Inside a hyperscale data center, optical fiber carries the vast majority of traffic that moves between servers, switches, and storage systems. The distances involved — from a few meters within a rack to several hundred meters across a campus — define a distinct engineering problem that is quite different from long-haul or metro networking. These links demand extreme simplicity, very low power consumption, low cost per gigabit, and massive volume deployability. Two design choices dominate this space: intensity modulation (IM) to generate the optical signal, and direct detection (DD) to recover it. The combination, universally abbreviated as IM/DD, removes the need for a local oscillator laser and coherent receiver hardware that long-haul systems require.

Within the IM/DD family, the format of choice for links operating at 100 Gb/s per lane and above is four-level pulse amplitude modulation (PAM-4). PAM-4 encodes two bits per symbol, halving the required electrical bandwidth compared with a binary non-return-to-zero (NRZ) signal at the same bit rate. As data center capacity generations have escalated from 100G to 400G to 800G and now toward 1.6 Tb/s per module, PAM-4 IM/DD has become the single most widely deployed optical modulation format in terms of shipped lane-count globally. IEEE 802.3dj, ratified to address 200 Gb/s, 400 Gb/s, 800 Gb/s, and 1.6 Tb/s Ethernet over a variety of reach classes, enshrines PAM-4 as the mandatory modulation format for virtually all direct-detect applications.

This article provides a complete engineering reference covering the physical principles of PAM-4 IM/DD, the architecture of a production link, quantitative performance analysis including the BER–SNR relationship and chromatic dispersion budget, the evolving modulator technology landscape through the state of the art as of 2026, and the deployment considerations that translate theory into working hardware. The goal is to serve the engineer who needs to design, specify, or troubleshoot these links, not merely describe them.

2.1 Amplitude Modulation and Direct Detection

Intensity modulation maps digital data onto the instantaneous optical power of a laser. A driver circuit translates electrical voltage levels into corresponding optical power levels at the output of the laser or external modulator. At the receiver, a photodiode converts optical power back into electrical current through the square-law relationship I = R·P, where R is the responsivity in A/W. This square-law detection is what makes IM/DD fundamentally different from coherent detection: the receiver responds only to intensity, discarding phase entirely. As a consequence, chromatic dispersion imposes a power-fading penalty in the C-band (around 1550 nm) that grows as the square of the symbol rate, which is one of the primary reasons O-band operation at approximately 1310 nm is preferred for most IM/DD data center links.

2.2 PAM-4 Signal Structure

PAM-4 uses four amplitude levels — typically represented as 0, 1, 2, and 3 in normalized units, or as optical power levels P₀, P₁, P₂, P₃ — that carry two bits per symbol. Gray coding maps adjacent symbol levels to codewords differing by exactly one bit, minimizing the impact of level-to-level symbol errors on the recovered bit error ratio (BER). Because each symbol carries two bits, a 112 Gbps link requires a symbol rate of only 56 GBaud, halving the electrical bandwidth demand compared with the equivalent NRZ signal. This electrical bandwidth reduction is the core motivation for adopting PAM-4 over the simpler on-off keying (OOK) or NRZ format at high data rates.

The penalty for this bandwidth efficiency is noise tolerance. In a PAM-4 signal, the spacing between adjacent amplitude levels is one-third of the total optical swing, compared with the full swing available in a two-level (OOK) system with the same total launch power. This reduced eye opening means PAM-4 requires approximately 9.5 dB more signal-to-noise ratio (SNR) than OOK to achieve the same BER in additive white Gaussian noise — a fundamental trade-off that drives the stringent optical power budget requirements in PAM-4 link design.

2.3 Forward Error Correction and Operating Thresholds

PAM-4 IM/DD links do not operate at the raw hardware BER threshold of 10⁻¹² or lower that older systems targeted before the receiver decision circuit. Instead, they operate above a forward error correction (FEC) threshold, relying on the FEC codec to reduce post-FEC BER to effectively zero. The industry has standardized on three primary pre-FEC BER thresholds for different FEC overhead levels. Hard-decision FEC with 7% overhead (HD-FEC) carries a pre-FEC BER threshold of 3.8×10⁻³; hard-decision FEC with 20% overhead targets approximately 1.4×10⁻²; and soft-decision FEC with 25% overhead extends to approximately 2.4×10⁻². The 7% HD-FEC threshold at 3.8×10⁻³ is the baseline for IEEE 802.3 short-reach specifications, while the higher-overhead FEC variants are adopted in research demonstrations pushing toward and beyond 400 Gb/s per lane.

A PAM-4 IM/DD link consists of five functional blocks: the transmitter optoelectronic assembly, the optical fiber, any inline amplification or filtering, the receiver optoelectronic assembly, and the digital signal processing (DSP) engine shared between transmitter and receiver. The figure below shows the canonical architecture of a single-lane 112 Gbps (56 GBaud PAM-4) link as deployed in current-generation 400G/lane DR4/FR4 transceivers and the emerging 800G/lane implementations.

3.1 Transmitter Components

The transmitter DSP performs three primary functions: FEC encoding, PAM-4 Gray mapping, and pre-emphasis filtering. Pre-emphasis (also called feed-forward equalization at the transmitter, or Tx-FFE) intentionally boosts high-frequency components of the transmitted signal to compensate for the bandwidth roll-off of the driver, modulator, and fiber channel. This is especially valuable because practical modulators have bandwidths that fall well short of the Nyquist frequency at high baud rates — for example, a 56 GBaud signal has a Nyquist frequency of 28 GHz, but production EML devices typically exhibit −3 dB bandwidths in the 40–60 GHz range, and the combined electrical path including driver and interconnects typically limits effective bandwidth to 35–45 GHz at these rates. A digital-to-analog converter (DAC) with a sample rate of twice the symbol rate (or higher) generates the pre-emphasized analog waveform that drives the modulator.

The modulator converts the electrical PAM-4 waveform into optical intensity variations. Three technologies compete in this space as of 2026. Electro-absorption modulated lasers (EMLs) integrate a distributed feedback (DFB) laser with an electroabsorption section on a single chip, offering compact size and maturity for rates up to approximately 200 Gb/s per lane. Silicon photonics (SiPh) Mach-Zehnder modulators leverage standard CMOS foundry processes to deliver low-cost, high-integration transceivers, with demonstrated bandwidths above 80 GHz enabling single-lane PAM-4 rates of 256–300 Gb/s and beyond. Thin-film lithium niobate (TFLN) Mach-Zehnder modulators represent the leading edge, with electro-optic bandwidths exceeding 110 GHz, drive voltages below 2 V, and demonstrated PAM-4 transmission at 420 Gb/s per lane as presented at OFC 2026 — the highest per-lane PAM-4 rate reported without a 225 GBaud SerDes at the time of that publication.

3.2 Fiber Channel and Wavelength Selection

The fiber medium for data center IM/DD links is almost always OM4 or OM5 multimode fiber for reaches up to 100–400 m with directly modulated VCSELs at 850 nm, or single-mode fiber (G.652/G.657) for reaches from 500 m to 10 km. Within single-mode links, the O-band around 1310 nm is strongly preferred because G.652 fiber has a dispersion zero near 1310 nm, which means the chromatic dispersion coefficient D is approximately 0 ps/(nm·km) at this wavelength. This near-zero dispersion eliminates the power-fading penalty that would otherwise make high-baud-rate IM/DD impractical. For links that must use C-band wavelengths — for example, to multiplex many channels onto a single fiber using conventional DWDM multiplexers — dispersion-shifted fiber (DSF, ITU-T G.653) or optical pre-compensation at the transmitter is required.

3.3 Receiver Components

The receiver front-end converts the arriving optical signal back into an electrical waveform. A Ge-on-Si PIN photodiode — or in some implementations a Ge-on-Si avalanche photodiode (APD) — converts photons to current with a responsivity typically between 0.7 and 0.9 A/W at 1310 nm. Recent work demonstrated SiGe photodetectors supporting 170 GBaud and achieving 300 Gbps net throughput under HD-FEC and 400 Gbps under 25% soft-decision FEC. A transimpedance amplifier (TIA) converts the photocurrent to voltage and provides the first stage of electrical gain, followed by an ADC that digitizes the waveform at approximately one or two samples per symbol for digital equalization. The receiver DSP applies a feed-forward equalizer (FFE) to correct residual linear impairments, and optionally a decision-feedback equalizer (DFE) or more sophisticated neural-network-based equalizer to address nonlinear distortions from the modulator or photodetector. FEC decoding then produces the final recovered bit stream.

4.1 BER–SNR Relationship for PAM-4

For PAM-4 in a channel dominated by additive white Gaussian noise (AWGN), the theoretical BER as a function of signal-to-noise ratio is given by a standard closed-form expression derived from the spacing between adjacent amplitude levels.

// PAM-4 BER in AWGN (Gray-coded, equal spacing)

BERPAM4  =  (3/4) · erfc( sqrt(SNR / 5) )

Where:
  SNR   = signal-to-noise ratio (linear, not dB)
         = (3 · Ppeak²) / (5 · σ²n)   [for equal-spacing PAM-4]
  erfc  = complementary error function
  σn   = rms noise amplitude (TIA thermal + shot)

// Penalty vs. OOK at same BER:
ΔSNRPAM4/OOK  ≈  10·log10(5)  =  +6.99 dB   (theory, equal power)
                                            +9.5 dB    (practical, incl. bandwidth penalty)

// For 7% HD-FEC threshold: BER = 3.8×10⁻³
// Solve: erfc(x) = 4/3 × 3.8×10⁻³ → x ≈ 1.57
SNRrequired  ≈  5 × (1.57)²  =  12.3  →  10.9 dB

// Worked example: 56 GBaud PAM-4 at HD-FEC threshold
  Target BER     = 3.8×10⁻³
  Required SNR   ≈ 10.9 dB   (electrical, at sampler)
  Launch power   ≈ +3 dBm    (typical EML, O-band)
  Fiber loss     ≈ 2.0 dB    (500 m @ 0.4 dB/km insertion loss)
  Coupling loss  ≈ 2.5 dB    (connectors + splices)
  Received power ≈ −1.5 dBm  → within sensitivity of PIN+TIA

4.2 TDECQ — Transmitter Dispersion Eye Closure

IEEE 802.3 uses Transmitter and Dispersion Eye Closure Quaternary (TDECQ) as the primary compliance metric for PAM-4 optical transmitters. TDECQ quantifies how much the eye diagram closes relative to a reference receiver's noise floor, expressed in dB. A lower TDECQ indicates a higher-quality transmitter. IEEE 802.3dj specifies TDECQ ≤ 3.4 dB for 200 Gb/s per-lane DR class transceivers. Recent demonstration hardware at OFC 2026 reported TDECQ values of 2.76–3.83 dB for TOSA (Transmitter Optical Sub-Assembly) designs operating at 180–210 GBaud, confirming that TFLN-based transmitters can meet 400 Gb/s per-lane specifications.

4.3 Optical Power Budget

// PAM-4 IM/DD Optical Power Budget

Pbudget  =  Plaunch  −  Psensitivity   [dBm]

Plaunch  =  laser output − modulator insertion loss − connector loss
Psensitivity  =  minimum received power to meet BER threshold

Losses to account for:
  Fiber attenuation   : α × L           [dB]   α ≈ 0.35 dB/km at 1310 nm
  Connector loss      : N × 0.5–1.0       [dB]   per mated pair
  Splice loss         : N × 0.1–0.2       [dB]   per mechanical splice
  Dispersion penalty  : ~0 in O-band      [dB]   (≤ 0.5 dB for 10 km at 56 GBaud)
  Aging margin        : 1–2               [dB]

// Practical example: DR1 (500m, 400G per lane, O-band)
  Launch power (TFLN TOSA, typ.)  =  +6 dBm
  Fiber loss (0.5 km × 0.35)      =  0.18 dB
  2× connector pairs               =  1.0 dB
  Aging margin                     =  1.5 dB
  Total loss                       =  2.68 dB
  Received power                   =  +3.32 dBm   (well above −5 dBm sensitivity)
  Margin to sensitivity            =  8.3 dB

4.4 BER Performance Chart

The chart below illustrates BER as a function of received optical power (ROP) for a representative 56 GBaud PAM-4 IM/DD link with three different equalizer configurations: a simple 5-tap feed-forward equalizer (FFE), a 31-tap FFE plus 5-tap DFE, and a neural network equalizer. The 7% HD-FEC threshold at BER 3.8×10⁻³ and the 20% HD-FEC threshold at BER 1.4×10⁻² are marked for reference. This pattern — neural network equalization providing 1–2 dB sensitivity improvement over a simple FFE — is consistent with experimental results from OFC 2026, where physically-assisted AI equalizers demonstrated measurable gains over conventional adaptive equalizers for bPAM-4 IM/DD systems.

5.1 IEEE 802.3 Reach Classes

IEEE 802.3dj defines several PAM-4 IM/DD reach classes that specify maximum fiber length, minimum launch power, maximum receiver sensitivity, and TDECQ limits. The DR (Data Rate) prefix indicates O-band single-mode fiber applications. DR1 (single lane, 500 m), DR4 (four lanes WDM over a single fiber pair, 500 m), FR4 (four lanes WDM over a single fiber pair, 2 km), and LR4 (four lanes, 10 km) cover the majority of data center campus and intra-data-center needs. The table below summarises the key specifications for these classes at 400G module level.

Table 1: Summary of IEEE 802.3dj PAM-4 IM/DD reach class specifications for 400G–1.6T modules. Per-lane rate is 100 Gb/s PAM-4 (56 GBaud) for all entries; 200 Gb/s per-lane variants (112 GBaud) are defined in the same standard for next-generation module architectures. Rx sensitivity is maximum acceptable (i.e., worst-case) value.

5.2 Multi-Path Interference and Connector Quality

A practical issue in data center PAM-4 links that does not appear in coherent systems is multi-path interference (MPI). Reflections from fibre connectors, fusion splices, and end faces create delayed, attenuated copies of the signal that interfere with the main signal at the photodetector. Because PAM-4 has finely spaced amplitude levels, even small amplitude perturbations from MPI cause disproportionate BER degradation. Research at OFC 2026 demonstrated that at MPI levels above −38 dB, the BER of a 53.125 GBaud PAM-4 signal at 1310 nm begins to degrade significantly. This sets practical requirements on connector return loss (typically ≥ 26 dB for UPC, ≥ 60 dB for APC connectors) and on the use of angle-polished connectors in high-power-budget, longer-reach links.

5.3 DSP Equalizer Selection

The choice of equalizer architecture has a direct impact on both link performance and ASIC power consumption. A simple feed-forward equalizer (FFE) with 5–15 taps adds only modest power overhead and handles linear impairments from bandwidth-limited optics well. Adding a decision-feedback equalizer (DFE) of 3–5 taps provides nonlinear compensation but introduces error propagation risk at high BER — undesirable near the FEC threshold. Maximum-likelihood sequence estimation (MLSE) with a Viterbi decoder provides near-optimal detection but scales in complexity as 4^m where m is the channel memory length, making it prohibitive above a few symbols of memory. Neural network equalizers — including feed-forward NN architectures, recurrent variants, and look-up table (LUT) approaches — have demonstrated 1–2 dB sensitivity gains over conventional FFE at practical complexity levels, a result confirmed experimentally at OFC 2026 using bipolar PAM-4 systems with physically-assisted AI.

Table 2: Comparison of modulator technologies for short-reach PAM-4 IM/DD links as of 2026. Bandwidth figures are −3 dB electro-optic bandwidth from representative published results. "Peak per-lane rate" refers to demonstrated PAM-4 transmission rates at OFC 2025–2026. CPO = co-packaged optics.

The trajectory for IM/DD in data centers follows two parallel paths: increasing per-lane rate and increasing aggregate capacity per fiber or module. On the per-lane dimension, demonstrations at OFC 2026 showed PAM-4 links operating at 420 Gb/s per lane using TFLN modulators and at 400 Gb/s per lane using GeSi electro-absorption modulators on standard 300 mm CMOS silicon photonic platforms — both without relying on 225 GBaud SerDes chips, which represent the next electrical interface generation. Once 225 GBaud SerDes become commercially available, 448 Gb/s per lane (PAM-4 at 224 GBaud) is anticipated to enable 3.2 Tb/s per module with 8 lanes.

On the aggregate capacity dimension, wavelength-division multiplexed IM/DD systems using optical comb sources are demonstrating impressive results. A 2026 demonstration used 25 comb lines from a quantum-dot mode-locked laser at 100 GHz spacing to carry 3.2 Tbps PAM-4, 4.3 Tbps PAM-6, and 5.76 Tbps PAM-8 over 2 km using a single semiconductor optical amplifier. While PAM-8 and higher formats increase spectral efficiency, they also impose increasingly strict SNR requirements and are sensitive to received optical power variations — PAM-4 tolerates approximately 1.5 dB ROP reduction before exceeding the FEC threshold, while PAM-8 tolerates only 0.2 dB at similar symbol rates, making it fragile for production deployment without careful power management.

Artificial intelligence applied to DSP represents perhaps the most significant near-term opportunity for improving PAM-4 IM/DD link performance without hardware changes. AI equalizers trained offline and deployed as look-up tables or lightweight neural network inference engines can handle the nonlinear transfer functions of high-speed modulators and detectors more accurately than linear FFE/DFE structures, with sensitivity gains of 1–2 dB translating directly to greater link margin or longer reach. The future evolution of PAM-4 IM/DD will therefore be as much a software and DSP discipline as a photonic hardware one.