Thermal throttling

Why does a laptop slow itself down at exactly the moment you are asking most of it?

Start a long video export and the machine is fast — for perhaps forty seconds. Then the fans reach their limit, the work rate visibly settles to something lower, and stays there. The obvious reading is that something is failing: the cooling is inadequate, the design is cheap, the chip is being protected from a flaw.

But hold that reading up to the light. If the chip could not sustain that speed, why was it ever allowed to run at it? Nothing forced the first forty seconds. Someone chose them.

Begin with where the heat comes from, because the shape of the answer is hidden in it. A digital circuit dissipates power mostly when it switches, and that dynamic power goes roughly as the capacitance switched, times the supply voltage squared, times the clock frequency. The voltage term is the one to watch: to run faster, transistors must switch harder, which needs a higher supply voltage, so raising the clock raises frequency and voltage together and power climbs faster than the clock does. Turn that around and you have the lever the whole mechanism uses — a modest cut in clock speed buys a disproportionately large cut in heat.

Where does that heat go? Out through the die, the package, a spreader, a heatsink, and finally into moving air. Every step resists heat flow, and in the steady state the temperature rise above ambient is the power dissipated multiplied by the total thermal resistance of that chain. Fix the cooling hardware and you have fixed the resistance. There is a temperature the silicon must not exceed — and in a phone or thin laptop a second, often stricter limit, which is how hot the outside may get before it is unpleasant to touch. So there is a power level the machine can shed indefinitely, and it is set by the cooling, not by the chip.

But now ask about the first forty seconds. Heat does not reach the fins instantly; it has to warm everything on the way. All that copper and aluminium has thermal mass, and while it is warming it absorbs energy without passing it on. For a short burst the chip may draw far more power than the cooling could ever remove, and the temperature climbs slowly because the surge is banked in the metal rather than crossing it. The burst limit is set by how much heat the assembly can store; the sustained limit by how fast it can dump.

Here is the honest design point, and it inverts the complaint. Almost everything a person does with a computer is bursty — opening an application, loading a page, applying a filter. Those take seconds, and in seconds the thermal mass is nowhere near full, so they run at the high number. A designer who set the clock to the sustainable value and left it there would build a machine slower at nearly everything, in exchange for consistency during the workloads you notice least. So the chip is deliberately specified to exceed its own sustainable clock in bursts and to fall back when a burst runs long. The fallback is not the failure. The burst is the feature, and the fallback is its price.

What closes the loop is measurement: sensors report die temperature continuously and a controller adjusts voltage and frequency to keep the reading just under the limits. It responds to ambient conditions too, so the same laptop throttles harder on a hot day, on a duvet, or with dust clogging the fins. Those are faults — but their signature is throttling that arrives sooner and settles lower than intended, which is a different observation from throttling at all.

Three refinements tie the steps together.

First, throttling is a response, not a cause. The chip is not slowing because it is hot in the way a tired thing slows; it is commanded to slow so that it will not become hot. Remove the controller and the same silicon would run until it damaged itself or shut down abruptly, which is what protection circuits used to do.

Second, "the cooling is inadequate" and "the design intends this" are not competing claims; they describe different quantities. The cooling sets the sustained level, the design sets the height and length of the burst above it, and a fanless tablet and a desktop tower may share a peak clock while differing enormously in what they hold.

Third — and this is where the popular framing goes astray — more cooling does not straightforwardly buy more speed. It raises the sustainable power, but the clock that power supports rises much more slowly than the power itself, because of the same voltage-squared relationship that made throttling effective in the first place. The steepness that lets a small slowdown save a lot of heat also means a lot of extra heat removal buys only a little extra speed. The curve cuts both ways, and it is the same curve.

• How vapour chambers and phase-change materials extend the burst by adding storage rather than by adding cooling.
• Why big-and-little core designs attack the same problem from the other end, by making the sustainable point itself more efficient.

Every term the collection defines is gathered in the glossary.