Compliance
What a bit error rate actually counts
The bit error rate is a bound on a probability, and the number of bits a bench must push before a clean run means anything is a consequence of it.
What a bit error rate counts
A bit error rate is a ratio, not a verdict: the number of bits received wrong divided by the number of bits sent. When a PCI Express document names a figure on the order of one error in 10 to the 12th, it is quoting a probability the link is designed to stay under, and reading that figure well is a small exercise in applied probability.
The count is honest but easily misread. A BER of 1e-12 does not promise an error every trillionth bit on a schedule; it states that errors are rare enough that the long-run rate stays below that bound. At 8 GT/s per lane, a bound of one error in 10 to the 12th bits corresponds to an expectation of an error every couple of minutes per lane on a fully loaded link, not to a guarantee that any given minute contains one. The figure is a ceiling on a rate, and the experiments that establish it run for a finite time, which means what is measured is an upper bound consistent with the observation, not a certainty about forever.
The habit of treating a counted risk as a readable object has a long pedigree. The history of how counting outcomes became a mathematics is kept by the Cabinet of Chance, which traces the path from dice and knucklebones to the pioneers of probability. A BER figure is a late descendant of the same question: how many ways can the die fall, and how often does the bad face come up.
Where does the number come from?
From test equipment that counts mismatches. A bit error rate tester sends a known pattern, typically a pseudo-random sequence, and compares what returns against what was sent. The ratio accumulates over the run, and the run length sets the confidence: to see whether a link meets a 1e-12 bound, the tester must move well over 10 to the 12th bits, which at serial rates is minutes rather than seconds. A shorter run that sees zero errors establishes less than it appears to.
The layers underneath the bit stream matter too. Errors detected by the link's own protection, the CRC at the data link layer, are corrected by replay and never reach the bit-count. The BER figure quoted at the electrical level is the raw channel, before the retry machinery hides what it fixes. The distinction between what the channel does and what the protocol forgives is the same distinction drawn on the page about encoding and usable bandwidth: the wire rate and the usable rate are different numbers, and so are the raw error rate and the observed one.
How does the figure relate to the eye?
The eye diagram and the BER are two views of the same margin. The eye, described on the page about reading a receiver eye, shows how much room the signal leaves around the decision point in voltage and in time. The BER is what happens when the room runs out: every sample that lands on the wrong side of the decision threshold is a counted error. A smaller eye is not itself a BER figure, but the statistical shape of the jitter and noise at the decision point is what the BER counts.
| Question | What the figure says | What it does not say |
|---|---|---|
| How often do bits flip | At most the quoted rate, as a long-run bound | When the next error will happen |
| Was the channel tested long enough | Only if the run moved far more than the inverse of the bound | A zero-error short run proves less than it looks |
| Is the user-visible error rate the same | No: link-layer replay removes detected errors | Nothing about corrected errors |
Why does the standard quote a bound rather than a measurement?
Because a measurement belongs to a setup and a bound belongs to a design. The standard cannot know which board, which cable, which temperature the link will meet, so it specifies a target the channel must be engineered to beat and leaves the demonstration to the compliance program described in how a compliance program works. The bound is the contract; the measurement is the evidence. Reading the contract as if it were the evidence is the most common misreading of the whole field.
The same figure also travels badly between contexts. A BER quoted for a raw electrical channel, a channel through a retimer, and a link with forward error correction are three different objects with the same notation. The correction machinery described in the literature on forward error correction, read on September 6, 2026, can change the observed rate by orders of magnitude, and a figure without its layer is a figure without a meaning.
Checks for reading a BER figure
- Ask which layer the figure measures: raw channel, link after replay, or end-to-end after correction.
- Check the run length: a bound of 1e-12 is only tested by moving well over a trillion bits.
- Convert the bound to time at the actual rate before calling it rare: at 8 GT/s a 1e-12 bound is not rare.
- Ask what the link does with a detected error before treating the raw rate as the visible one.
Common mistakes
- Reading 1e-12 as an error every trillionth bit on a schedule rather than a bound on a rate.
- Accepting a zero-error short run as proof of a long bound.
- Comparing BER figures measured at different layers of the stack as if they were the same object.
- Quoting the bound without the rate that turns it into time.
A bit error rate is a probability with a unit attached, and the discipline of reading it is the discipline of all counted risk: name the denominator, name the run length, and name the layer before naming the number.
The page behind these facts
The definition of the rate, its measurement by known patterns and the relationship between confidence and run length are summarized in the Wikipedia article on bit error rate, read on September 6, 2026. The replay machinery that separates the raw channel from the visible error rate is described in the PCI Express article and the cyclic redundancy check article, read the same day.