Smartphone
100x smaller every decade [Nakagawa08]
mm-Scale Computing
1 per Enterprise
1 per Company
1 per Professional
1 per person
Ubiquitous
1 per Family
1 per Engineer
Laptop
Ubiquitous
9 9
9
University of Michigan 9
mm-Computing: Application Areas
Medical Surveillance and micro robotics
mm-Scale Computing
Infrastructure
Textiles
Environment
10 10
10
University of Michigan 10
1998 2000 2002 2004 2006 2008 2010 20120.1
1
10
100
1000
10000
100000Si
ze (m
m3 )
Where Are We?
[Pister99]
[Sylvester11]
[Blaauw10]
[Pister01]
Mica2Dot
[Park06]
[Roggen06]#
Timeline
[Nakagawa08]
[Hill03]
[Chee08]
Mica2mote
[Hollar00]
Goal - 1mm3 complete sensor
Kris Pister coins “Smart Dust” term
Intel mote#
11 11
11
University of Michigan 11
Why aren’t we further along? Long device lifetime
vs. small form factor The 3 most important
things in miniaturization Circuits just are not
there yet
101 102 103 104 105 106 107 108 109 101010p
100p1n
10n100n
1µ10µ
100µ1m
10m100m
110
1001k
decade
hour
Powe
r Bud
get,
W
Desired Lifetime, s
minute
weekday
yearmonth
1 mm3 Harvester
12 12
12
University of Michigan 12
Status of Miniature Sensors Where are we now? 1cc in the research labs 30-50cc commercial motes
Goal: 1mm3 sensor node 1000x improvement Enables host of new applications Tiny wireless sensor nodes that
can be distributed anywhere and everywhere
Challenge: battery size
10,238:1 5hrs of lifetime
1:1 5 years of lifetime
~10,000x
13 13
13
University of Michigan 13
Battery Trends New battery chemistries
are rare Li-ion ≤7%/year expected
improvement (energy density)
Energy capacity limited by safety and cost
1956 Eveready Alkaline Battery ~72mAh
1992 Rechargeable
Alkaline Battery
2000 mAh
1989 Lithium
Iron Disulfide Battery
3100 mAh
1896 Columbia Dry Cell
Zinc Carbon Battery
2005 LSD NiMH
Battery 2300 mAh
1961 Nickel
CadmiumBattery
1100 mAh
1960 Zinc
Carbon Battery
1100 mAh
[Broussely04]
14 14
14
University of Michigan 14
Energy Harvesting
µW/cm3
Photovoltaic (outside)
15,000#
Air flow 380
Vibration 200
Temperature 40#
Pressure Var. 17
Photovoltaic (inside)
10#
[Courtesy: Jan Rabaey, S. Roundy]
Microturbine
Air Flow
Solar
Capacitive Vibration
Harvesting
Capacitive Vibration
Harvesting
Temperature Gradient
Piezoelectric
# fundamental metric is µW/cm2
15 15
15
University of Michigan 15
Harvesting Improvements Limited Energy harvester efficiency gains are modest Fundamentally limited by harvesting source
16 16
16
University of Michigan 16
Past Michigan Sensor Designs
Phoenix 2 Design (2010) - 0.18 µm CMOS
- Commercial ARM M3 Core - Solar harvesting / PMU
-28 µW/MHz
Subliminal 2 Design (2007) - 0.13 µm CMOS
- Study variability effects - 3.5 µW/MHz
Subliminal 1 Design (2006) -0.13 µm CMOS -investigate Vmin
-2.60 µW/MHz
Phoenix 1 Design (2008) - 0.18 µm CMOS
- Minimize sleep current - 2.8 µW/MHz / 30pW sleep power
processor
memory
244µm
305µm
122µm
181µm
processor
memory
244µm
305µm
122µm
181µm
IOPM (2011) - 0.18 µm CMOS
- MEMS/CDC - Solar / PMU
- Wireless comm
EE Times 20 Hot Technologies for 2012
17 17
17
University of Michigan 17
Microsystem functions include sensing, processing, storage, and transmission All components must be re-examined to fit within power
envelope defined by power sources and power management
mm3: How Do We Get There
Sensors and Front End
Microprocessor Memory
Wireless Communication
Power Management
Wakeup Controller
Antenna
Power Sources
MEMS Sensors
Wakeup Controller /
Timers
18 18
18
University of Michigan 18
How? Best reported energy
efficiencies for: Microprocessors/microcontrollers Timers Static memories CMOS image sensors Voltage references Signal processing cores
Near-threshold computing is the key Voltage scaling in CMOS circuits
must be re-established
Tradeoffs abound Speed, area, jitter, SNR, etc. We seek 10X reductions in power/
energy with reasonable tradeoffs
Vopt
Ene
rgy
/ Ope
ratio
n Lo
g (D
elay
)
Supply Voltage 0 Vth Vnom
~3 -10X
~2X
~5-10X
Vbal Vmax
19 19
19
University of Michigan 19
Microsystem functions include sensing, processing, storage, and transmission All components must be re-examined to fit within power
envelope defined by power sources and power management
mm3: How Do We Get There
Sensors and Front End
Microprocessor Memory
Wireless Communication
Power Management
Wakeup Controller
Antenna
Power Sources
MEMS Sensors
Wakeup Controller /
Timers
20 20
20
University of Michigan 20
100p
1n
10n
100n
1µ
10µ
100µ
1m
Powe
r Con
sum
ptio
n (W
)
Time (min)0 20 40 60
100 ms
1 ms
Ex: Why Timers Are So Important
Power consumptions in various modes of ULP sensor
1
Base Station
Sensor Node
IdleSensor measure
TX / RX
Asymmetric RF communication does not require precise timing
21 21
21
University of Michigan 21
100p
1n
10n
100n
1µ
10µ
100µ
1m
Powe
r Con
sum
ptio
n (W
)
Time (min)0 20 40 60
100 ms
1 ms
Why Timers Are So Important
Power consumptions in various modes of ULP sensor
1
Base Station
Sensor Node
IdleSensor measure
TX / RX
Meas. 10%
TX/RX 11%
Idle 79%
Energy 0.91 µJ/hr
Asymmetric RF communication does not require precise timing
22 22
22
University of Michigan 22
100p1n
10n100n
1µ10µ
100µ1m
Powe
r Con
sum
ptio
n (W
)
Time (min)
100p1n
10n100n
1µ10µ
100µ1m
0 20 40 60
0 20 40 60
IdleSensor Measure Synch. Mismatch
TX / RX
Timer
Why Timers Are So Important
Power consumptions in various modes of ULP sensor
1
Sensor Node
Sensor Node
Asymmetric RF communication does not require precise timing Symmetric RF communication requires precise timing Energy penalty for mismatch can dominate energy budget
23 23
23
University of Michigan 23
100p1n
10n100n
1µ10µ
100µ1m
Powe
r Con
sum
ptio
n (W
)
Time (min)
100p1n
10n100n
1µ10µ
100µ1m
0 20 40 60
20 40 60
Mismatch (1s)
0
IdleSensor Measure Synch. Mismatch
TX / RX
Timer
Why Timers Are So Important
Power consumptions in various modes of ULP sensor
1
Sensor Node
Sensor Node
Asymmetric RF communication does not require precise timing Symmetric RF communication requires precise timing Energy penalty for mismatch can dominate energy budget
24 24
24
University of Michigan 24
100p1n
10n100n
1µ10µ
100µ1m
Powe
r Con
sum
ptio
n (W
)
Time (min)
100p1n
10n100n
1µ10µ
100µ1m
0 20 40 60
20 40 60
Mismatch (1s)
0
IdleSensor Measure Synch. Mismatch
TX / RX
Timer
Why Timers Are So Important
Power consumptions in various modes of ULP sensor
Synch. Mismatch
97%
Energy 103 µJ / hr
1
Sensor Node
Sensor Node
Asymmetric RF communication does not require precise timing Symmetric RF communication requires precise timing Energy penalty for mismatch can dominate energy budget
25 25
25
University of Michigan 25
Keeping Time with Picowatts Crystal oscillators bulky and
power hungry RC oscillators preferable,
exhibit accuracy vs. power tradeoff
Low power commercial crystal oscillator ~400nW, 5x3mm
[Micro Crystal Switzerland RV-2123-C2]
26 26
26
University of Michigan 26
Keeping Time with Picowatts Crystal oscillators bulky and
power hungry RC oscillators preferable,
exhibit accuracy vs. power tradeoff
Vin Vs
Vclk
ML1
MS1
MS2
MS3
MS4
MS5
MS6MC1
INV1 INV2
TINV
Vinv
Vin Vs Vout
MC2
vx
MI1
MI2
MI3
MI4
Gate leakage current based
timer 1pW!
Low power commercial crystal oscillator ~400nW, 5x3mm
[Micro Crystal Switzerland RV-2123-C2]
27 27
27
University of Michigan 27
Putting it together: 1.5mm3 microsystem Continuous intraocular pressure monitoring Wireless communication Energy-autonomy Device components
• Solar cell • Wireless transceiver • Cap to digital converter • Processor and memory • Power delivery • Thin-film Li battery • MEMS capacitive sensor • Biocompatible housing
28 28
28
University of Michigan 28
Active Mode PowerCDC 7.0 µWTransceiver 47.0 mW
• µP @ 100 kHz 90.0 nW
Energy/Day
6.3 µJ
1.7 µJSCVR 116.9 nW 2.2 µJ
Standby Mode PowerCDC 172.8 pWTransceiver 3.3 nW
• WUC 62.0 pW
SCVR 174.8 pW• 4kb SRAM 9.8 pW
Energy/Day14.9 µJ
846.7 nJ15.1 µJ
5.2 µJ
Time/Day19.2 sec
134.4 µsec
19.2 sec19.2 sec
Time/Day24 hours24 hours
24 hours24 hours
24 hours
134.8 µJ
285.1 µJ
IOP Monitor Power Budget Measure IOP every 15 minutes DSP with 10k processor cycles @ 100 kHz per measurement Daily wireless transmission of 1344b raw IOP data
5.3 nW average power 1 month lifetime with no harvesting
29 29
29
University of Michigan 29
Other Biosensor Applications Pressure is an early biomarker
for tumor health Implant sensor during biopsy 1mm size makes delivery
through same needle feasible
Track pressure to determine response of tumor to chemotherapy Adjust medication after 1 – 2
weeks if necessary Earlier indicator than size of
tumor (6 – 9 weeks)
Also: Smart orthodontics, intracranial pressure, others
Implanted Biosensors
(pressure, pH)
Wireless data transfer
30 30
30
University of Michigan 30
Power from the Bottom Miniature Benthic Microbial
Fuel Cells Generate electricity from
microbes in marine sediment – anaerobic conditions Low harvest: 14uW/cm2
Traditionally requires large deployments – at high cost mm-sensor node
enables a small “dart” – cheap and fast deployment
UUVs, etc Work with SPAWAR
32 32
32
University of Michigan 32
1mm3 Platform
Battery and Solar Harvesting
Processor
Memory
Timer
MotionDetect
Imager
Wireless Communication
ImagingSolar
Harvesting
Processingand Wireless
New Applications
Biology
Pika
Medicine
ICP
Security
Conclusions Applications of mm-scale computing are endless and often
unimaginable today But first the hardware must get there (which it is)
Power minimization is paramount Few nW avg power 1m comm range ~Indefinite lifetime Re-think entire sensor
system from bottom up
Chip in cells?