Design for low power depends primarily on the characteristics of the design being implemented. If it is a multi-million gate design, we cannot implement any technique that is gate specific; it has to be a global technique. This post lists the global techniques that work at that scale, explains each in plain terms, and then puts numbers on them.
Where the power goes
A CMOS chip burns power in two ways. Dynamic power is spent every time a node switches, charging and discharging capacitance; it depends on how often things toggle, how much capacitance there is, the clock frequency and, most strongly, the supply voltage. Leakage power is spent all the time, even when nothing switches, because transistors never turn fully off. Global techniques attack one or both.
The global techniques
- Multi-Vdd, variable Vdd and multi-Vth seem to be a good global solution. Run blocks that do not need full speed from a lower supply (multi-Vdd), adjust the supply at run time (variable Vdd), and use high-threshold, low-leakage cells on paths that have timing slack while keeping fast low-threshold cells only on critical paths (multi-Vth).
- Reducing the clock speed will result in low power consumption, but at the cost of performance. Dynamic power falls in proportion to frequency. On its own, though, it does not reduce the energy needed to finish a task, because the task simply takes longer.
- Using power header and power footer transistors cuts down power. These are switches placed in the supply (header) or ground (footer) path of a block of logic, not of individual gates. Turning them off disconnects the block and removes almost all its leakage. This is called power gating.
- You could separate the design into blocks which can go into sleep mode. Partition the chip into power domains so that idle blocks can be gated off or held at a low retention voltage while the rest keeps running.
- Another solution is variable VDD and variable frequency (as Intel or AMD do). This means you adapt VDD and frequency to the necessary performance. This is dynamic voltage and frequency scaling (DVFS), and it is the single most effective technique because voltage enters the power equation squared.
- Gated clocks and logic addressable clocks. Stop the clock to registers whose contents are not changing. The disadvantages noted at the time were timing problems due to improper latching of signals, and difficulty in testing. Modern flows solve both with integrated clock-gating cells (a latch plus an AND gate, so the gated clock cannot glitch) and a test-enable input that forces the clock on during scan.
- Negative edge triggered flops. The original list suggested that a negative-edge flop built as (NOR + INV) is about 1.5 gates against 2.5 gates (AND + INV) for a positive-edge flop, so negative-edge flops would have fewer gates, less glitching and hence lower power. Treat this as a gate-level rule of thumb rather than a general fact: in a standard-cell library, compare the area and power figures of the actual positive-edge and negative-edge flip-flop cells before choosing, and remember that mixing clock edges halves the time available on paths between them.
Any more thoughts and ideas are welcome.
How these fit together
Most large chips combine them. Clock gating is applied almost everywhere because it is cheap. Multi-Vth cell swapping happens automatically during synthesis and optimization. Power domains, gating and DVFS are architectural decisions made early, because they need extra hardware: switches, isolation cells, level shifters, retention registers and a power controller. The power intent is written in a standard format, the Unified Power Format (IEEE 1801, UPF), so that synthesis, place and route and verification tools all see the same domains and rules.
Deep dive: the numbers and details
This part holds the power equations and worked examples. Skip it if the overview is all you need. All numbers are illustrative values chosen to show the method.
The power equation
P_total = alpha x C x V^2 x f + I_leak x V
alpha is the switching activity, C the switched capacitance, V the supply and f the clock frequency. Short-circuit current during switching adds a smaller third term.
Frequency scaling against DVFS
Start at 1.0 V and 1.0 GHz. Two ways to cut power:
| Operating point | Dynamic power | Energy per operation |
|---|---|---|
| 1.0 V, 1.0 GHz (baseline) | 1.00 | 1.00 |
| 1.0 V, 0.7 GHz (frequency only) | 0.70 | 1.00 |
| 0.8 V, 0.7 GHz (DVFS) | 0.8^2 x 0.7 = 0.45 | 0.8^2 = 0.64 |
Lowering frequency alone saves power but not energy. Lowering voltage with it saves both, which is why DVFS matters for battery life.
Multi-Vth and leakage
Subthreshold leakage falls by one decade for each increase in threshold voltage equal to the subthreshold swing S. At room temperature S is at least about 60 mV per decade and is typically higher in real devices. With S = 85 mV per decade, raising Vth by 100 mV cuts leakage by 10^(100/85), about 15 times, at the cost of slower cells. Swapping cells on non-critical paths to the high-Vth version therefore saves a large share of leakage for little timing cost.
Clock gating
Take a 1,000-flop register bank whose contents change in only 10 percent of cycles. Without gating, every flop's clock pin toggles every cycle. With one clock-gating cell enabled only when the bank is written, the clock pins toggle in 10 percent of cycles, so their clock power drops by about 90 percent, less the small power of the gating cell itself.
Power gating: header sizing and break-even time
- Switch sizing: the header must pass the block's peak current with a small voltage drop. For 100 mA and a 30 mV budget, the total on-resistance must be at most 0.03 / 0.1 = 0.3 ohm, usually built from many switch cells in parallel.
- Break-even time: switching a block off and on costs energy (recharging its supply network and restoring state). If the block leaks 5 mW and a sleep and wake cycle costs 2 microjoules, the block must stay off for at least 2 x 10^-6 / 5 x 10^-3 = 400 microseconds to save anything.
- Supporting cells: isolation cells clamp the outputs of a powered-off block to known values; retention flops keep key state on an always-on supply; level shifters are needed wherever a signal crosses from a low-voltage domain to a higher one.
Common mistakes
- Gating clocks with plain AND gates instead of integrated clock-gating cells.
- Forgetting isolation on the outputs of a switchable domain, so a powered-off block drives floating values into live logic.
- Measuring power with unrealistic activity; dynamic power depends on alpha, so use activity from real workloads.
- Choosing power gating for blocks whose idle periods are shorter than the break-even time.
Further reading
- Low Power Design Techniques on the ASIC-System on Chip-VLSI Design blog.
- Unified Power Format (IEEE 1801) overview.


Your comments will be moderated before it appears here.