(See also Keyboard-powered computers.)
E-ink displays are capable of supporting continuous, comfortable reading at average power levels under 3 milliwatts. For this to be useful, we need a microcontroller to drive them at comparable power. But 5–20 watts might be available sometimes, and we’d like to be able to take advantage of it.
To build a computer that provides rapid UI response on very low average power (<10mW, >100 seconds per joule) I think you need, among other things, multiple processors: a low-speed, low-power “peripheral processor”, probably designed as an embedded microcontroller, for normal real-time interaction, plus a higher-speed “central processor” that powers up occasionally, or when lots of power is available, to handle heavier-duty computation.
Simply slinging pixel data around is a minimum; at 1 bit per pixel and 400 pixels per character (due to high-res screens), you need to send 50 bytes of data to the framebuffer to draw a single letter or space. That’s probably about 100 instructions, so 100 instructions/ glyph × 350 wpm × 6 glyphs / word = 3500 instructions/second. But that's like a theoretical minimum.
WordStar on a 2MHz (≈0.5MIPS) 8-bit CPU would sometimes fall behind your typing a bit. This suggests that one (8-bit) MIPS is adequate to provide real-time word processing.
A nanojoule (1000 pJ) per instruction is a microwatt per MIPS, which would be about 0.56 μA/MIPS (the unit microcontroller vendors like to use) at 1.8V. That’s also 1000 giga-ops-per-second per watt. Low-power microcontrollers are usually 2–5 times better than this.
Atmel ARMs can do 250 pJ per 32-bit instruction and 2μW in standby. However, they take forever to wake up, like tens of microseconds (and a microsecond is 12 to 48 instructions!). A full reboot from OFF takes 2.2 milliseconds.
At this energy cost, 10W would pay for 40 billion instructions per second, but that would require nearly a thousand microcontrollers.
These are Cortex-M0+s with single-cycle hardware (32×32!) multipliers. And also a tiny four-cell FPGA called “CCL”, “configurable custom logic”.
http://www.atmel.com/technologies/lowpower/picopower.aspx makes it hard to tell how low power can go. They say 1.62V.
http://www.atmel.com/Images/doc8349.pdf gives an example of how to lower power consumption by lowering clock frequency from 8MHz to 1.8432MHz, transmitting data at 115,200 baud instead of 19,200 baud, lowering the brownout voltage from the usual 2.7V to 1.8V, and going to sleep, increasing run-time off a capacitor from 6 seconds to 217 seconds. However, they never specify the capacitance!
http://www.atmel.com/products/microcontrollers/arm/sam-l.aspx#saml21 says “power consumption down to 35 µA/MHz in active mode and 200nA in Sleep mode”. http://www.atmel.com/Images/Atmel-42385-SAM-L21_Datasheet.pdf says 48MHz (the Performance Level 2 model) or 12MHz (the PL0 model), 256kB flash, 40kB SRAM, 1.62 to 3.63 V, and typically 25 to 100 μA/MHz with the regulator in buck mode, lower for PL0 than PL2. 75 μA/MHz at 3.3V in the PL0 model seems like the worst-case performance. Normalizing that, we get 250pJ/inst. Also gives numbers in the 1 to 10 μA for STANDBY and BACKUP states, depending on temperature, and a bit lower for OFF.
More easily available is the https://www.digikey.com/product-detail/en/ATSAMD10D13A-MUT/ATSAMD10D13A-MUTCT-ND/5226477 Atmel ATSAMD10D13A-MUT (US$2.24); datasheet at http://www.atmel.com/Images/Atmel-42242-SAM-D10_Datasheet.pdf. 8 kB flash, 4 kB RAM, 22 GPIOs. Worst case 75 μA/MHz (+284μA), which normalizes to the same 250pJ/inst. Wakeup time is 4–20μs.
See also Notes on the STM32 microcontroller family, but the STM32L low-power ARM microcontroller series is supposedly 144 pJ/insn, and can reasonably do duty cycles down to 0.02%, while using 2 mA at 1.8 V at 16 MHz, and the STM32F0 range is 250 pJ/insn.
The familiar ATMega328 used in the Arduino isn’t the highest-tech or lowest-power microcontroller out there, but it’s very well known, and it’s simple to get running. It runs at up to 20MHz, with most instructions in a single cycle, but only 8-bit ALU operations. In power-save mode, with a real-time clock enabled, it uses 0.75μA (at 1.8V, so 1.4 μW); at 8MHz and 5V, it uses up to 12mA, so 60mW, working out to 7500 pJ/instruction. This is about 30 times worse than the AT-SAM family and MSP430 family, even before you take into account that these are 8-bit operations.
Consequently, at 20MHz, this processor will gobble 150mW.
The von Neumann MSP430 family is in the range of 300–500 pJ per instruction, 10–25 16-bit MIPS, and 220–4000nW idle, which means instruction consumption exceeds idle consumption above about three instructions per second or one wakeup per minute. At a power budget of 5mW, an MSP430 could average 12 16-bit MIPS.
These processors are generally the processors of choice for very-low-duty-cycle applications.
http://www.ti.com/lit/wp/slay015/slay015.pdf TI in 2012 explaining why they think MSP430 is better at low-power than Microchip XLP.
http://www.greenarraychips.com/home/documents/greg/WP003-100617-msp430.pdf says that in 2010 the F18 consumed 7 pJ per instruction or 450 pJ per 17×17-bit multiply; an MSP430 (at 8MHz) consumed 330 pJ per instruction or 2310 pJ per 16×16 multiply. It also gives a sleep/wakeup time of 4 ns for the F18, with an idle power consumption of 100 nW, and a sleep/wakeup time of 5μs for the MSP430, with an idle (LPM4 sleep state) power consumption of 3600 nW.
http://johann-glaser.blogspot.com.ar/2012/10/msp430-launchpad-with-debian.html says that an MSP430G2231 has 2kB flash, 128 bytes RAM, and runs at 16MHz (performing 16-bit operations). (Also, it gives a nice howto for getting a TI Launchpad device working with Debian.)
http://www.ti.com/tool/msp430-gcc-opensource is TI’s deprecated but actively maintained port of GCC.
https://en.wikipedia.org/wiki/TI_MSP430A says the FRAM series of MSP430s has 320 nA RAM retention and 82 μA/MIPS, which is half the power consumption of other MSP430s, with up to 2 kB of RAM. Other MSP430s have up to 512kB of ROM or 10kB of RAM. Normalizing these numbers using 3.3V, which ought to be about right, we get 271 pJ/insn and 1100 nW. The MSP430G2xx series has sub-microsecond wakeup, 0.1μA RAM retention, and 220μA/MIPS at 2.2V, or 480 pJ/insn and 220nW idle.
http://www.eit.lth.se/fileadmin/eit/courses/eit095f/Hanson_Variability_J_2008.pfd describes an 8-bit processor that works at 350 mV and uses 3.5 pJ/instruction, two orders of magnitude less than the MSP430. https://web.eecs.umich.edu/~taustin/papers/VLSI-subliminal.pdf describes it again more briefly. https://web.eecs.umich.edu/~taustin/papers/VLSI06-sublim.pdf describes another one that reaches 2.6 pJ/inst at 833kHz. None of these processors are commercially available.
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.25.1097&rep=rep1&type=pdf says the CoolRisc 81 uses 22 pJ/inst. In 2001!
http://icri-ci.technion.ac.il/files/2013/07/Karlheinz-Meier_red2.pdf Neurons use about 10 fJ (0.01 pJ) per neural spike (action potential) per synapse. It is not yet feasible to use them to build a portable computer.
A solar-powered computer could still be somewhat useful at night if it can store some energy. But standard batteries are flaky and break after a few years, so fuck them. And we’re talking about a super-low-power device, one that might use under 10mW when in active use, so even storing 1 joule of energy is potentially useful (100 seconds), 10 joules could last you 20 minutes, and 100 joules could last you three hours.
But how can you store 100 joules without batteries? I haven’t found a good way.
If you can reduce the power consumption below 1mW, then a smaller amount of energy storage would be feasible.
http://www.mouser.com/pdfDocs/Cymbet-WP-EnerChips-vs-Supercaps.pdf claims that supercapacitors discharge at 10%–20% per day, while EnerChips (theoretically on the market but not in stock at Digi-Key) discharge at 1%–2% per year.
These solid-state batteries use a thin-film solid with ionic conduction as their electrolyte, so they are fully solid-state, despite the mobility of ions within them.
http://www.nature.com/nmat/journal/v14/n10/full/nmat4369.html discusses some new research on solid-state electrolytes for lithium batteries in more detail. http://news.mit.edu/2015/solid-state-rechargeable-batteries-safer-longer-lasting-0817 is a press release about the paper.
Solid-state capacitors are the physically stablest medium for storing a few joules of energy for short periods of time. They are low in capacity and expensive. To store 1J, a capacitor needs:
| E < | C > |
| 3V | 0.2 F |
| 10V | 20000 μF |
| 50V | 800 μF |
| 100V | 200 μF |
| 300V | 22 μF |
| 1000V | 2 μF |
Ceramic capacitors generally do not have the necessary energy capacity.
https://www.digikey.com/product-detail/en/C3216X5R0J476M160AC/445-1428-2-ND/569054 TDK Corporation C3216X5R0J476M160AC is 47μF, 6.3V, 3.20mm x 1.60mm, US$0.12. CV²/2 = 1 mJ.
https://www.digikey.com/product-detail/en/AMK432BJ477MM-T/587-4368-1-ND/5405565 Taiyo Yuden AMK432BJ477MM-T is 470μF, 4V, 4.50mm x 3.20mm, US$5.34. CV²/2 = 4 mJ.
https://www.digikey.com/product-detail/en/JMK325ABJ227MM-T/587-3980-1-ND/4950534 Taiyo Yuden JMK325ABJ227MM-T is 220μF, 6.3V, 3.20mm x 2.50mm, US$2.56. CV²/2 = 4 mJ.
https://www.digikey.com/product-detail/en/B58033I5206M001/495-6706-5-ND/5039882 EPCOS (TDK) B58033I5206M001 is 20μF, 500V, 33.00mm x 22.00mm, 31g, US$89.90. CV²/2 = 2.5J. Designed for industrial power inverters.
Some film capacitors are designed for high voltages and consequently have substantial energy capacity. You could hook up the output side of an ordinary switchmode power supply (the part after the rectifier) designed for 240VAC line current to many of these, because in normal use a 240VAC-input SMPS’s capacitors will be charged up to 340V. http://www.righto.com/2014/05/a-look-inside-ipad-chargers-pricey.html gives an overview of these power supplies, and points out that you can get a crappy one with 5V output for US$3. http://www.onsemi.com/pub_link/Collateral/SMPSRM-D.PDF explains a bit about how you could charge them from a low-voltage source such as solar panels.
https://www.digikey.com/product-detail/en/DS371506-CA/P9669-ND/821966 Panasonic Electronic Components DS371506-CA is 50μF, 370V, 50.50mm diameter × 112.00mm, US$7.87. CV²/2 = 3.4 J. This capacitor is designed for shunting a motor. https://media.digikey.com/pdf/Data%20Sheets/Panasonic%20Capacitors%20PDFs/AC%20Film%20Caps.pdf is the datasheet; specifies insulation resistance >1000 MΩ (giving τ > 14h), explains it's an “oil wound metallized polypropylene film” capacitor.
https://www.digikey.com/product-detail/en/B32778G1276K/495-3933-ND/1884954 EPCOS (TDK) B32778G1276K is 27μF, 1300V, 57.50mm × 35.00mm × 50.00mm, US$14.48. CV²/2 = 23 J. http://en.tdk.eu/inf/20/20/db/fc_2009/MKP_B32774_778.pdf is the datasheet; gives lifetime expectancy of 100k hours (11 years) to 1M hours (110 years) at room temperature, but insulation resistance such that τ is only 10k seconds (3 hours).
https://www.digikey.com/product-detail/en/C44AFGP6200ZE0J/399-5955-ND/2704609 Kemet C44AFGP6200ZE0J is 200μF, 400V, 76.00mm diameter × 140.00mm, US$55.44. CV²/2 = 16 J.
https://www.digikey.com/product-detail/en/B25620B1317K322/B25620B1317K322-ND/3489148 EPCOS (TDK) B25620B1317K322 is 310μF, 1320V, 116.00mm diameter × 103.00mm, US$82.83 (in quantity 16!). CV²/2 = 270 J.
A few tantalum capacitors have reasonable energy capacity.
https://www.digikey.com/product-detail/en/T491B476K010AT/399-9728-2-ND/3724805 Kemet T491B476K010AT is 47 μF, 10V, 3.50mm x 2.80mm, US$0.12. CV²/2 = 2 mJ.
https://www.digikey.com/product-detail/en/T491X476K035AT/399-3821-2-ND/818681 Kemet T491X476K035AT is 47 μF, 35V, 7.30mm x 4.30mm, US$0.94 (in quantity 500!). CV²/2 = 30mJ.
https://www.digikey.com/product-detail/en/109D107X9060F2/718-1226-ND/1559882 Vishay Sprague 109D107X9060F2 is 100μF, 60V, 7.92mm x 20.22mm, US$41.91. CV²/2 = 180mJ.
https://www.digikey.com/product-detail/en/TWDE503M006CB0Z0700/478-9404-ND/4990384 AVX Corporation TWDE503M006CB0Z0700 is supposedly 50 millifarads, but actually mislabeled; it’s 50 microfarads. 6.3V, 9.52mm x 26.97mm, US$69.20. CV²/2 = 1 mJ. Also “wet tantalum”.
https://www.digikey.com/product-detail/en/592D228X06R3X8T20H/592D228X06R3X8T20H-ND/2802652 Vishay Sprague 592D228X06R3X8T20H is 2200μF, 6.3V, 14.50mm x 7.37mm, US$3.69 (in quantity 500!). CV²/2 = 40mJ.
https://www.digikey.com/product-detail/en/TWAE228K025CBSZ0000/478-4966-ND/1879539 AVX Corporation TWAE228K025CBSZ0000 is 2200 μF, 25V, 9.52mm x 26.97mm, US$102.73, wet tantalum. CV²/2 = 700mJ.
https://www.digikey.com/product-detail/en/M39006%2F22-0660/1012-1025-MIL/2773731 Vishay Sprague [MIL] M39006/22-0660 is 56 μF, 125V, 10.31mm x 28.60mm, US$76.61. CV²/2 = 400mJ.
https://www.digikey.com/product-detail/en/TWAE157K125SBDZ0000/478-7301-ND/3451926 AVX Corporation TWAE157K125SBDZ0000 is 150 μF, 125V, 10.31mm x 26.97mm, US$87.20. CV²/2 = 1 J.
These are newish (like, the last 30 years). I’m not sure they can be trusted to last a long time; the numbers I’m seeing on Digi-Key’s page are like “1000 hours at 70°”. http://www.murata.com/~/media/webrenewal/products/capacitor/edlc/techguide/electrical/c2m1cxs-053.ashx talks about the failure modes; it says they degrade little by little over time, although even after 5 years their 4.2V 470 mF supercap holds 2 J; they claim that their flat packages do a better job of keeping moisture out and electrolyte in than cylindrical packages, showing a degradation to 80% of capacity after 3000h (4 months). It also mentions that leakage current is around a microamp. On p.20 it shows a dryup lifetime graph, showing that even their DMT products will dry up in only 20 years at 70°, while their DMF products will dry up in only 5 years at 40°; on p.23 it explains that at 5.5V, their DMF will break after only 500 hours at 70° or 6400 hours (9 months) at 40°, while the DMT will last dramatically longer, like 40,000 hours at 70°.
https://www.digikey.com/product-search/en?pv13=1538&FV=fff40002%2Cfff8000c&mnonly=0&newproducts=0&ColumnSort=0&page=1&quantity=0&ptm=0&fid=0&pageSize=500 Nichicon JJD0E608MSEH is 6000 farads, 2.5V, 76.20mm × 168.00mm, US$326. CV²/2 = 18750 J. This sounds like science fiction.
https://www.digikey.com/product-detail/en/JUWT1105MCD/493-4330-ND/2538684 Nichicon JUWT1105MCD is 1F, 2.7V, 6.30mm diameter × 10.50mm, US$0.86. CV²/2 = 3.6 J.
https://www.digikey.com/product-detail/en/DSK-3R3H224U-HL/604-1020-1-ND/970232 Elna America DSK-3R3H224U-HL is 0.22F, 3.3V, 6.80mm × 3.00mm, US$2.16. CV²/2 = 1 J.
https://www.sparkfun.com/products/10319 is a 24.5 mm × 3.0 mm 3.6V LIR2450 rechargeable coin cell, rated for 110 mAh, which is 1430 J. It costs US$3. https://electronics.stackexchange.com/questions/218655/running-3-3v-mcu-from-lir2032-lithium-ion-button-cell cites a LIR2032 (20 mm × 3.0 mm) cell with 40 mAh at 3–4.2 V, 520 J.
Horowitz & Hill (the 1989 edition with the micropower chapter) says that commonplace batteries (9-volt and AA batteries) “give nearly their full shelf life at drain currents less than 20μA.” At 2 V and 200 pJ/insn (reasonable ballpark figures, though the STM32L does better) this works out to an average of 200 000 instructions per second. At 1400mAh for an ordinary 1.5 AA cell, of which you would need two (or a boost converter), this is 6 years, which is indeed comparable to the battery’s shelf life. At 1 MIPS, the two-cell battery only lasts a year.
I bought a couple of rechargeable NiMH AA batteries the other day for some absurd price, like US$10 or something. They claim 2000 mAh, 1.2V, which is 8.6 kJ (per cell). At the full 3.5 mW power of a 16 MHz STM32L at 1.8V (see Notes on the STM32 microcontroller family), that would run for 28 days.
A cell of the https://en.wikipedia.org/wiki/Zamboni_pile yields 0.8 volts, but only a few nA, and the Clarendon Dry Pile has been ringing its bell since 1840. The cell is perhaps 20mm in diameter, but <1mm thick; you could easily stack three of them to get 2.4 volts, perfectly adequate to run a microcontroller on without regulation. In particular, the Clarendon Dry Pile produces about 1 nA at 2 kV, consuming about 1 μg from each of its cells over the last 144 years, according to Croft’s 1984 paper, “The Oxford electric bell.” He doesn’t give dimensions but the dry pile seems to be about 30mm in diameter, i.e. 700mm².
However, at least STM32s have a minimal current consumption of a few hundred nA (180 nA in standby mode with no RTC, 410 nA in standby mode with RTC). This would require, say, 500 times the cross-sectional area of the Clarendon battery: perhaps a 3-meter roll of 120mm-wide battery. If the total thickness were 1mm, the battery would be 350 mℓ. This seems like a small price to pay for a battery life measured in centuries, as long as the device doesn’t have to be portable.
ISSI’s IS62C256AL-45ULI-TR is 256 kilobits of parallel 5V 45ns SRAM for US$1.43 down to US$1.11 in quantity 500. It uses 150 μW, or up to 15 mW in operation.
Microchip’s 23K256-I/SN is 256 kilobits of SRAM in a little bitty 3.3V 8-pin 20MHz SPI SOIC for US$1.09 down to US$1.05. It uses 4 μA max for standby, which works out to 13.2 μW.
ISSI’s IS62/65WVS2568FALL / IS62/65WVS2568FBLL is similar, 256 kiloBYTES of SRAM in a 3.3V SPI/SDI/SQI SOIC for US$2.20, but only available in quantity, running on 4 μA for standby.
The Adesto AT25SF041-SSHD-T is four megabits of 104MHz 2.5–3.6V Flash in the same 8-pin SOIC format, and it's much cheaper, US$0.36. Being Flash, it uses no power for “standby” but a lot of power to erase (10 mA for 500 ms to erase 32 kB ≈ 500 nJ/byte), but “only” 4 mA to read. It has quad outputs, so 104 MHz is pretty decent speed actually.
ISSI’s IS62C1024AL-35QLI-TR is a megabit of parallel 5V 35ns SRAM for US$2.10 down to US$1.62 in quantity 500. It uses 20 μW, or up to 100 mW in operation.
In PSRAM, ISSI’s IS66WV51216EBLL-55TLI is 8 megabits of parallel 55ns 2.5–3.6V DRAM, with all the refresh nonsense hidden inside the chip, for US$3.20 down to US$2.31 in quantity 1000. Because of the refresh nonsense, it uses more power, but not as much as you’d think — in CMOS standby mode (where presumably it retains the data) it uses 100 μA, which works out to 250–360 μW. That is about 80 times as much as the Microchip part, but it also has about 32 times as much memory, so the penalty isn’t that bad.