Skip to content
Tutorials Center
My learning

Topics

Lesson 55 of 67 ยท Real-World Machines and Performance

Power, Energy and Thermal Limits

Learn why chips stopped getting faster by clock speed: dynamic power, voltage and frequency scaling, dark silicon and big.LITTLE, with Python calculations.

Every operation a processor performs turns electricity into heat. A laptop cooled by a thin fan, a phone with no fan at all and a data center paying a huge electricity bill all face the same constraint: you can only put so much power into a chip and carry the heat away. This limit, not the number of transistors, now shapes how processors are designed.

In this lesson you will learn the difference between power and energy, the formula for dynamic power, why lowering voltage is so powerful, how this killed the race for higher clock rates, and how designers respond with multicore chips, dark silicon and heterogeneous designs such as big.LITTLE. Everything is backed by small Python calculations. It builds on Measuring Performance.

Power versus energy

These two words are often mixed up, but they answer different questions.

  • Power is the rate of using energy, measured in watts (W). One watt is one joule per second. It determines heat and cooling: a chip drawing 100 W must shed 100 W of heat.
  • Energy is power over time, measured in joules (J) or watt-hours (Wh). It determines the battery life of a phone and the electricity bill of a data center.
TEXT
energy = power x time

A faster chip may use more power but finish sooner, so total energy can go either way. When comparing designs, ask which quantity actually matters: peak power for cooling, energy for batteries and bills, or energy times delay when you want both speed and efficiency.

Where the power goes

A CMOS chip burns power in two main ways:

  • Dynamic power is used when transistors switch. Every switch charges or discharges tiny capacitors.
  • Static (leakage) power is used even when nothing switches, because modern transistors are so small that current leaks through them. It grows with temperature and is a large share of the total in advanced processes.

The dynamic part follows a famous formula:

TEXT
P_dynamic = alpha x C x V^2 x f

where alpha is the activity factor (the fraction of the circuit that switches each cycle), C is the total switched capacitance (roughly, how much circuitry there is), V is the supply voltage and f is the clock frequency.

Two lessons hide in this formula. Power is linear in frequency and in activity, but quadratic in voltage. So voltage is the big lever. And voltage and frequency are linked: a transistor needs a higher voltage to switch fast enough for a higher clock. Over the working range, voltage rises roughly in proportion to frequency, which makes dynamic power grow roughly with the cube of frequency.

Python The dynamic power formula

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

A 20 percent slowdown that also lowers the voltage cuts dynamic power roughly in half, because 0.8 x 0.8 x 0.8 is about 0.51. This trade, giving up a little speed for a large power saving, is the foundation of almost every power-management trick below.

The end of the clock-rate race

For about thirty years, chip makers followed a lucky pattern called Dennard scaling: each time transistors shrank, they could switch faster at a lower voltage, so a chip with twice as many transistors, clocked higher, used about the same power per area. Frequency climbed from megahertz to gigahertz almost for free.

Around 2005 this broke. Voltage could not keep dropping, because leakage grew as transistors got smaller. Without a falling voltage, doubling the transistor count and raising the clock meant much more power per square millimeter. Chips hit what engineers call the power wall: roughly 100 to 150 W per processor package, which is about what air cooling can remove. Since then clock rates have stayed in the 3 to 6 GHz range, and extra transistors have gone into more cores, bigger caches and specialized units.

Dynamic voltage and frequency scaling

Because power falls so fast with voltage, processors adjust both on the fly. This is DVFS (dynamic voltage and frequency scaling). The operating system and hardware pick a point from a table of voltage-frequency pairs depending on load. When you are idle, the chip drops to a low state; when you start a game, it climbs. Under heavy use it may also boost above its normal clock for a short time until temperature or the power budget forces it back down (thermal throttling). Deeper sleep states switch off whole blocks or stop the clock completely when the core has nothing to do.

DVFS creates an interesting question: is it better to run a job fast and then idle (race to idle), or to run it slowly all the time? The answer depends on the leakage power. The next example finds the most energy-efficient speed for a fixed job with a toy voltage-frequency table.

Python Which speed uses the least energy?

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Look at the energy column: it falls at first, then rises. At very low frequencies the job takes so long that leakage power, which is paid for the entire runtime, dominates. At very high frequencies the V-squared term dominates. The best point is somewhere in the middle, and where exactly depends on how leaky the process is. This is why a phone often does short bursts at moderate frequency and then goes quickly back to sleep.

Parallelism as a power strategy

Now apply the cube law to multicore. Suppose one core at frequency f delivers some performance. Two cores, each at f/2 with proportionally lower voltage, deliver about the same total throughput (if the work parallelizes perfectly) but use far less power:

Python One fast core or several slow cores?

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

In the ideal case, four cores at a quarter of the speed use a sixteenth of the power for the same throughput. That is the reason phones and servers have so many cores. But the ideal needs perfectly parallel code. With only 90 percent parallel work, four cores give 3.08 times at full clock, as Amdahl's law predicts, and the serial 10 percent runs on a core that must stay fast. Software determines how much of the hardware's promise you actually see.

Dark silicon

Here is a sobering consequence. Transistors keep getting smaller, so a chip can hold more of them every generation, but the power budget per chip stays about the same. If you switched everything on at full speed at once, the chip would overheat. The portion that must stay idle (or run at reduced speed) at any moment is called dark silicon.

Python How much of the chip can be on?

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

(The numbers are made up to show the trend, not measured from a real chip.) Designers respond in several ways: run more blocks at lower voltage and speed, build specialized accelerators that do one job (video decoding, AI math, encryption) with far less energy than a general core, and turn blocks on only when needed. Dark silicon is a main reason today's chips contain so many different units instead of one giant CPU.

big.LITTLE and heterogeneous cores

Different tasks need different amounts of speed. Scrolling a list needs little; compiling code needs a lot. Heterogeneous processors therefore combine fast, power-hungry big (performance) cores with slower, efficient LITTLE (efficiency) cores. ARM's big.LITTLE and DynamIQ designs, Apple's performance/efficiency cores and Intel's P-core/E-core designs all follow this idea. A scheduler moves threads to the right kind of core: background work goes to the small cores, bursts of heavy work to the big ones.

flowchart LR
  OS["Scheduler"] -->|"light / background"| L["Efficiency cores<br/>low power, lower speed"]
  OS -->|"heavy / latency-critical"| B["Performance cores<br/>high power, high speed"]
  L --- C["Shared cache and memory"]
  B --- C

A scheduler matches each task to a core type.

The final calculation compares energy for different kinds of work on two core types, using illustrative figures: a big core that is three times as fast but uses eight times the power of a little core.

Python Big core or little core?

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

The little core wins on energy for every task (0.50 J per unit against 1.33 J), but takes three times longer. If the user is waiting, use the big core; if the work is background or has a loose deadline, the little core saves battery. Smart scheduling also reduces peak heat, which lets the big cores boost longer when they are needed.

Check your understanding

Quick check ยท 5 questions

of 5 answered You scored / 5 ยท perfect!

  1. 1. What is the relationship between power and energy?

    Power (watts) is the rate of using energy; energy (joules) is power over a period of time.
  2. 2. In P = alpha x C x V^2 x f, which change reduces dynamic power the most for the same relative reduction?

    Power is quadratic in voltage, so a 10% drop in voltage cuts power by about 19%, against 10% for the other terms. If the frequency drops along with it, the saving is larger still.
  3. 3. Why did clock rates stop rising around 2005?

    Dennard scaling ended. Higher frequency without lower voltage raised power density beyond what cooling can handle.
  4. 4. What does dark silicon mean?

    More transistors fit on a chip than the power budget can supply, so part of the chip must stay idle or slow at any moment.
  5. 5. A little core uses less energy per unit of work than a big core but is slower. When should the scheduler prefer it?

    Heterogeneous scheduling puts work that can wait on efficient cores to save energy, and keeps big cores for bursts that need speed.

Summary

  • Power (watts) is a rate and sets cooling needs; energy (joules) is power times time and sets battery life and cost.
  • Dynamic power is alpha x C x V^2 x f. Voltage is quadratic, and since voltage tracks frequency, power grows roughly with the cube of frequency.
  • The end of Dennard scaling created the power wall, so chips gained cores and accelerators instead of clock speed.
  • DVFS, boost, throttling and sleep states adjust voltage and frequency to the workload; there is an energy-optimal speed because of leakage.
  • Dark silicon, accelerators and big.LITTLE-style heterogeneous cores are designs for a world where power, not transistors, is scarce.

Continue with Inside Modern Processors: x86, ARM and RISC-V.