Millimeter-scale Computing

Transcript

1 Millimeter-Scale Computing Dennis Sylvester University of Michigan Joint work with David Blaauw and many excellent PhD students University of Michigan 1 1 Modern Computing Landscape  Cloud 2  Social networking, Google Docs  Mobile  Netbooks, tablets, smart phones  Sea of sensors  Ubiquitous computing  The next explosion University of Michigan 2 2 The Challenge & Opportunity 3  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? University of Michigan 3 3 Ex App: Continuous IOP Monitoring 4 Iris Optic Nerve Cornea High Intraocular Pressure Anterior Chamber Lens Retina Implant Location  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?? University of Michigan 4 4 Bell’s Law 5 Mainframe Inflation Adjusted Price (1000s of USD) 100000 Workstation New class of computing systems every decade 10000 1000 Laptop 100 10 1 [Bell et al. Computer, 1971, Bell, ACM, 2008] Mini Computer Personal Computer 0.1 Smartphone 0.01 1950 University of Michigan 1960 1970 1980 1990 2000 2010 2020 5 5 Bell’s Law – Corollary 6 Mainframe 10T 1T Workstation 100G 100x smaller every decade [Nakagawa08] 10G 1G Size (mm3) 100M Laptop 10M 1M 100k 10k 1k Mini Computer 100 Personal Computer 10 1 Smartphone 100m 1950 University of Michigan 1960 1970 1980 1990 2000 2010 2020 6 6 Bell’s Law – Production Volume 7 Mainframe 1 per Enterprise 10T 1T Workstation 100G 1 per Engineer 10G 100x smaller every decade [Nakagawa08] 1G Size (mm3) 100M Laptop 10M 1 per Professional 1M 100k 10k 1k 100 Mini Computer 1 per Company Personal Computer 10 1 per Family 1 1 per person Smartphone 100m 1950 University of Michigan 1960 1970 1980 1990 2000 2010 2020 7 7 Bell’s Law – Production Volume 8 Mainframe 1 per Enterprise 10T 1T Workstation 100G 1 per Engineer 10G 100x smaller every decade [Nakagawa08] 1G Size (mm3) 100M Laptop 10M 1 per Professional 1M mm-Scale Computing 100k 10k 1k 100 Mini Computer Ubiquitous 1 per Company Personal Computer 10 1 per Family 1 Ubiquitous 1 per person Smartphone 100m 1950 University of Michigan 1960 1970 1980 1990 2000 2010 2020 8 8 mm-Computing: Application Areas 9 Surveillance and micro robotics Medical mm-Scale Computing Environment Infrastructure Textiles University of Michigan 9 9 Where Are We? 100000 10 Mica2mote Mica2Dot [Roggen06]# [Hollar00] 10000 Intel mote# [Chee08] [Park06] Size (mm3) 1000 [Nakagawa08] [Pister99] [Pister01] [Hill03] 100 [Blaauw10] 10 [Sylvester11] 3 1 Goal - 1mm complete sensor Kris Pister coins “Smart Dust” term 0.1 1998 2000 2002 2004 2006 2008 2010 2012 Timeline University of Michigan 10 10    Long device lifetime vs. small form factor The 3 most important things in miniaturization Circuits just are not there yet Power Budget, W Why aren’t we further along? 11 1k minute 100 hour 10 1 day week 100m month 10m year 1m decade 100µ 10µ 1µ 100n 10n 1 mm3 Harvester 1n 100p 10p 101 102 103 104 105 106 107 108 109 1010 Desired Lifetime, s University of Michigan 11 11 Status of Miniature Sensors   12 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 ~10,000x 10,238:1 5hrs of lifetime University of Michigan 1:1 5 years of lifetime 12 12 Battery Trends  13 New battery chemistries are rare  Li-ion ≤7%/year expected improvement (energy density)  Energy capacity limited by safety and cost [Broussely04] 1896 1956 Columbia Eveready Dry Cell Alkaline Zinc Carbon Battery Battery ~72mAh University of Michigan 1960 Zinc Carbon Battery 1100 mAh 1961 Nickel CadmiumB attery 1100 mAh 1989 Lithium Iron Disulfide Battery 3100 mAh 1992 Rechargeable Alkaline Battery 2000 mAh 13 2005 LSD NiMH Battery 2300 mAh 13 Energy Harvesting 14 Capacitive Vibration Harvesting Temperature Gradient Capacitive Piezoelectric Vibration Harvesting Solar µW/cm3 Photovoltaic (outside) 15,000# Air flow 380 Vibration 200 Temperature 40# Pressure Var. 17 Photovoltaic (inside) 10# # Microturbine Air Flow fundamental metric is µW/cm2 [Courtesy: Jan Rabaey, S. Roundy] University of Michigan 14 14 Harvesting Improvements Limited   15 Energy harvester efficiency gains are modest Fundamentally limited by harvesting source University of Michigan 15 15 Past Michigan Sensor Designs processor 244µm 16 Subliminal 1 Design (2006) -0.13 µm CMOS -investigate Vmin -2.60 µW/MHz 122µm memory 305µm Phoenix 1 Design (2008) - 0.18 µm CMOS - Minimize sleep current - 2.8 µW/MHz / 30pW sleep power 181µm Subliminal 2 Design (2007) - 0.13 µm CMOS - Study variability effects - 3.5 µW/MHz Phoenix 2 Design (2010) - 0.18 µm CMOS - Commercial ARM M3 Core - Solar harvesting / PMU -28 µW/MHz IOPM (2011) - 0.18 µm CMOS - MEMS/CDC - Solar / PMU - Wireless comm EE Times 20 Hot Technologies for 2012 University of Michigan 16 16 mm3: How Do We Get There   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 Power Management Sensors and Front End Microprocessor Memory Wakeup Wakeup Controller / Controller Timers Wireless Communication MEMS Sensors Power Sources University of Michigan Antenna 17 17 How?        Near-threshold computing is the key   Microprocessors/microcontrollers Timers Static memories CMOS image sensors Voltage references Signal processing cores 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 University of Michigan Vopt Energy / Operation Best reported energy efficiencies for: Vbal Vmax ~5-10X ~2X Log (Delay)  18 ~3 -10X 0 Vnom Vth Supply Voltage 18 18 mm3: How Do We Get There   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 Power Management Sensors and Front End Microprocessor Memory Wakeup Wakeup Controller Controller/ Timers Wireless Communication MEMS Sensors Power Sources University of Michigan Antenna 19 19 Ex: Why Timers Are So Important Idle Sensor measure Power Consumption (W) 1m Base Station Sensor Node 20 TX / RX 1 ms 100µ 10µ 100 ms 1µ 100n 10n 1n 100p 0 20 40 Time (min) 60 Power consumptions in various modes of ULP sensor  Asymmetric RF communication does not require precise timing University of Michigan 20 20 1 Why Timers Are So Important Idle Sensor measure Power Consumption (W) 1m Base Station Sensor Node 21 TX / RX 1 ms 100µ TX/RX 11% Meas. 10% 10µ 100 ms 1µ Idle 79% 100n 10n 1n Energy 100p 0 20 40 Time (min) 60 0.91 µJ/hr Power consumptions in various modes of ULP sensor  Asymmetric RF communication does not require precise timing University of Michigan 21 21 1 Why Timers Are So Important Power Consumption (W) Idle Sensor Measure Sensor Node Sensor Node 1m 100µ 10µ 1µ 100n 10n 1n 100p 0 1m 100µ 10µ 1µ 100n 10n 1n 100p 0 20 40 60 20 40 60 22 TX / RX Synch. Mismatch Timer Time (min) Power consumptions in various modes of ULP sensor    Asymmetric RF communication does not require precise timing Symmetric RF communication requires precise timing Energy penalty for mismatch can dominate energy budget University of Michigan 22 22 1 Why Timers Are So Important Power Consumption (W) Idle Sensor Measure Sensor Node Sensor Node 1m 100µ 10µ 1µ 100n 10n 1n 100p 0 1m 100µ 10µ 1µ 100n 10n 1n 100p 0 Mismatch (1s) 20 40 60 20 40 60 23 TX / RX Synch. Mismatch Timer Time (min) Power consumptions in various modes of ULP sensor    Asymmetric RF communication does not require precise timing Symmetric RF communication requires precise timing Energy penalty for mismatch can dominate energy budget University of Michigan 23 23 1 Why Timers Are So Important Power Consumption (W) Idle Sensor Measure Sensor Node Sensor Node 1m 100µ 10µ 1µ 100n 10n 1n 100p 0 1m 100µ 10µ 1µ 100n 10n 1n 100p 0 Mismatch (1s) 20 40 60 24 TX / RX Synch. Mismatch Timer Synch. Mismatch 97% Energy 103 µJ / hr 20 40 60 Time (min) Power consumptions in various modes of ULP sensor    Asymmetric RF communication does not require precise timing Symmetric RF communication requires precise timing Energy penalty for mismatch can dominate energy budget University of Michigan 24 1 24 Keeping Time with Picowatts   Crystal oscillators bulky and power hungry RC oscillators preferable, exhibit accuracy vs. power tradeoff 25 Low power commercial crystal oscillator ~400nW, 5x3mm [Micro Crystal Switzerland RV-2123-C2] University of Michigan 25 25 Keeping Time with Picowatts   Crystal oscillators bulky and power hungry RC oscillators preferable, exhibit accuracy vs. power tradeoff 26 Low power commercial crystal oscillator ~400nW, 5x3mm [Micro Crystal Switzerland RV-2123-C2] Gate leakage current based timer Vinv 1pW! MS4 MC1 MI4 MC2 MS6 MS3 Vin Vin Vs MS2 vx MS5 MS1 University of Michigan ML1 Vs MI3 Vout INV1 MI2 INV2 TINV 26 Vclk MI1 26 Putting it together: 1.5mm3 microsystem 27  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 University of Michigan 27 27 IOP Monitor Power Budget    28 Measure IOP every 15 minutes DSP with 10k processor cycles @ 100 kHz per measurement Daily wireless transmission of 1344b raw IOP data Power Time/Day Energy/Day 19.2 sec 134.4 µsec 19.2 sec 134.8 µJ Transceiver SCVR 7.0 µW 47.0 mW 116.9 nW • µP @ 100 kHz 90.0 nW 19.2 sec 1.7 µJ Standby Mode Power Time/Day Energy/Day CDC Transceiver 172.8 pW 3.3 nW 24 hours 24 hours 14.9 µJ 285.1 µJ SCVR 174.8 pW 9.8 pW 24 hours 24 hours 62.0 pW 24 hours 15.1 µJ 846.7 nJ 5.2 µJ Active Mode CDC • 4kb SRAM • WUC 6.3 µJ 2.2 µJ 5.3 nW average power  1 month lifetime with no harvesting University of Michigan 28 28 Other Biosensor Applications   29 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 Wireless data transfer Implanted Biosensors (pressure, pH)  Earlier indicator than size of tumor (6 – 9 weeks)  Also: Smart orthodontics, intracranial pressure, others University of Michigan 29 29 Power from the Bottom  30 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 University of Michigan 30 30 Status: M3 Michigan Micro Mote University of Michigan 31 31 31 Conclusions  32 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 Battery and Solar Harvesting 1mm3 Platform Solar Harvesting Imaging Processor Imager Memory Motion Detect Timer system from bottom up Processing and Wireless Chip in cells? University of Michigan New Applications Wireless Communication Biology Medicine Pika ICP 32 Security 32