A browser sees how much work gets done each second, not why. The processor records why it lowered its clock, such as a power, temperature or current limit, and only the operating system can read that.
- Shapes still narrow it down. A level stretch and then one step down is consistent with a power budget running out; a fall spread across the run is consistent with heat building. Each has look-alikes, and neither is proof.
- Another program sharing the cores is the one cause a run measures directly: each worker times its own batches of work, and a run that loses 3% or more of its working time to pauses is flagged.
- To tell heat from a power limit for certain, log temperatures, clocks and limit reasons with a sensor tool during the same run.
The short answer
Read the shape first, then the run quality notes, then confirm with a sensor log. The shape tells you which causes fit; the run quality notes tell you whether another program was in the way; a sensor log is what names the limit.
The shapes the report names
The report places a processor curve in one of a fixed set of shapes, each with a confidence. These are the shapes that describe a change; the full set, with how confidence is worded, is on the methodology page. How a curve is assigned to a shape is not published.
- Held: The closing level stays close to the opening plateau, inside the run's own noise.
- Plateau, then a step down: A level stretch, then a fall concentrated in one transition, then a lower level.
- Gradual slide: A fall spread across the run rather than taken in one step.
- Oscillating: The level rises and falls repeatedly rather than settling.
- Dipped, then recovered: A fall partway through, then a return to near the opening level.
- Erratic: Movement with no consistent direction or rhythm.
A power limit
Many processors run above their sustained power for a short budget and then settle. On a curve that looks like a level stretch and one step down to a lower level. SystemCheck shows a short opening boost and its step on their own and does not count them against the machine; a step later in the run is read as part of the curve.
A step has look-alikes. A window that lost focus can have its work moved to efficiency cores, a laptop can switch from mains to battery, and a background scan can start. The report lists these beside the shape. A sensor log that shows the power limit reached at the moment of the step is what settles it.
Heat
Heat builds over minutes rather than seconds, so a fall that is spread across the stage rather than taken in one step is the shape heat tends to draw. The Deep run holds the load longer and re-tests the machine hot, which is the better test of cooling than a one-minute stage.
Starting state matters more than most people expect. A machine that has just finished a game may show almost no fall because it had already given up its headroom; a machine that has sat idle for an hour has more to lose. Start every run you compare from the same state.
A busy machine
This is the cause a run can measure. Each worker times every batch of its work; a batch does the same work every time, so on a quiet core its time barely moves, and heat or a power limit slows every batch alike. When the operating system hands the core to another program, that one batch takes several times as long. The share of working time lost that way is reported, and a run at 3% or more is flagged as sharing its cores.
On our validation laptop, in 20-second processor stages: three quiet runs measured 70.7 to 73.6 million operations a second and lost 0.2 to 0.5% of their time to pauses; beside a four-thread load the rate fell to 52.2 and the paused share rose to 14.7%; beside a twelve-thread load, 31.5 and 49.4%. An earlier signal, the share of time between batches, moved by only 0.01 beside a load that cost 28% of the rate, which is why it was replaced.
The limits are stated in the methodology: pauses shorter than the browser's timer resolution (1 ms in Firefox and Safari) are not seen, and a load on the graphics chip alone is not seen by the processor stage.
What settles it
A sensor tool that logs temperatures and clocks and, where the processor reports them, its limit reasons, run during the same test. If the clock falls as the temperature reaches its limit, heat is the cause; if it falls as package power reaches its limit while the temperature has room, the power budget is. Run the same test twice under the same conditions before reading either way.
Terms used here
Questions
- Is thermal throttling the same as power limit throttling?
- No. Thermal throttling lowers the clock to stay under a temperature limit. A power limit lowers it to stay under a power budget, often after a short boost allowance runs out, whatever the temperature. Both make a curve fall, which is why a curve alone cannot tell them apart for certain.
- Can a browser test tell me which one my laptop has?
- It can say which shapes the curve fits and whether another program was sharing the cores. It cannot read temperatures, power or the processor's limit reasons, so it names what a shape is consistent with and leaves the cause to a sensor log.