I was just looking at the ATTiny13A on Digi-Key, the one in an 8-pin SOIC. It costs 33–40¢, depending on quantity, has six GPIO pins, runs at anything up to 20 MHz (or 0.128, 4.8, or 9.6 MHz on the internal RC oscillator), and can source or sink 20 mA per output pin, up to a max of 60 mA across all the pins. Its input pin bias current is 1 μA. It has the usual assortment of AVR peripherals, cut down a bit: one 8-bit timer/counter with two PWM channels, a separate watchdog, a 10-bit ADC with an internal voltage reference (same speed as the one on ATMega328, 15ksps at max resolution), and a comparator. It only has a kilobyte of Flash and 64 bytes of RAM. It’s 5.4 mm × 5.35 mm × 2.2 mm in the SOIC version.
So far this sounds like a pretty wimpy microcontroller, even if it is an AVR. But I think maybe instead of comparing it to microcontrollers I should compare it to transistors.
(Incidentally, the same chip also comes in a much smaller MLF package with 20 pins, which gives you 18 GPIOs instead of 6. And the ATTiny5 costs 17¢ now, but is limited to half the memory, 12MHz, and 8 bits of ADC resolution, but has a 16-bit timer — in 2 mm × 2 mm.)
(And the STM32F031x4/6 comes in a WLCSP25 package that is 2.5 mm × 2.4 mm and has 20 GPIOs, or 18 GPIOs and an I²C interface, or 16 GPIOs and an SPI interface. It can handle somewhat less current or voltage than the AVRs but is not at all wimpy, rather the opposite. I think it costs 56¢ but I may have it confused with a larger package of the same chip.)
For example, a 2N7002T-2 MOSFET costs US$0.51 (or down to 9.4¢ in quantity 1000), about 30% more than the microcontroller (again, in quantity 1). It can switch 115 mA at up to 60 volts, a much higher voltage than the microcontroller but less than twice the current, and its input bias current is 0.1 μA, ten times higher. It can do analog things the AVR can’t. But you need 10 volts to turn it all the way on, it can only reasonably be used up to a few megahertz, and its biggest disadvantage is that it has only one input and one output.
Depending on the circumstances, you could replace as many as six 2N7002s (US$3) with a single ATTiny13A (US$0.40). That’s nearly an order of magnitude cheaper, and it’s smaller too.
But maybe it’s unfair to compare the microcontroller to power MOSFETs, even lightweight ones like the 2N7002. Maybe you should be comparing it to small-signal transistors, like a 300MHz MMBT3904, which only costs 10¢ and can handle 200 mA at up to 40 V, or a 12¢ BC849, which is 30V and 100mA, which is good to 100MHz. But those still cost more per output pin than the microcontroller. And typically they come in a SOT-23 package, which is 2.9 mm × 1.3 mm × 1 mm, larger than many of the microcontroller packages mentioned above.
It seems like if you just want to turn things on and off, even very fast, the microcontroller is best up to 60 mA at 5 volts — 300 milliwatts — and perhaps 5 megahertz out, 7 kilohertz analog in. If you need higher voltage, higher current, analog output, or higher frequency, then maybe you should go with discrete transistors or perhaps other ICs.
This suggests a real potential niche for microcontrollers with open-drain GPIO outputs, which could potentially directly switch much higher-power loads — I suspect 60 V at 20 mA wouldn’t be unreasonable, and would be over a watt without dissipating any more power in the μC itself than an ATTiny does. Hmm, and any I²C/TWI interface already has open-drain hardware... I wonder if it’s possible to take advantage of this?
The STM32 line of microcontrollers has 5V-tolerant GPIO pins that can be configured as open-drain, but the datasheet warns not to try to run them over 5.5V, and gives Vdd+4V as the “absolute maximum”. So 60 V is probably unwise.
An interesting question is how you design things to make this potential advantage real. What kind of software do you run on the microcontroller? How much autonomy do you give it? How do you divide up the functionality of a complex device among different microcontrollers?
A difficulty with these super tiny devices with 4 or 6 GPIOs is that you don’t have much left over for communication. An ATtiny48, by contrast, has 24 or 28 GPIOs (depending on the package), and three of them suffice for SPI at 3 Mbps. In between 21 remaining GPIOs, you could plausibly charlieplex 441 LEDs, which would be a pretty decent chunk of transistors, or run 10 tiny DC motors in an H-bridge configuration. These packages, not counting the DIP, range from 7 mm square to 4 mm square.
Other possible chips are reasonable alternatives to microcontrollers. There are I²C GPIO extenders, for example, and there are op-amp chips, and there are shift registers, and maybe you could use JTAG. The LM7321 op-amp comes in a SOT-23 (1.75 mm wide, 3.05 mm long, not counting the pins) or dual in a 8-pin VSSOP (3 mm square) or SOIC.
An amusing possible alternative is to use some kind of voltage regulator — Horowitz & Hill suggest treating an LM317 as a cheaper alternative to the LM395 “protected transistor”, with “-1.2 V base-to-emitter voltage and 50 μA pullup base current,” and mention using 7805s in a similar way despite their annoyingly large base-to-emitter voltage. But you should also be able to use a buck or boost converter in this way, generating a sort of class-D amplifier.
GPIO extenders include the Microchip MCP23008, which Digi-Key classifies as “Interface — I/O Expanders”, a category with 1643 chips in it. These are mostly SPI and I²C with 8 or 16 pins and can sink or source a few tens of mA. They tend to cost about 50–100¢. Examples include the 129¢ TI TCA9534APWR, a 400kHz I²C 16-TSSOP (4.4mm square) with 8 push-pull GPIOs that runs anywhere from 1.65–5 V; the 144¢ Semtech SX1502I087TRT, a 400kHz I²C 20QFN (3mm square) with 8 push-pull GPIOs of 8mA source, 24mA sink capacity at anywhere from 1.2 to 5.5 V; and the 150¢ TI TCA9535PWR, a 400kHz I²C 24TSSOP (4.4mm) with 16 push-pull GPIOs of 10 mA source, 25 mA sink at anywhere from 1.65 to 5.5 V; the MCP23008-E/ML, a 113¢ I²C 20QFN (4 mm square) with 8 push-pull GPIOs of 25 mA source or sink at anywhere from 1.8V to 5.5V; and the obsolete 5MHz SPI NXP PCA9701D,112, with 16 input-only pins, in a 7.5mm 24-SOIC. (All prices from Digi-Key in quantity 1.)
The thing to notice here is that the price per GPIO here is 16¢, 18¢, 9.4¢, 14¢, and ∞¢, respectively. While this is reasonable compared to the price of a whole discrete transistor, it compares unfavorably with the price of GPIOs on a full-fledged microcontroller. And the size, too, compares unfavorably.