How to Read a Keyboard Latency Chart Without Fooling Yourself

There are two ways to read a latency chart. One is to find the biggest number that flatters the product you already like and quote it. The other is to work out what the chart is actually plotting, where the clock starts and stops, and how many samples produced it. The first method takes seconds and produces confident nonsense. The second takes a few minutes and usually changes your mind.

Charts are persuasive precisely because numbers look neutral. A figure printed to two decimal places suggests a rig that resolves hundredths of a millisecond, which is rarely true. Learning to read the method behind the number is the difference between a chart that informs you and one that sells you something.

What question is the chart actually answering?

The first thing to establish is the interval being measured. A chart that starts at the switch and ends at the host answers a different question from one that starts at the USB report and ends at the pixel. Both are legitimate, and they can differ by many milliseconds on the same board.

If the chart does not say where the clock starts, treat it as decoration. That single omission is the most common reason two reputable sources publish incompatible figures for the same hardware, and it is never an accident on the part of whoever commissioned the chart.

Is it an average, a median, or a worst case?

This matters more than the headline number. An average hides the tail, and the tail is what you feel when a press occasionally lands late. A median is more robust but hides the same thing. A worst case tells you the bound, which is usually what a timing-sensitive user cares about.

The most informative charts plot a distribution rather than a single value. From a distribution you can read the centre and the spread at once, and you can see whether a board is consistently mediocre or usually excellent with occasional lapses. Those two feel very different in the hand.

How many samples produced the figure?

A number from five presses is an anecdote. Timing measurements are noisy, and the noise does not average out until you have collected a meaningful run. We treat anything under a few dozen samples as indicative only, and we say so when we publish it.

Sample count also tells you whether the chart can support a claim about outliers. Detecting a problem that occurs one press in fifty requires far more than fifty presses, which is why rare glitches almost never appear in published figures even when they exist.

What was the rest of the system doing?

A measurement taken on an idle machine with no background work is a best case, and most real machines are not idle. Background load pushes the host-side portion of the chain around, and the effect lands inside the interval the chart is measuring. An unstated host condition makes the figure hard to reproduce.

The same applies to the display. If the chart’s endpoint is a pixel change, the refresh rate and the frame pacing of the test machine matter enormously. A figure taken at one refresh rate does not transfer to another, and a chart that omits it is incomplete.

Why do two charts of the same board disagree?

Usually because of the interval definition, and sometimes because of sample size or host condition. Rarer causes include different firmware versions, different units from different batches, and different connections. None of these makes either chart dishonest, but only one of them answers your question.

Checking these four things before comparing charts eliminates most apparent contradictions. When the methods match, independent measurements usually agree within a surprisingly small margin, which is a good sign that the underlying phenomenon is real.

What does a suspiciously round number suggest?

That it was rounded, or estimated, or derived rather than measured. Measurements produce messy values, and a table full of numbers ending in zero usually means someone tidied them. Tidying is not dishonesty, but it does tell you the precision is lower than the presentation implies.

Excessive decimal places are the opposite signal and equally unreliable. Two decimal places on a millisecond figure implies resolution that consumer capture equipment does not have, and quoting it suggests the author has not thought about the limits of the rig.

How should you compare two boards from different charts?

You usually cannot, and it is better to admit that than to subtract two numbers you cannot reconcile. If both sources publish their method and the methods match, a comparison is defensible. If either source is silent about the interval, the honest answer is that the comparison is unknown.

The practical alternative is to look for a source that measured both boards on the same rig. Single-source comparisons are far more reliable than cross-source arithmetic, even when the single source has fewer decimal places.

What should a trustworthy chart always state?

Where the clock starts and stops, how many samples were taken, what the host and display conditions were, and whether the figure is a mean, a median or a bound. Those four items take one paragraph to write and turn a number into evidence.

It should also state what was not controlled. Every rig has limits, and a source that admits them is generally more reliable than one that implies the measurement is exact. That kind of honesty is easy to spot once you start looking for it.

Does any of this change what you should buy?

Often, because the ranking by worst case differs from the ranking by average on real hardware. A board that wins on the mean can lose badly on consistency, and for timing-sensitive work consistency is the property that matters. Reading distributions instead of headline figures changes decisions.

It also changes how much a given improvement is worth. Once you know which slice of the chain a chart covers, you can tell whether a faster board, a faster display, or a cleaner host is the change that will actually help you.

Where should you go after reading the chart?

What role does the display play in a keyboard chart?

A large one, whenever the endpoint is a pixel. The panel’s refresh interval quantises the result, and processing modes add their own delay on top. Two charts of the same keyboard taken on different displays will disagree by more than the keyboards differ.

This is why a chart should state the display and its mode. Without that, a figure describing the keyboard is actually describing the keyboard plus an unknown display, and the unknown part is often the larger of the two.

Why do some charts show a single press?

Because capturing one press is easy and capturing a distribution is work. A single press has no spread, so it cannot show consistency, and consistency is what separates good boards from merely fast ones. A chart built on one sample is an illustration rather than a measurement.

Look for the sample count before trusting any claim about outliers. Detecting a problem that happens rarely requires many samples, and a chart that does not report how many it took has not earned the right to make that claim.

How should you handle conflicting claims between two sources?

Compare methods before comparing numbers. If the intervals, sample sizes and conditions match, and the figures still disagree, then one of them has a problem with its rig. If the methods do not match, there is nothing to reconcile, and both can be right.

Prefer the source that publishes more about its method. Transparency correlates strongly with care, and a source that describes its limits is generally measuring something real rather than assembling a figure for effect.

What should you do when no method is stated?

Treat the figure as a rough ordering and nothing more. It tells you that the device was measured and that someone thought the result worth publishing, and it does not tell you what was measured. That is still slightly better than a marketing claim with no measurement behind it.

Where possible, find a second source that does state its method and check whether the two agree. Agreement across methods is weak evidence, and disagreement is a signal that at least one of them is measuring something narrower than you assumed.

Does the number of devices in a chart matter?

Yes, because a chart with one device cannot be a comparison. Comparing against a reference device measured on the same rig is what makes a figure interpretable, and a chart without one gives you a number with no scale attached to it.

Look for a familiar device in the set if you can find one. Anchoring an unfamiliar figure against something you know is the fastest way to judge whether it is good.

Reproduce it, at least roughly, on your own hardware. Even a crude home measurement tells you whether your system behaves like the test rig, and the cases where it does not are the interesting ones. A chart that disagrees with your own desk is a prompt to measure more, not to argue.