Today, with off-the-shelf microcontrollers, you could build a computer, somewhat superior in speed and screen size to a Sun 4, that runs off a solar cell from a solar calculator. Memory is somewhat trickier. The hardware BOM cost is about US$30 or US$40.
Computational power would not be the power bottleneck, even with a fairly powerful CPU; rather, memory power (to erase Flash especially) and display power are likely to outweigh the CPU.
First, a word on the class of machine that desktop software was designed for.
The Unix workstation revolution was sparked by what Raj Reddy called the “3M computers”: one MIPS, one megapixel, one megabyte of RAM, for under a megapenny. This was what you needed for a PARC-style GUI. The SUN (later called the Sun-1) was one of these, as were the NeXT, the Apollo, and the doomed PERQ. These machines all came with dozens of megabytes of disk.
The diskless Sun-3/60 I used to use as an X terminal is about 3 MIPS (InfoWorld, 1988-11-14, “Sun Shows Tempest Versions of Sun-3/4 Workstations”, by Scott Mace) with an 1152×900 display (I think it has a bwtwo and a cgfour in it, but I think I was using a monochrome monitor). By the end of the 1980s, various Sun 4 models (based on the 1987 SPARC family) exceeded 10 MIPS; the contemporary DECStation 3100 was 20 MIPS, with DEC forcing its competitors down to a price of a thousand dollars per MIPS. Some DECStations at UNM provided internet access to 50 or 100 concurrent users when I started using them around 1994.
By that time, much more than a megabyte of RAM was commonplace on these workstations. There was a memory-price bubble from about 1992 to about 1996 during which a RAM cartel held prices steady at around US$40 per megabyte before finally collapsing to the Moore’s-law trend line at about US$10 per megabyte; before that, memory was halving in price every couple of years, so, in 1990, it was around US$80 per megabyte. My Sun-3/60 from 1988 has, IIRC, a full complement of 48 megs of RAM, which at US$160 per megabyte would have cost Sun about US$8000, a bit out of line with the cost of the rest of the machine — but I suspect that RAM was added later.
But 16 megs would be an entirely reasonable amount.
So, our target is 10 MIPS, 1 megapixel, 16 megabytes of RAM, and let’s say a gigabyte of disk.
An E-ink display needs about 1 mW on average to keep it updated for continuous comfortable reading. (This number comes from the Keyboard-powered computers, which says that a 6" E-Ink display needs about 750 mW during a 120 ms screen update, for 122×91 mm at 167 dpi, 190 nJ/pixel, 8100 nJ/mm²; 8 20-em lines of 7-point text work out to 49 mm × 20 mm, or 8 mJ per update, and about 50 words, so about 0.16 mJ per word; at 350 words per minute that’s almost a milliwatt. Larger text is more power-hungry.)
They can be a little hard to come by; the Seeed Studio one I found on DX is AR$390 (US$26) and only 200×96 pixels at 111 dpi (57 mm × 29 mm), enough for 12 lines of 40 characters. Old reflective passive-matrix feature phone replacement displays might be a viable alternative, and they are very cheap if you can find them (US$3, say). The Nokia 2730 display costs AR$78 (US$5) on MercadoLibre at the moment; it’s QVGA resolution (which I guess is 320×240) and does 200 dpi 18-bit color. More recent smartphone displays are probably too power-hungry.
Various kinds of off-the-shelf low-power microcontrollers consume a bit less than a nanojoule per instruction, which has been stable for a decade or more and might finally start trending down again.
Processing power can scale down smoothly with available energy up to a point. Atmel picoPower ARMs, which go up to 48MHz (at 1 insn/cycle), do 1–10 μW in standby (and only 250 pJ/insn) but take 4–20 μs to wake up. Some MSP430s are below 0.3 μW idle, while others are in the same range as the Atmel chips, but wake up much faster. (These numbers come from file low-power-micros in this repo.)
ARMs are ubiquitous these days, and they’re power-efficient and do a lot of work per clock cycle and per instruction. However, it might make bootstrapping quicker to use a different processor architecture, one that has existing self-hosted assemblers, compilers, and development environments available as relics from the homebrew personal computer era that have some chance of running in the very limited RAM on a microcontroller. That basically means 6502, 8080, or 8086, and nowadays that basically means 8080, specifically Z80.
Zilog — now a fabless semiconductor house — still makes Z80s (now the “eZ80” and “ZGATE” lines) for microcontroller applications. Toshiba has a Z80-compatible microcontroller line called the 870/C. Unfortunately all of these are fairly overpriced and underpowered. A typical part is the EZ80F93AZ020SG, which costs US$6.14, runs at 20MHz, and has 64K of Flash and 4K of on-chip RAM, but supports off-chip RAM too. It’s not designed for low-power operation — it sucks 180 milliwatts when not sleeping, working out to a rather hefty 9 nanojoules per instruction, assuming one instruction per cycle. Even Zilog is switching to ARMs nowadays.
For a while, many MP3 players used Z80-derivative DSP chips like the ATJ2085; some of them were called "S1mp3", and an alternative firmware was developed for them. Digi-Key doesn't carry the ATJ2085.
Nowadays it probably makes more sense to try to run a Z80 emulator in software on an ARM core than to run an actual Z80. An LPC4310FET100.551 runs you US$7.74 from Digi-Key, has 168 kB of RAM, is ROMless (it has a bootloader in ROM that loads code from an external NOR Flash or other device at startup), runs its two ARM cores at 204MHz, and runs down to 2.2 volts, and even then it uses only 160 milliwatts, less than the EZ80F93AZ020SG.
I walked by a cheap-shit electronics store today and saw a bunch of small solar calculators for AR$40 (US$2.60). It occurred to me to wonder how much computational power you could run off one of their solar cells.
I think they use inefficient amorphous silicon photovoltaic cells rather than the standard 16%-efficient polycrystalline silicon used for rooftop and utility-scale solar panels, and the cells look like they’re about 9 mm × 36 mm. This means about 320 milliwatts of sunlight falls on them even in direct sun. If we suppose it's 9% efficient, as Wikipedia’s “Thin-film cell” article says of production thin-film amorphous (silicon?) cells, then that’s 29 mW, which at the 1.7-V a-Si bandgap voltage would be 17 mA.
At a nanojoule per instruction, 29 mW is 29 MIPS!
You have: 9 mm * 36 mm * 1000 W/m^2 * 9% / nanojoule
You want: MHz
* 29.16
Can you extend the usefulness of a wimpy solar cell with some solid-state energy storage?
Off-the-shelf ceramic capacitors can store in the range of 1–4 mJ; you can get a 1 mJ 6.3 V 47 μF ceramic capacitor, holding 1 mJ, for US$0.12. If you were to try to stuff the calculator with US$2.60 worth of these, you would have 21 mJ of storage, enough for about 21 million instructions (or 84 million on the picoPower chips). This is enough for a minute or so of CPU word processing, but it’s not enough to update the display.
I bought some solar garden lights at the supermarket a few months ago for AR$40 (about US$2.90 I think; XXX check this). They have a 300 mAh 1.2 V NiCd AA cell inside powering a white LED and a polycrystalline solar cell on top that’s about 40 mm × 40 mm; they claim it provides light for about 8 hours, although I haven’t tested how much current it takes. It probably doesn’t have a MPPT controller (the only component visible on the tiny PCB is a through-hole ⅛ W resistor, although there’s a tiny blob of epoxy that might be hiding something). But it probably doesn’t need one, in the sense that the cell should produce about a quarter watt in full sun, which would fully charge the cell in 1.4 hours:
You have: (300 milliamp hours 1.2 volts) / (40 mm * 40 mm * 1000 W/m^2 * 16%)
You want: hours
* 1.40625
So it’s probably limited by the battery, not the solar cell, so it probably fully charges the battery to its full charge of 1.3 kJ (!!), which is 1.3 trillion nanojoules and thus about 1.3 trillion instructions, which would be 1.3 million seconds (about two weeks) at 1 MIPS, or about two weeks of continuous E-ink reading. ¼ W at 1 nJ/insn is of course 250 MIPS.
Unfortunately, the microcontrollers I mentioned above are plenty fast, have very little memory. A typical example is the NXP MKW01Z128 mentioned in License-free femtowatt UHF radio transceiver ICs under a μJ per bit: a 48 MHz 32-bit RISC in-order Cortex-M0 CPU with a built-in 600 kbps license-free RF transceiver, with 128 KiB of Flash and 16 KiB of RAM. (It uses about 1.2 nanojoules per instruction if it’s receiving data at the same time.)
For many year, it has been normal for a desktop computer to have about as much RAM as it can access in a second, which works out to have been very roughly a megabyte per megahertz. You will note that this microcontroller is nowhere close to this balance, with 48 MHz (and about 48 MIPS) but 0.144 megabytes, a factor of some 300 away from the balance for a desktop computer. This is a very common problem with current microcontrollers.
At this point, old 36-pin and 72-pin SIMMs are e-waste that can be freely scavenged or bought for prices in the neighborhood of a dollar or less, typically ranging from 4 MB to 32 MB. But these run on 5 V and typically have standby currents of a few milliamps and operating currents of over an amp, which means they use on the order of 5000 mW running at full speed and tens of milliwatts just to retain data. So they aren’t really an option.
Chips to look at:
https://www.digikey.com/product-detail/en/issi-integrated-silicon-solution-inc/IS42S16100H-7TL/706-1446-ND/5683868 https://www.digikey.com/product-detail/en/issi-integrated-silicon-solution-inc/IS62WV1288BLL-55HLI/706-1046-ND/1555419 https://www.digikey.com/product-detail/en/microchip-technology/23LC1024-I-SN/23LC1024-I-SN-ND/3543084 https://www.digikey.com/product-detail/en/microchip-technology/23LCV1024-I-SN/23LCV1024-I-SN-ND/3543093 https://www.digikey.com/product-detail/en/on-semiconductor/N01S830HAT22I/N01S830HAT22IOS-ND/6166720 https://www.digikey.com/product-detail/en/fujitsu-electronics-america-inc/MB85RS1MTPNF-G-JNERE1/865-1255-1-ND/4022688 https://www.digikey.com/product-detail/en/rohm-semiconductor/MSM51V17405F-60T3-K/MSM51V17405F-60T3-K-ND/2695010 https://www.digikey.com/product-detail/en/adesto-technologies/AT25SF041-SSHD-T/1265-1131-1-ND/4824165 https://www.digikey.com/product-detail/en/issi-integrated-silicon-solution-inc/IS25LQ080-JNLE-TR/706-1463-1-ND/5872437 https://www.digikey.com/product-detail/en/fremont-micro-devices-usa/FT25H04S-RT/1219-1190-1-ND/5875686 https://www.digikey.com/product-detail/en/winbond-electronics/W25X40CLZPIG/W25X40CLZPIG-ND/3008616 https://en.wikipedia.org/wiki/Programmable_metallization_cell https://www.pjrc.com/mp3/simm/simm.html
There are reasonably inexpensive SRAM chips out there that can provide a megabyte or two of RAM for a few dollars, but nothing bigger. DRAM is thousands of times cheaper (right now Pricewatch lists DDR3-1600 DIMMs at US$30 for 8 GiB, under US$4/GiB, which would be 6¼¢/16 MiB) but very power- hungry. Even just retaining data in DRAM uses a lot of power.
(You will note that US$4/GiB today in 2016 and US$40/MiB in 1992 is a factor of ten thousand in 24 years, which is a bit over 13 doublings — a doubling time of about 22 months, reasonably close to the Moore’s trend line I mentioned.)
Flash memory, by contrast, can typically retain data for ten years while turned off and dissipating only tiny amounts of energy it had stored while turned on, and always at least a few months. And it's even cheaper than DRAM: Pricewatch lists 64GB SDHC cards at US$18, or US$0.28 per GB, almost 15 times cheaper than DRAM. It even uses a little less power to read than SRAM or DRAM do. But erasing Flash is enormously more power-hungry than erasing SRAM or DRAM.
(Other kinds of nonvolatile memory — FeRAM/FRAM, MRAM, CBRAM, PCM RAM — have broadly similar characteristics to Flash.)
It’s tempting to think of Flash as having similar access characteristics to a hard disk, but that’s only really true for erasing. Reading written data is very fast, and can typically be done a byte at a time only a few times slower than reading entire blocks. This means that with the appropriate data structures, Flash can replace much of the RAM in our 1990 workstation, as well as its disk.
Erasing and writing data to Flash is about a thousand times slower than reading it.
Writing to erased blocks is somewhat fast, Even when you have to write by pages, they’re 256-byte or 264-byte pages, and writing them takes about 700 microseconds — you can do 1400 pages per second on a single cheap Flash chip, compared to the 50 random spinning-rust disk writes you could do in 1990, or the 120 you can do nowadays. Sequentially, you could write a megabyte per second (2048 pages) to pre-erased disk at the time, or a third of a megabyte per second (1400 pages) to pre-erased Flash now. And these Flash chips are so cheap that it makes sense to use several of them, multiplying your bandwidth and allowing block erases to be concurrent with reads and writes.
Many Flash chips have bit-serial interfaces, which makes them much easier to hook up to a microcontroller. None of the SRAM or DRAM chips I’ve been able to find on Digi-Key have bit-serial interfaces, which means that driving them from a cheap microcontroller is going to require hooking up some kind of shift register to control the address bus.
So maybe we can get by with only a megabyte of real RAM and, say, 32 gigabytes of Flash.
There’s a distinction I’m ignoring here, which is that the Flash that’s fast to read a byte from is NOR, and a megabyte of it costs about US$0.46. So 32 gigabytes of it would be about US$15000, which is significantly more than the US$9 of NAND Flash. In practice you probably want all three types: maybe a couple megabytes of real RAM (US$8), 16 megabytes of NOR Flash (US$7), and 32 gigabytes of NAND Flash (another US$7).