Heat and throttling

Short peak performance differs from sustained performance. Stress mode runs CPU multicore, GPU or both for five or ten minutes, recording a score every fifteen seconds. Stability is the minimum recorded score divided by the maximum. Throttling onset is inferred at the first of two consecutive rounds at least ten percent below the peak seen so far. This is not a hardware temperature sensor or a direct thermal-limit signal. Background work and power limits can cause similar drops. Pauses discard the current interval and may let the device cool; consecutive-drop detection restarts across pause boundaries. Keep room conditions, case, stand and charging state comparable. Stop if the device becomes uncomfortable to use.

Before measuring, write down the conditions you want to compare: operating system and browser versions, power connection, energy mode, external display and device model. Wait for downloads, video playback, backups and updates in other apps to finish. Keep the screen on and this tab visible throughout the run. Hiding the tab discards the current round and repeats it when you return, so frequent interruptions create different conditions from a continuous run.

Repeat the measurement two or three times under the same conditions and look at the range rather than selecting only the highest score. Allow the device to return to a comparable state between runs. If a small difference changes direction on repetition, treat it as variation rather than proof of superiority. Check valid test counts and exclusion reasons too. Missing support does not necessarily indicate defective hardware, but it changes what the overall score summarizes.

Change one condition at a time. Changing the browser, power mode and cooling together makes it hard to identify the cause of a difference. Use quick mode to check that the test runs, then use full mode for detailed comparisons. When sharing, include score version, mode, GPU backend and whether AI was included. Read the raw units: higher throughput is usually better, while lower latency is better. Normalized scores account for that direction. Rounded display digits alone cannot establish an improvement. These measurements describe web workloads; they do not guarantee game frame rates, battery life or performance in every desktop application.