Notes on measuring honestly.

Longer arguments about measuring hardware

The SystemCheck blog has 8 long-form posts on how PC performance gets measured and where the usual conventions mislead: Soft shadows as a GPU load: the A/B method; What a measurement revision is; Your boost is not your speed: what the first 10 seconds hide; Ten runs on one laptop: how repeatable is a browser test?; Why the WebAssembly kernel scores 5.4x the JavaScript one; Why peak benchmarks mislead; What browser benchmarks can and cannot measure; Thermal decay, explained. Each one explains a decision behind the free browser stress test, with the reasoning and the limits stated.

The posts are the reasoning; the rest of the site is where it is applied. The methodology page is the specification the posts argue for, and the changelog records every change to it. The guides turn the same ideas into practical steps, the glossary defines the terms, and the fix articles and diagnosis pages start from a symptom. Read a post when you want to know why the test works the way it does, and run the test when you want the answer for your own machine.

Benchmarking has a culture problem: the number that gets quoted is almost always a peak, held for a few seconds in a cool room, and the number that decides whether a machine is pleasant to use is the one it still holds several minutes later. These posts are the longer arguments behind that position, and behind the decisions in how SystemCheck measures.

They also mark the boundary honestly. A benchmark that runs inside a browser cannot read a temperature, a fan speed or a wattage, and a good deal of what is written about PC performance quietly assumes access to all three. Where we infer a cause from the shape of a curve rather than reading a sensor, these posts explain how that inference is made and how far it can be trusted.

If you want the method rather than the argument, the methodology page sets out each stage, and the guides section is the practical counterpart: what to do about throttling, how to read a result without fooling yourself, and how to tell a processor limit from a graphics one.

Good places to start

• Soft shadows as a GPU load: the A/B method
• What a measurement revision is
• Your boost is not your speed: what the first 10 seconds hide
• Ten runs on one laptop: how repeatable is a browser test?
• Why the WebAssembly kernel scores 5.4x the JavaScript one
• Why peak benchmarks mislead