Performance Check

doc

Measure browser timing, responsiveness, and performance signals to diagnose instability, fingerprint risk, and device profile drift with actionable workflows.

Two laptops of the same model, running the same browser, can be told apart by how fast they execute the same code. Performance fingerprinting measures that difference. It is a fingerprint made of timing, and it is harder to spoof than a header because it depends on the machine actually doing the work.

What a performance check measures

MeasurementWhat it reflectsStability
Core count and memory tierHardware class, coarsely rounded by the browserFixed per device
JavaScript execution speedCPU generation, thermal state, background loadModerate; noisy between runs
Timer resolution and driftOS scheduling and browser throttlingStable within a build
Floating-point behaviourHardware maths implementationFixed per CPU generation
WebGL or WASM throughputGPU and driver, or JIT qualityHigh per device
Network round-trip estimatePath latency to the serverLow; changes constantly

The first and last rows are the ones usually reported, and they are the least distinctive. Core count is shared by millions of machines; latency changes with every request. The rows in the middle are where the discriminative power lives.

Why timing is hard to fake

A user agent is a string: rewriting it is trivial. Performance is a measurement: to change the result you have to change how the machine runs, which means changing the machine.

  • Software throttling is visible. Slowing a benchmark to imitate a weaker device also slows everything else, and the throttled values form their own recognizable cluster.
  • Averaging defeats noise injection. Repeated measurements converge. Random jitter added to each sample disappears in the mean.
  • Cross-checks catch inconsistencies. A machine that reports fast CPU figures and slow GPU figures is unusual in a specific way, and that combination is itself a signal.

This is why performance checks appear in bot detection and in device-integrity systems: they are expensive to imitate convincingly at scale.

The two uses, and the tension

  1. Capacity planning. Deciding whether a site can ship a feature — a heavy chart, a video pipeline — based on what the visitor's machine can actually do.
  2. Classification. Distinguishing real devices from emulated ones, and repeat visitors from new ones, without a cookie.

The same numbers serve both, and the check page cannot tell which use is intended. What it can do is report the measurement with its noise, so the number is not mistaken for a verdict.

Reading a result

  • Single-run numbers are unreliable. A benchmark on a machine with a browser tab compiling in the background will read low. Compare two runs before drawing a conclusion.
  • Thermal state matters. A laptop on battery after an hour of work is throttled; the same laptop plugged in and cool is not. This is a real variable, not a measurement error.
  • Reported memory is rounded. The browser buckets it deliberately, so close values are not distinguishable.
  • Latency is not a performance property of the device. It describes the path. Do not read it as hardware speed.

What cannot be hidden

You can reduce background load and plug in the machine, which improves the numbers honestly. You cannot make a fast machine report as slow, or the reverse, without changing what it is. Performance is one of the few fingerprint surfaces where the measurement is the property.

Common misreads

  • "My score is low, so my device is failing." A low score can mean a throttled laptop, a busy browser, or a modest machine. Three causes, one number.
  • "The numbers are random." They vary, but they cluster. Repeated sampling recovers the stable part.
  • "Latency is part of the fingerprint." It is a property of the network path, which changes constantly and is not device-specific.
  • "Changing my resolution changes my score." It can, by changing how much work the browser does per frame — an indirect effect, not a disguise.

FAQ

Is performance fingerprinting used for tracking?

Yes, in device-integrity and bot-detection systems. It is also used for capacity planning, which is legitimate. The measurement is identical; the purpose differs.

Can a VPN hide my performance profile?

No. The work runs locally. A tunnel changes the path latency and nothing about execution speed.

Why do two runs give different numbers?

Background load, thermal throttling and scheduler noise. Comparing medians over several runs reduces the effect but does not eliminate it.

Are the core and memory values exact?

No. Both are bucketed by the browser to limit their use as identifiers.