Power and energy decide how hot a chip runs, how long a battery lasts and how much a data center pays for electricity. This first part covers why power matters, the difference between power and energy, the three components of CMOS power, and the main techniques designers use to cut each one. The material follows the power chapter of the University of Waterloo ECE 327 course notes by Mark Aagaard, credited below.
Why power and energy matter
For phones, laptops and other battery devices the reason is obvious: every milliwatt saved is battery life gained. The case is just as strong for plugged-in equipment. The ECE 327 notes, written in the early 2000s, give some memorable examples from that period:
- For PC microprocessors, every watt above about 40 W was estimated to add roughly $1 to manufacturing cost, because of the heat sink, fan and packaging needed to remove it.
- Energy was put at about a quarter of the operating expense of a server farm.
- Sandia National Laboratories had to build a dedicated electrical substation for its Teraflops supercomputer, built from more than 9,000 Pentium Pro processors.
- Fast processors of that era could run hot enough to damage themselves, which is why the Pentium 4 used thermal throttling and early Athlon parts had reliability problems with overheating.
- The notes even predicted "power viruses": malicious code that keeps a phone running at full power to drain its battery.
The specific dollar figures belong to their time, but the pressure has only grown since.
Power is not the same as energy
People say "power reduction" for two different goals:
- Power (watts) is about heat. Peak and average power set the package, heat sink, cooling and supply design.
- Energy (joules) is about battery life and electricity bills. It is power multiplied by the time a task takes.
The two can move in opposite directions. Halving the clock frequency halves dynamic power, but the task takes twice as long, so the dynamic energy for that task stays the same, and leakage energy actually goes up because the chip is on for longer.
Where the power goes
Total CMOS power has three parts:
- Switching power: charging and discharging the capacitance of wires and transistor gates every time a signal changes. This is the useful work of computing.
- Short-circuit power: during a transition both the pull-up (PMOS) and pull-down (NMOS) transistors are briefly on, so current flows straight from supply to ground. It does no useful work.
- Leakage power: current that leaks through transistors that are meant to be off. It also does no useful work.
Switching and short-circuit power together make up dynamic power, which depends on clock speed and on how often signals change. Leakage is static power, which is drawn whether or not the clock is running.
How often signals change is captured by the activity factor. The ECE 327 notes use a rule of thumb that roughly 20% of signals switch in a given clock cycle. Glitches, the extra transitions that happen while a combinational output settles, raise the activity factor and must be counted.
How designers reduce power, in outline
Each technique attacks one or more terms of the power equation.
Circuit and technology level (what the notes call analog): lower capacitance, for example with silicon-on-insulator (SOI) processes; lower wire resistance, for example with copper interconnect; and lower supply voltage.
- Dual VDD: a higher supply for performance-critical logic and a lower one elsewhere, or a supply that changes over time depending on the workload.
- Dual Vt: low-threshold transistors, which are fast but leaky, only on critical paths, and high-threshold transistors, which are slower but leak less, everywhere else.
- Special circuits: flops, latches and logic styles designed to run fast at low power.
- Adiabatic circuits: circuits that recover charge instead of dumping it to ground, trading speed for lower power.
- Clock tree design: the notes put clock generation and distribution at up to 30% of total power, so the clock network deserves its own attention.
Logic and architecture level (what the notes call digital): fewer gates (less capacitance), lower activity and lower clock frequency.
- Multiple clocks: a fast clock only where performance needs it, a slower clock elsewhere.
- Clock gating: stop the clock to blocks that are idle.
- Data encoding: choose Gray, one-hot or binary encodings to reduce the number of bits that toggle.
- Glitch reduction: balance path delays or restructure logic so outputs do not toggle several times per cycle.
- Asynchronous circuits: remove the global clock altogether.
Deep dive: the numbers and details
This part holds the equations, a worked example and a technique-by-technique summary. Skip it if the overview is all you need.
The power equations
Following the ECE 327 notes, with activity factor defined as the average number of 0 to 1 or 1 to 0 transitions per signal per clock cycle:
P_total = P_switch + P_short + P_leak
P_switch = 1/2 x ActFact x f_clk x C_load x V_dd^2
P_short = ActFact x f_clk x t_short x I_short x V_dd
P_leak = I_leak x V_dd
The factor of one half comes from the energy of a capacitor. Charging C from 0 to V draws CV2 from the supply: half is stored in the capacitor and half is lost as heat in the pull-up transistor. Discharging dumps the stored half through the pull-down. So each transition costs CV2/2.
Texas Instruments writes the same physics for standard logic parts in terms of a power-dissipation capacitance Cpd plus the external load, and a static term PS = VCC x ICC; its application report also notes that typical CMOS logic data sheets specify a maximum ICC of 10 to 40 µA (TI, CMOS Power Consumption and Cpd Calculation, SCAA035B).
Worked example: one block
Take a logic block with 2 nF of total switched capacitance, an activity factor of 0.2, a 500 MHz clock and a 1.0 V supply (illustrative numbers):
P_switch = 0.5 x 0.2 x 500e6 x 2e-9 x 1.0^2 = 0.100 W
Drop the supply to 0.9 V and nothing else changes: 0.100 x 0.81 = 0.081 W, a 19% saving from a 10% voltage cut. That quadratic term is why voltage is the strongest lever.
Now the clock network. A clock rises and falls once every cycle, so its activity factor is 2 in this convention. A 100 pF clock net at 500 MHz and 1.0 V:
P_clock = 0.5 x 2 x 500e6 x 100e-12 x 1.0^2 = 0.050 W
A clock net with only one-twentieth of the logic's capacitance burns half as much power as all of that logic, because it switches ten times as often. This is how clock trees reach the large share of total power quoted above, and why clock gating pays off.
The voltage and delay trade-off
Lower voltage slows the circuit, because there is less drive current to charge each load. The ECE 327 notes use the approximation:
delay ~ V_dd / (V_dd - V_t)^2
With Vt = 0.3 V, delay at 1.0 V is proportional to 1.0 / 0.49 = 2.04, and at 0.9 V to 0.9 / 0.36 = 2.50. So the 10% voltage cut that saved 19% of switching power makes the logic about 22% slower. Lowering Vt to recover the speed raises leakage, which is the trade-off that dual-Vt design manages path by path.
Power versus energy, in numbers
Energy per clock cycle is P / f. In the block above, 0.100 W at 500 MHz is 0.2 nJ per cycle. Run the same block at 250 MHz and switching power falls to 0.050 W, but energy per cycle is still 0.2 nJ, so a fixed task costs the same switching energy. Only a voltage reduction, or fewer transitions, cuts energy per operation. If the block also leaks 10 mW, that leakage costs 0.02 nJ per cycle at 500 MHz and 0.04 nJ per cycle at 250 MHz.
Which term each technique reduces
| Technique | Term reduced | Main cost |
|---|---|---|
| Lower Vdd (dual VDD, voltage scaling) | V2 in switching, V in short-circuit and leakage | Slower logic, level shifters between domains |
| Dual Vt | Ileak | Slower high-Vt cells, extra library work |
| SOI, smaller gates, fewer gates | Cload | Process cost, area or function trade-offs |
| Clock gating | ActFact of clock and registers | Gating cells, timing and test care |
| Multiple clocks | fclk in slower regions | Clock domain crossings |
| Data encoding | ActFact on buses | Encode and decode logic |
| Glitch reduction | ActFact from spurious transitions | Extra registers or delay balancing |
| Faster edges | tshort in short-circuit power | Stronger drivers add capacitance |
Encoding example
A binary counter flips bit 0 on every increment, bit 1 on every second increment, bit 2 on every fourth, and so on, so it averages close to 2 bit changes per count. A Gray-code counter changes exactly 1 bit per count. On an address bus that mostly steps through sequential addresses, Gray coding roughly halves the switching activity of that bus.
Further reading
- Mark Aagaard, ECE 327 Digital Systems Engineering course notes, University of Waterloo, chapter 9 "Power Analysis and Power-Aware Design" (Internet Archive copy, 2020)
- Texas Instruments, CMOS Power Consumption and Cpd Calculation (SCAA035B)


Your comments will be moderated before it appears here.