Electronics Guide

Articles in This Category

Why More Bits per Symbol

The relationship at the center of this subject is short: the bit rate equals the symbol rate multiplied by the base-two logarithm of the number of distinct symbols. Two levels carry one bit per symbol, four levels carry two, eight levels carry three. Because the highest fundamental frequency a baseband link produces is approximately half the symbol rate, often called the Nyquist frequency, adding levels lowers the frequency at which the channel must still deliver usable signal.

That matters because insertion loss is not flat. Conductor loss climbs roughly with the square root of frequency and dielectric loss climbs roughly in proportion to it, so a channel that is merely lossy at 10 GHz can be effectively opaque at 30 GHz. Trading symbol rate for levels moves the link back onto the gentler part of that curve. Information theory frames the same choice: for a fixed bandwidth, capacity grows only with the logarithm of the signal-to-noise ratio, so every additional bit per symbol demands a disproportionately larger share of margin. Practical modulation design is the search for the point on that curve where the channel's available bandwidth and the receiver's achievable signal-to-noise ratio are both used efficiently.

Multi-Level Amplitude Signaling

Pulse amplitude modulation is the dominant multi-level family in wireline links because it keeps the signal at baseband and requires no carrier, mixer, or phase reference. The transmitter drives one of several discrete amplitudes each unit interval, and the receiver compares the incoming voltage against a set of thresholds.

The PAM-4 Trade-off

PAM-4 carries two bits per symbol on four levels and therefore halves the symbol rate needed for a given bit rate. The price appears immediately in the eye diagram. Where two-level signaling shows one eye spanning the full swing, PAM-4 shows three stacked eyes, each one third as tall, and the receiver must resolve all three. Gray coding limits the damage: adjacent levels are assigned code words that differ in a single bit, so the most probable error, a slip into a neighboring level, corrupts only one bit rather than two.

PAM-4 is also far less forgiving of nonlinearity than binary signaling. Unequal level spacing directly consumes margin in one of the three eyes, which is why Ethernet transmitter specifications include a level-separation mismatch parameter alongside the usual amplitude and jitter limits. Reflections, crosstalk, and residual intersymbol interference all become proportionally more damaging once the eye is one third as tall, so PAM-4 links lean heavily on transmitter feed-forward equalization, receiver continuous-time and decision-feedback equalization, and, at the highest rates, full digital signal processing behind a multi-gigasample analog-to-digital converter.

Adoption is now broad. Ethernet moved to PAM-4 for its 50 and 100 Gb/s electrical lanes, with 100 Gb/s lanes running at 53.125 gigabaud, and IEEE 802.3df, approved in 2024, built 800 Gb/s Ethernet from eight such lanes. The IEEE P802.3dj project carries the same signaling to roughly 106 gigabaud for 200 Gb/s lanes and a 1.6 Tb/s aggregate rate, with parallel electrical work in the OIF Common Electrical Interface projects. The retained P prefix is not incidental: 802.3dj is still an open task force rather than a published amendment, so its parameters should be read as drafts until the standard is approved. PCI Express adopted PAM-4 at PCIe 6.0, which reached 64 GT/s, and the PCIe 7.0 specification released in June 2025 doubles that to 128 GT/s on the same modulation.

Three Levels, and More Than Four

Four is not the only useful count. USB4 Version 2.0 uses PAM-3, which offers three levels and thus slightly more than 1.58 bits per symbol in principle; the specification maps 11 binary digits onto 7 ternary symbols and runs the wire at 25.6 gigabaud to reach 40 Gb/s per lane and 80 Gb/s in each direction across two lanes. Ternary signaling is a deliberate middle ground: it keeps a taller eye than PAM-4 while still holding the baud rate below what a passive cable of useful length can no longer support.

Twisted-pair Ethernet went further in the other direction long ago. 1000BASE-T transmits five-level symbols on four pairs at 125 megabaud, combined with four-dimensional trellis coding, and 10GBASE-T uses sixteen levels with a two-dimensional constellation and transmitter precoding. Those channels tolerate many levels because their bandwidth is tightly constrained by the cable while sophisticated echo and crosstalk cancellation keeps the noise floor low. On lossy board and cable channels, the returns fade quickly: each doubling of the level count adds one bit per symbol but costs roughly 6 dB of signal-to-noise ratio, which is why board-level links have stopped at four levels rather than climbing to eight or sixteen.

Controlled Intersymbol Interference

Partial-response signaling takes the opposite approach to equalization. Rather than fighting the intersymbol interference a low-pass channel imposes, it accepts a precisely known amount of it and designs the receiver to remove it by arithmetic. Duobinary, the classic case, adds each symbol to its immediate predecessor. A binary input therefore produces a three-level output, and the resulting transfer function places a null at half the symbol rate and concentrates the transmitted energy at lower frequencies, exactly where a lossy channel is most transparent.

The naive receiver for such a scheme suffers error propagation, because each decision depends on the previous one. Precoding at the transmitter solves this: the data is differentially encoded before the partial-response filter, which makes each received symbol decodable on its own. More elaborate partial-response classes, paired with maximum-likelihood sequence estimation using the Viterbi algorithm, let the receiver evaluate whole sequences against the known channel memory instead of slicing symbols independently.

Optical transport used duobinary at 10 and 40 Gb/s because the narrower spectrum improves tolerance to chromatic dispersion and allows tighter wavelength spacing. In electrical links, the same reasoning applies wherever the channel already behaves as a strong low-pass filter, since matching the transmitted spectrum to the channel is cheaper than boosting the high frequencies that the channel has already destroyed.

Coding as Part of the Link Budget

Multi-level signaling raises the raw error rate, and no amount of equalization drives an electrical link to a raw error rate of one in a quadrillion. Forward error correction closes the gap. The transmitter appends structured redundancy, and the receiver uses it to detect and correct errors without asking for retransmission, converting a moderate raw error rate into an extremely low corrected one.

High-speed Ethernet standardized on a Reed-Solomon code commonly called KP4: RS(544,514) over 10-bit symbols, which appends 30 parity symbols to 514 data symbols and corrects up to 15 symbol errors per codeword. The overhead is a little under 6 percent, and the code takes a pre-correction error rate on the order of one in ten thousand down below one in a quadrillion. That single code is what makes 100 Gb/s PAM-4 lanes practical, and it is specified as part of the 400 Gb/s Ethernet physical layer rather than left to the implementer.

Correction is never free. Encoding and decoding consume power and add latency, which is tolerable in a network interface but costly in a processor or memory interconnect. PCI Express answered that constraint by pairing a deliberately lightweight forward error correction with a strong cyclic redundancy check and link-level retry, so that the common case is corrected in place with minimal delay while rare uncorrected errors trigger a replay. Longer-reach optical systems move in the other direction, adopting soft-decision decoding and stronger codes such as low-density parity-check families, where the extra latency and power buy reach that no equalizer could supply.

The strongest designs treat coding and modulation as one problem rather than two. Trellis-coded modulation, used in 1000BASE-T, expands the constellation and constrains the allowed sequences so that the coding gain more than repays the added levels, which is the general principle behind coded modulation in both wireline and wireless systems.

Orthogonality and Multi-Carrier Methods

Orthogonal signaling rests on a different property. Two waveforms are orthogonal when their product, integrated over a symbol period, is zero, so a receiver that correlates against one recovers nothing from the other. That property lets many signals occupy the same band simultaneously and still be separated.

The most widespread application is orthogonal frequency-division multiplexing, and its wireline sibling, discrete multitone. Both divide a wide band into many narrow subcarriers spaced at the reciprocal of the symbol period. Each subcarrier sees a channel that is nearly flat across its own narrow slice, so equalization reduces to one complex coefficient per subcarrier instead of a long time-domain filter, and a cyclic prefix absorbs the channel's delay spread so that successive symbols do not overlap. Digital subscriber lines, DOCSIS cable systems, Wi-Fi, LTE, and 5G New Radio all rest on this structure, and modern versions add per-subcarrier bit loading so that clean parts of the spectrum carry denser constellations than noisy parts.

Multi-carrier methods have not displaced single-carrier PAM inside equipment, and the reasons are instructive. Summing many subcarriers produces a high peak-to-average power ratio, which demands linear drivers and wastes headroom. The transforms at both ends cost power, area, and latency. Orthogonality itself is fragile, degrading with frequency offset and phase noise. Board-level channels are also smoothly low-pass rather than notched, so the main advantage of multi-carrier signaling, graceful handling of deep frequency-selective fades, has little to offer. Orthogonality remains central elsewhere, however, underpinning code-division multiple access, the spatial streams of multiple-input multiple-output radio, and the polarization- and phase-multiplexed constellations that coherent optical transceivers use on long-haul fiber.

Choosing a Scheme

Selecting a modulation is a budget exercise, not a matter of taste. For a target bit rate, the designer compares the channel's insertion loss at each candidate Nyquist frequency against the signal-to-noise ratio that each scheme requires after equalization, then confirms that the remaining margin covers reflections, crosstalk, jitter, and manufacturing variation.

Two-level signaling stays the correct answer wherever the channel is short and clean. It offers the tallest eye, the simplest receiver, the lowest latency, and the least power, and it needs no forward error correction at moderate rates. PAM-4 becomes attractive precisely when loss at the two-level Nyquist frequency exceeds what equalization can recover, which on typical printed circuit board channels arrives somewhere above 25 to 50 Gb/s per lane depending on materials and reach. Partial-response methods suit channels whose response is already steeply low-pass, and ternary schemes such as PAM-3 occupy the middle ground where a passive cable must remain viable. Forward error correction is effectively mandatory once per-lane rates pass the 50 Gb/s mark, and multi-carrier methods win where the channel is frequency-selective in a way that board traces are not.

Every step in complexity is also paid for in silicon. A PAM-4 receiver with digital equalization needs a fast, moderately deep analog-to-digital converter and a substantial digital back end, and both scale badly with speed. The best link designs therefore push the channel, the connector, and the material as far as they will go before reaching for extra levels, because a better channel is almost always cheaper than a bigger receiver.

Conclusion

Advanced modulation is the discipline of spending margin where it is most plentiful. Multi-level amplitude signaling converts noise margin into bandwidth relief, partial-response methods reshape the transmitted spectrum to match a channel rather than fight it, forward error correction converts a small amount of throughput into a large amount of reliability, and orthogonal signaling converts computation into immunity from frequency-selective channels. None of these techniques works alone at the rates modern systems demand; a contemporary 100 Gb/s lane combines PAM-4 signaling, multi-stage equalization, and Reed-Solomon correction as a single co-designed system. The topics in this category examine each element in turn, from the two-level baseline through the multi-level, coded, and orthogonal methods that extend a link far past the rate its bare channel would allow.

Related Topics