Thermal decay, explained

Decay is the percentage gap between a system's early peak throughput and the plateau it settles into under sustained load.

• It is a behavioural measurement, not a temperature reading. SystemCheck derives it from throughput over time because browsers expose no thermal sensor API.
• The shape of the decay curve identifies its cause: a gradual slope over 20-60 seconds points to heat, an immediate hard plateau points to a power limit, and a jagged trace points to background contention.
• Lower decay is not automatically better. A machine that starts slow decays less, and sustained throughput is the figure that determines how long a job takes.
• Decay is a diagnostic that tells you where to look, then hand off to HWiNFO or your vendor's utility for the actual temperatures and power figures.

The definition

Decay is the distance between two points on a throughput curve: the early peak, reached while the machine is cold and boost budgets are full, and the sustained plateau it settles into once thermal and power equilibrium arrive. Expressed as a percentage of the peak, it summarises how much of a machine's headline capability survives contact with a real workload.

SystemCheck computes it from a retained throughput series rather than from two sampled instants. Stage 1 runs one Web Worker per logical processor under a continuous floating-point load and keeps throughput across the whole run; Stage 2 does the equivalent for GPU raster work and frame time; Stage 3 for path-tracing samples. Decay falls out of those series directly.

Why it is a proxy and not a reading

This needs saying without hedging, because the distinction is the whole basis on which the number should be trusted. Browsers expose no API for temperature, power draw, fan speed, or clock frequency. There is no permission prompt, no flag, no vendor extension that unlocks it. SystemCheck does not read your thermal sensors, and it will not display a temperature it did not measure.

What it measures is work completed per unit time, continuously, under a load heavy enough to saturate the hardware. When that rate falls and then stabilises lower, the machine has hit a limit. Decay quantifies that. It is a behavioural consequence of thermal or power limiting, which makes it a genuine signal about a real effect, and it stops precisely at the point where naming the physical cause would require sensor data the page does not have.

That boundary is also the practical handoff. Decay tells you the machine is limiting and roughly how hard. HWiNFO64, HWMonitor, or your vendor's utility (Ryzen Master, Intel XTU, MSI Afterburner, AMD Adrenalin) tell you the die reached 98°C, or that package power capped at 35W, which is what you need to know which fix to apply.

Reading the shape, not just the number

Three causes produce three recognisable curves, and distinguishing them is most of the diagnostic value.

• Thermal: a gradual slope over roughly 20 to 60 seconds as heat accumulates, settling into a stable lower plateau, with full recovery after a minute or two of idle.
• Power limit: a fast drop within the first few seconds to a hard, flat plateau. Onset is too quick to be heat accumulating. Extremely common on laptops, and the signature case is a machine that behaves completely differently on battery.
• Contention: an irregular, jagged trace rather than a curve, usually accompanied by a poor consistency figure. Something else is taking cores in bursts. This is not a hardware property and should not be read as one.

The trap: low decay is not the goal

It is easy to start optimising for the decay percentage, and it is the wrong target. Decay is a ratio, and a ratio can be improved by making the numerator worse. A machine capped to a low clock from the first second will show near-zero decay while doing less work than a machine that boosts hard and settles 20% down, and the second machine finishes the job first.

The figure that determines how long your export takes is sustained throughput, in absolute terms. Decay tells you how much was left on the table relative to what the chip could briefly do. Both are worth having; only one is the goal. When you evaluate a change, compare sustained throughput first and treat the decay number as an explanation of the result rather than as the result.

What a healthy curve looks like

On a well-cooled desktop under a saturating all-core load, expect a short peak, a settling period of twenty to forty seconds, and a plateau within a few percent of the peak, held flat for the rest of the run. Consistency should be high throughout. That is a machine whose cooling is comfortably ahead of its power delivery.

On a laptop, expect a sharper drop and a plateau 10-25% below peak, and treat that as the design working rather than failing. Thin machines are engineered around a sustained budget well below their burst budget. What should concern you on either is a plateau that keeps sagging rather than stabilising, a second run that starts materially lower than the first, or a jagged trace with a poor consistency figure. The first two indicate the cooling solution cannot shed the energy at all; the third indicates you were not measuring the hardware in the first place.

Questions

Does SystemCheck measure my temperature to calculate decay?

No. Decay is computed entirely from throughput over time. Browsers expose no thermal sensor API, so no web page can read temperature. Decay is a behavioural proxy for thermal and power limiting, and SystemCheck reports it as exactly that.

What decay percentage should I aim for?

Aim for high sustained throughput rather than for a low decay percentage. Under 5% is typical on a well-cooled desktop and 5-15% is normal on laptops, but a low ratio achieved by starting slower is not an improvement. Compare absolute sustained throughput before and after any change.

My decay is high only on battery. Why?

That is a power limit rather than a thermal one. Many laptops enforce a substantially lower sustained power budget on battery, and the signature is a fast drop to a hard plateau within seconds rather than a gradual slope. Cooling changes will not affect it; plugging in will.