Every link is a negotiation between how little power you can transmit and how much error you can repair. Forward error correction is where that negotiation is settled: it buys back the decibels the link budget cannot afford, at a cost in latency, silicon area and power that almost always dominates the receiver.
A receiver needs a certain signal-to-noise ratio to recover data. You can supply it with transmit power, with antenna gain, or with coding — and coding is usually the cheapest of the three, because power costs amplifier efficiency and thermal design while antenna gain costs mechanics and beamwidth.
The other reason is that retransmission is not always available. On a satellite link the round trip makes an automatic repeat request expensive, and on a broadcast link there is no return path at all. Where the channel is one-way or the delay is long, correction has to happen forward.
The consequence is that the decoder frequently dominates the receiver. On the FPGA satellite modem programme the error-correction block consumed a disproportionate share of both logic and memory, and its architecture effectively chose the device for the entire design.
They are not interchangeable. Block length, target error rate, latency budget and decoder cost decide, and modern standards use more than one.
| Family | Where it is used | Why it was chosen there |
|---|---|---|
| Convolutional / Viterbi | Legacy cellular, deep-space heritage, short control channels | Simple decoder with good performance at short lengths; superseded for bulk data but still efficient where blocks are tiny |
| Turbo | 3G and 4G data channels | Near-capacity performance at moderate block lengths; iterative decoding cost was acceptable at the throughputs of the era |
| LDPC | 5G NR data, DVB-S2/S2X, Wi-Fi, storage | Excellent performance with a decoder that parallelises well, which is what makes multi-gigabit throughput achievable |
| Polar | 5G NR control channels | Provably capacity-achieving and better than LDPC at the short block lengths control signalling uses |
| Reed-Solomon | Optical, storage, and as an outer code in concatenated schemes | Corrects bursts rather than scattered bit errors, which is why it pairs with an inner code that leaves burst residue |
5G NR uses LDPC for data and polar for control in the same air interface, because block length changes which code wins.
No return path, or a return path too slow to use. Everything is decided at the transmitter.
| Constraint | What it forces | Consequence |
|---|---|---|
| Long round trip | Retransmission is impractical for most traffic | Coding is sized for the worst usable condition rather than the average |
| Very low signal-to-noise ratio | Long block lengths and low code rates | Large decoder memories and more iterations, both of which cost silicon |
| Varying conditions | Adaptive coding and modulation across the pass or the weather | The receiver must track and switch modes without losing lock |
| No return path at all | Every recipient must decode from the same transmission | The margin is set by the worst receiver in the population, not the typical one |
Concatenation is common here: an inner soft-decision code close to capacity, and an outer code clearing the residual errors the inner decoder leaves behind.
Adding structured redundancy at the transmitter so the receiver can repair errors without asking for a retransmission. It buys signal-to-noise margin — coding gain — which is usually cheaper than supplying the same margin with transmit power or antenna gain.
Because block length changes which code wins. LDPC performs excellently and parallelises well at the long block lengths data channels use, which is what makes high throughput achievable. Polar performs better at the short block lengths control signalling uses. Using one code for both would compromise one of them.
Memory, not logic. Iterative decoders pass messages between nodes and need large memories to hold them, and the HARQ soft buffer holds failed transmissions for combining. On real designs the decoder frequently dominates device selection for the whole receiver.
Feeding the decoder a confidence value per bit — a log-likelihood ratio — rather than a hard one or zero. It is worth several decibels of coding gain. How many bits those values are quantised to is a direct trade between performance and memory, and it is decided per stage.
Hybrid automatic repeat request: combining forward error correction with retransmission. With incremental redundancy the soft values of a failed transmission are retained and combined with the retransmission rather than discarded, which is why the soft buffer is often the largest memory in the receiver.
It corrects bursts rather than scattered bit errors, so it pairs naturally as an outer code with an inner soft-decision code that leaves burst residue behind. Concatenation of that kind is common in satellite, broadcast and storage.
It cannot wait for many decoding iterations or a long interleaver, so it accepts a lower code rate and more transmit resource instead. The right coding decision for bulk data is the wrong one for ultra-reliable low-latency traffic on the same air interface.
Where this is implemented.
CASE STUDYThe decoder that chose the device.
TECHNOLOGYWhere coding sits in the layer 1 chain.
TECHNOLOGYThe link budget coding is buying back.
PRODUCTThe datapath building block.
CAPABILITYClosing a decoder in the budgeted device.
Tell us the specification, the constraint and the deadline. Programmes that cross silicon, radio, embedded and AI are where Faststream is strongest.