Forward Error Correction (FEC)

What Forward Error Correction Does

FEC adds calculated redundancy to the data before it is transmitted, so the receiver can detect and repair errors on its own. Nothing is retransmitted, no round trip is involved and the upper protocol layers never learn that anything happened. At hundreds of gigabits over a link with real propagation delay, asking for a resend is not a viable error strategy, so correcting in place is the only option.

The scheme that dominates modern Ethernet is a Reed-Solomon code written RS(544,514) and universally called KP4. It is the layer that turns a raw optical link making errors continuously into an Ethernet service that appears error-free.

How the Code Works

The encoder groups the data into 10-bit symbols, takes 514 of them and emits 544 — thirty parity symbols, costing about 5.8 per cent of overhead. Reed-Solomon corrects up to half the parity count in symbol errors, so a KP4 codeword survives fifteen corrupted symbols wherever they fall.

Because the unit of correction is a ten-bit symbol rather than a bit, KP4 is naturally good at the burst errors PAM4 produces: a single bad decision on a four-level eye corrupts up to two adjacent bits, and a decision-feedback equaliser can propagate that into a short run. Ten bits go wrong, one symbol is spent, and the codeword still decodes.

The Cliff, and Why It Is Not a Gentle Slope

With random errors, a pre-FEC bit error rate of about 2.4e-4 is the point at which RS(544,514) still delivers an output error rate of roughly 1e-15. Below that, output errors are so rare you cannot measure them in a working day. Above it, the correction capacity is exhausted and the output error rate climbs almost vertically.

That shape is the operational message. There is no warning band in the post-FEC counter — it reads zero, zero, zero, then traffic is broken. All of the early warning lives in the pre-FEC counter, which is why the useful discipline is to trend pre-FEC BER over days rather than glance at it during an incident. The newer clauses reflect that: 802.3 specifies the PMD at a pre-FEC BER of 2.4e-4 and states the end-to-end requirement as a frame loss ratio rather than a post-FEC BER, with the P802.3dj objective set at 1e-13 or better at the MAC interface.

What Changed in Practice

On an NRZ link, any pre-FEC error was a defect worth investigating. On a PAM4 link it is the designed operating point: a healthy 400G lane sits somewhere around 1e-6 to 1e-5 pre-FEC and reports nothing post-FEC. Treating that as a fault is the fastest way to replace working modules.

FEC also moved around inside the system. A conventional module runs the encode and decode in its DSP; an LPO link leaves it to the host, so the host has to support the required mode. And at 224G per lane the drafts add a concatenated inner code for the harder reaches, which is another parameter to match at both ends. The standard lookup lists which FEC each PMD assumes.

Related Terms

Bit error rate is what FEC is measured against, and TDECQ is the transmitter test calibrated to the same error target.