112-Gbit/s PAM4 transmission is a high spectral efficiency modulation method using four amplitude levels, enabling 56 Gbaud operation for data center, PON, and fronthaul applications.
Advanced electro-optic modulators and equalization techniques—including FFE, DFE, MLSD, and machine learning methods—overcome hardware and channel impairments in these systems.
System performance is optimized by balancing drive voltage, insertion loss, and DSP complexity to achieve low BER across varied distances from 1 km to 80 km.
Pulse amplitude modulation with four levels (PAM4) at 112 Gb/s (56 Gbaud) is a fundamental technology for next-generation optical interconnects and access networks, providing high spectral efficiency within stringent power and complexity constraints. 112-Gbit/s PAM4 transmission occupies a key regime for both short-reach (intra/inter-data center), passive optical network (PON), and fronthaul applications, requiring advanced electro-optic devices, digital signal processing (DSP), and equalization techniques to overcome hardware and channel impairments. The following exposition reviews leading architectures, device technologies, equalization methodologies, system performance, and future directions based on published experimental demonstrations and analysis.
1. Device Technologies and Modulator Architectures
PAM4 at 112 Gb/s has been realized using several modulator/transmitter types: silicon dual-drive Mach-Zehnder modulators (DD-MZM) (Ruan et al., 2018), travelling-wave MZMs (TW-MZM) (Yang et al., 2018), silicon-organic hybrid (SOH) MZMs (Kieninger et al., 2020), directly-modulated long-wavelength VCSELs (Kerrebrouck et al., 2018), and electro-absorption modulators (EAMs) (Eiselt et al., 2018). Key parameters include electro-optic 3-dB bandwidth, drive voltage (Vπ​), insertion loss, and monolithic integration capability.
3. Digital Signal Processing and Equalization Techniques
Mitigating bandwidth constraints, chromatic dispersion, and device nonlinearity is critical in 112 Gb/s PAM4 systems.
Linear Equalization: Feedforward equalizers (FFE) are implemented with 41–51 taps (typically decision-directed LMS or Sato algorithms) at the receiver, with transmitter-side pre-emphasis using 11 taps for EMLs/VCSELs (Eiselt et al., 2018, Kerrebrouck et al., 2018, Shao et al., 2024).
Nonlinear and Advanced Equalization:
Decision-Feedback Equalizers (DFE): Employed with directly-modulated VCSELs/EMLs; e.g., 7+7 taps achieve FEC threshold in most MCF cores (Kerrebrouck et al., 2018).
Maximum Likelihood Sequence Detection (MLSD): Two-tap post-filter plus MLSD enables ISI mitigation; especially necessary with SSB+KK architectures and limited modulator bandwidth, with Euclidean branch metrics (Ruan et al., 2018, Yang et al., 2018).
Partial-Response PAM4 and MLSE: Encoders implement channel memory ($1+D$), decoded by 4-state MLSE, slightly improving BER at the cost of DSP complexity (Eiselt et al., 2018).
Machine Learning-based Equalizers:
Kolmogorov–Arnold Networks (KAN) and deep CNNs model nonlinearities (EAM/SOA), outperforming FIR and CNN baselines at a given computational budget (Fischer et al., 2024). For example, a 2-layer KAN achieves BER <10−3 at ROP∼−4.6 dBm using only 321 multiplications/symbol.
Fourier Convolution-based Network (FConvNet): Leverages frequency-domain attention, multi-periodicity partitioning, and Inception-style convnets; achieves 2 dB sensitivity gain over 51-tap Sato equalizers and 1 dB over MLAs at BER =5×10−3, at only 1.7k multiplications/symbol (Shao et al., 2024).
Equalizer Type
Complexity (mults/sym)
BER (∼−2 to −4 dBm ROP)
41-tap FFE
41–51
10−3–10−2
KAN-2
321
2.5×10−4–10−3
CNN-2
321
7.0×10−4–10−3
FConvNet (wl=64)
1,714
3.2×10−3 (2 dB gain over Sato)
Pre-equalization, post-filtering, and advanced sequence detection are all integral to optimizing 112 Gb/s PAM4 link performance.
4. System Performance and Transmission Metrics
Empirical performance of 112 Gb/s PAM4 depends on device bandwidth, channel impairment, DSP, and equalizer sophistication:
Short-range (1–2 km): BER <3.8×10−3 at ROP −11.5 dBm (1 km) and −8.5 dBm (2 km), corresponding to standard FEC thresholds (Yang et al., 2018).
Long-reach (80 km SSMF, SSB+KK): BER 2.46×10−3 with net data rate ∼102 Gb/s (including FEC) (Ruan et al., 2018).
PON/Fronthaul (2–10 km): BER 4×10−3 at −4 dBm (10 km), with ∼4 dB margin over FEC threshold (Eiselt et al., 2018).
SDM/VCSEL: At 56 Gbaud/1 km, five out of seven MCF cores achieve BER <3.8×10−3 using pre-equalization and DFE (Kerrebrouck et al., 2018).
Optimal launch power, OSNR, and equalizer parameters are set per link; for 80 km links +8 dBm launch (−9 dBm Rx), OSNR penalty is 3–3.2 dB over short links (Ruan et al., 2018).
5. Trade-Offs, Implementation, and Limitations
Major design trade-offs include:
Bandwidth-Limited Devices: TW-MZM and SOH-MZM devices often have intrinsic ∼21–40 GHz bandwidth, requiring pre-emphasis, post-filtering, and MLSD. Aggressive DSP mitigates penalties but elevates complexity and power usage (Ruan et al., 2018, Yang et al., 2018, Kieninger et al., 2020).
Equalization Complexity: Linear FFE/DFE and MLSD suffice below 100 multiplications/symbol, while ML methods (KAN, CNN, FConvNet) achieve higher sensitivity at moderate complexity (∼300–3k mults/sym) (Fischer et al., 2024, Shao et al., 2024).
Crosstalk (SDM/MCF): MCF cross-talk is <−45 dB over 100 km but core variation limits aggregate throughput (Kerrebrouck et al., 2018).
SSB+KK Cost/Benefit: SSB removes CD nulls but demands higher carrier-to-signal power ratio (CSPR∼16.6 dB) and OSNR (∼41.1 dB at 80 km), elevating transmitter power requirements (Ruan et al., 2018).
Link Loss and FEC: Tolerated loss budgets >20 dB (e.g., SOH-MZM), with 7% HD-FEC for <3.8e-3 BER, and net rates ≈104–112 Gb/s depending on configuration (Kieninger et al., 2020).
6. Future Directions and Research Opportunities
Further improvement of 112 Gb/s PAM4 systems will involve:
High-speed modulator technology: Extension to >30–70 GHz EO bandwidth via optimized doping/segmented electrodes for both MZM and VCSELs, reducing the need for aggressive DSP (Ruan et al., 2018, Kieninger et al., 2020).
Integrated transmitter/driver circuits: Co-packaged drivers and optical modulators to improve swing, reduce RF loss, and power consumption.
Advanced equalization: Extension of KAN and FConvNet/DNN approaches to adaptively compensate for multi-dimensional impairment at low complexity, with on-chip implementation via LUTs or DSP slices (Fischer et al., 2024, Shao et al., 2024).
Higher-order modulation: Exploration of PAM-6/8 and hybrid SSB-OFDM for spectral efficiency exceeding 112 Gb/s/lane (Ruan et al., 2018).
Photonic integration: Monolithic integration of laser, amplifier, and modulator for reduced coupling loss and overall cost.
Adaptive and nonlinear compensation: Improved chirp management for VCSELs and robust nonlinear equalization for EAM/SOA-based links in high-loss or high-penalty environments.
These directions aim to meet stringent BER, power, complexity, and integration requirements for intra- and inter-data center interconnects, backhaul/fronthaul, and next-generation PON, supported by ongoing device and DSP innovation.
