Sleep
Maybe the most obvious strategy for power saving is for the device to be in a low power"sleep" state as much as possible, assuming it cannot actually be turned off completely. You obviously need a means of waking it which can be periodically by time or an event (i.e. interrupt) such as a button push or physical movement. You then have various parameters to juggle with such as how often the device wakes up and what it does when it wakes up. You can also have dynamic sleep intervals whereby the device sleeps for varying times depending what is happening. For example, with a fitness monitor you may want to monitor and log data quite frequently when there is activity but when the activity slows or stops, the monitor interval could be increased, thus saving battery power.
Inter-IC communications
I2C has pullup resistors which waste power - SPI doesn't need pullup resistors. Pin capacitance consumes power between ICs. Minimize the amount of data you transfer if you can. If there are four lines of 5pF each running at 20MHz with a 3.3V supply, it will draw 660μA due to pin capacitance alone (current = 0.5CVf where C is the pin capacitance, V is the supply voltage and f is the frequency). The pin capacitance will be both the sending and receiving all the IC capacitance combined. This current may be dwarfed by the actual IC current, but not always - it depends on the function of the device. It is a contributory factor to why highly integrated chips can be the most efficient - access to peripherals, RAM and flash memory is all internal so no pin capacitance is involved when accessing those peripherals or memory. A large part of pin capacitance is due to the ESD protection diodes.
Power Supply Efficiencies
Choosing a switching regulator for a switched mode power supply is a key factor in maximizing efficiency, particularly synchronous regulators where efficiencies of over 95% are possible. However, it is not just headline efficiency or even standby efficiency that is necessarily the most critical. It is necessary to look at the current in different modes for the device and determine the contribution to overall power consumption from each mode after taking account of the switching regulator efficiency at each current level. There are some quite impressive regulators around though, such as the new Analog Devices ADP5301 Step-down Regulators. The quiescent current is down as low as 180nA when not-switching but still operating in hysteresis mode. It will switch for a short burst to add charge to the output capacitor using the inductor at very light loads then return to just the quiescent current. The low quiescent current can give you efficiencies as high as 80% at 1μA depending on the input and output voltages. You are more likely to see lower figures than that optimum value, but still above 40%. It also delivers up to 0.5A and has a single pin programmable output with a fixed resistor. It is very impressive compared to older regulators which would take a few milliamps with no load.
If you are using a switching regulator with an external MOSFET, bear in mind that the MOSFET switching time can result in significant losses. The transition from non-conducting to conducting is the time when a switching MOSFET dissipates the most power. When it is turned fully on the voltage drop will usually be very small and hence power dissipation will be low. However, partly turned on there will be a significant voltage drop across the MOSFET accompanied by significant current. You therefore want to minimize the time that transistor spends in that state by choosing a fast switching device and low gate capacitance. You obviously want a low ON resistance as well.
Power supply shutdown
See if you can keep power supply capacitors small if the power supplies are shut down in sleep mode. It takes energy to charge them and if the power supply is shut down when in a sleep mode then the energy in the capacitors is wasted energy (unless it will still be there when you need it next). For example, a 1μF capacitor on the power supply of circuitry which is shut down 100 times per second will consume 165μA at 3.3V (same calculation as before). Many ICs will take less than that in shutdown or sleep mode, so it is often better to keep circuitry powered but in a sleep state than to actually do power switching to save power. The exception would be if the device you were using didn't have a sleep mode or if its sleep mode was not very low current. If you can use 100nF instead of 1μF you could save a lot of energy.
Low supply voltage
Devices will consume less power at lower voltage even if they don't consume less current. So, if a microcontroller is powered by 1.8V instead of 3.3V power consumption will be half for the same current. Usually digital devices will also consume less current at lower voltage as well so the power is further reduced. Watch out for the maximum clock speed also reducing though - it is not uncommon for the maximum clock speed to be lower at lower voltages. So, while the current will be lower it will take twice as long to run the microcontroller's code.
For example, while the Microchip nanoWatt XLP PIC24F16KA102 microcontroller running at 2MHz consumes 695μA at 3.3V, it only requires 363μA at 1.8V, which is 70% less power - a massive saving. However, at 3.3V the microcontroller can run at 32MHz, while at 1.8V the maximum clock speed is 8MHz.
Clock Speeds
Choose the clock speed to suit the application, not just the fastest possible. This probably won't be known until the code is finished. Most microcontrollers have an adjustable clock multiplier, allowing the clock speed to be changed by the application code. If the microcontroller code is the limiting factor in terms of executing the code and going back to sleep, then the fastest may be the most efficient. However, if something else will actually be slowing things down and the microcontroller will be "marking time", the microcontroller can be woken by a periodic interrupt instead.
An alternative approach would be to dramatically slow down the processor when it is just "waiting" and then speed it back up when there is something to do. Ideally you would sleep instead but clock startup times and the power wasted while waiting for the clock to stabilize can mean sleep is not the best option in some circumstances.
Clock Startup
Power is wasted while waiting for the system clock to stabilize. If code can be run while the clock is still settling then that can help, but if there is any other part of the circuitry relying on a stable, accurate clock the processor has to wait, resulting in wasted energy. Some microcontroller manufacturers' clocks are quicker to start than others.
32-bit Microcontrollers
Does every design really need a 32-bit processor? They seem to be everywhere but the work actually being done by them can be minimal. A 16-bit or 8-bit microcontroller may be more efficient. It does depend on what code is running and the efficiency of the compiler. If there is a Bluetooth or Internet interface then more memory may be required and a 16-bit or 32-bit processor should be used. An TCP/IP stack with a web server requires a 32-bit processor. Creative software writing can minimize the resources needed and minimize power consumption.
Wider data width processors consume more power in a number of ways. Accessing 32-bit RAM and Flash memory is more power hungry than accessing 16-bit memory. Also, leakage current increases with wider memory. That points towards keeping memory size to a minimum both by efficient code structuring and writing, and also not choosing a processor with massive amounts of excess memory.
RF power
If there is a RF wireless interface, e.g. Bluetooth Smart, consider the transmitting distance. Not only are there power savings by transmitting with lower power, the receiver sections can have adjustable sensitivity and will take less power when sensitivity is set lower. For example, the Nordic Semiconductor nRF52832 has a high receiver sensitivity mode taking 10.9mA but normal sensitivity only takes 6.1mA.
Custom IC
The ultimate lower power solution might be a full custom IC design with only the needed circuitry. Circuitry that never drives off-chip consumes a lot less power. However, it is the slowest and most expensive way of developing a product. It is also probably why a Frederique Constant Smartwatch battery lasts 2 to 3 years and a Fitbit Flex lasts 3 to 5 days.
Summary
There are a lot of aspects to consider when designing a low power system. It can be a good idea to create a spreadsheet with all the parts of the system listed, current consumption, required duty cycles, voltages and total power consumed. Then for each design scenario calculate the projected power consumption. The only way to find out some of the information needed is to actually build something and test it.
