Designing for manufacture: PAM-4 transmitters using segmented-electrode Mach-Zehnder modulators

As PAM-4 signalling becomes the norm in datacentre optics, there is increasing focus on finding optimal design and implementation of the PAM-4 transmitter PICs. We use Synopsys’ OptSim software to simulate a design using segmented-electrode Mach-Zehnder modulators and investigate how manufacturing variations affect its performance.

By Jigesh Patel, product manager, and Pablo Mena, R&D engineer, Synopsys

In datacentre optics, 4-level Pulse Amplitude Modulation (PAM-4) signalling is gradually overtaking non-return-to-zero (NRZ) signalling [1-3]. Although both schemes use intensity modulation and direct detection, PAM-4 encodes two bits into four intensity levels, reducing bandwidth requirements for a given data rate by half. In other words, with PAM-4, transmission of a 40G signal requires components with 20 GHz bandwidth (corresponding to a symbol rate of 20 GBaud).

PAM-4 strikes a good balance between data carrying capacity and cost, since it requires less complicated digital signal processing than coherent transmission methods. Since lowering energy consumption is an economic and social imperative, designers need to be mindful of energy efficiencies when exploring technologies. This article explores the design of a PAM-4 transmitter chip that provides higher bandwidth compared to NRZ, avoids using energy-inefficient electronics, and achieves PAM-4 signalling in the optical domain by using segmented-electrode phase-shifters.

Figure 1. Schematic of a PAM-4 transmitter using SE-MZM

Commonly used PAM-4 transmitter PICs often suffer from poor energy efficiency and larger footprint. Traditionally, these transmitters incorporate a Mach-Zehnder modulator (MZM) being driven by an electrical digital-to-analogue converter (DAC) with an electrical driver. But the electronics required for this is highly energy inefficient – an ever-increasing concern in datacoms. Although there are alternative designs with nested modulators and drivers, a drawback is that they typically have larger footprints.

To overcome these challenges, we can accomplish a DAC-less design using the inherent DAC capabilities of segmented phase shifters [4]. Although conventional travelling-wave MZMs (TW-MZMs) can reduce drive voltage (VπL) due to a longer interaction length (L) thereby improving energy efficiency, the longer electrodes result in higher radio frequency (RF) losses and a mismatch between the group velocities of RF and optical signals. This mismatch, in turn, impacts modulation bandwidth [5].

Figure 2. SE-MZM PAM-4 transmitter output signal (left) and spectrum (right).

The segmented approach offers the advantage of longer interaction lengths without increasing loss, by shifting the velocity matching to electronic timing circuits, which control the timing of applied electrical signals to match the optical delay between segments [6]. Figure 1 shows a schematic of a PAM-4 transmitter using an SE-MZM made from discrete PIC elements. The design is built in Synopsys OptSim – an advanced photonic circuit simulation tool, which is part of the Synopsys electronic-photonic design automation (EPDA) toolset.

The topology comprises bidirectional PIC elements such as a 1-input 2-output optical splitter and a 2-input 1-output optical combiner with user-defined power ratio, and two pairs of travelling-wave optical phase shifters. The latter are used to implement the segmented MZM. Their lengths are binary weighted, meaning that each binary word can be applied directly. This minimises the number of segments needed, making integration less complicated.

The first segment’s length is one third of the total MZM length, while the second segment is two thirds of the total length. The 20G bit sequence has already been split into separate bit patterns corresponding to the most and least significant bits (MSB and LSB, respectively). The top driver modulates the first MZM segment using the LSB pattern, and the bottom driver modulates the second MZM segment using the MSB pattern.

Each travelling-wave phase shifter has an optical waveguide and a surrounding electrical transmission line that can change the waveguide’s refractive index and propagation loss. The interaction between the electrical and optical signals is distributed along the propagation direction. The waveguide’s thermal behaviour – and hence also the thermal behaviour of the modulator – can be modelled with the derivative of effective refractive index, the parameter VπL, and propagation loss as functions of temperature [7].

Figure 3. Optical eye diagram at the transmitter output.

Each MZM arm also has a phase tuner near the combiner, which sets the modulator at quadrature. A simple 90-degree phase-shifter model from OptSim in one of the arms could have performed the same job. However, for a packaged product, external controls allow for any additional tuning that might be needed, for example due to environmental factors or manufacturing tolerances.

The continuous-wave (CW) light is modulated by an electrical signal derived from pseudorandom binary signal (PRBS) data followed by an electrical driver. The inverter model provides a push-pull electrical bias to one of the electrodes relative to the other. A scope placed at the output of the SE-MZM PIC observes the PAM-4 modulated optical signal and its spectrum, depicted in Figure 2. Meanwhile, Figure 3 shows the four-level PAM-4 optical eye diagram at the transmitter’s output.
Impact of inter-segment distance deviations.

Figure 4. Schematic of a PAM-2 transmitter using SE-MZM.

While this segmented design concept offers a promising way forward for energy-efficient, small-footprint PAM-4 transmitters, several factors could impact the transmission signal quality. It’s particularly important to understand how deviations from the design introduced by the manufacturing processes can affect functionality.

For simplicity, we use an SE-MZM PAM-2 transmitter in OptSim, although the discussion also applies to higher-level PAM signals, including PAM-4. We analyse how performance varies with deviations in key parameters such as segment-to-segment distance, driver time-delay, and response time. All of these parameters can be affected by manufacturing and packaging process variations and have a direct impact on the overall yield of SE-MZM-based high-speed transmitter PICs.

In SE-MZMs, electrodes are split in segments of specific lengths to achieve the desired level of amplitude modulation. Whether the phase shifter lengths are binary weighted or thermometer coded [8], variations in the manufacturing process can result in deviations in the inter-electrodes, thereby affecting the MZM’s bias and bandwidth. OptSim is an excellent platform with which to conduct Monte Carlo analyses and achieve design for manufacturing [9].

Figure 4 shows a PAM-2 transmitter using an SE-MZM made from discrete PIC elements in OptSim. Similarly to the PAM-4 transmitter design discussed earlier, the topology comprises bidirectional PIC elements such as a 1-input 2-output optical splitter and 2-input 1-output optical combiner with user-defined power ratio, and pairs of travelling-wave optical phase shifters.

Figure 5. Optical eye diagrams at the transmitter output.

OptSim’s bidirectional waveguide model connects the phase shifters and implements deviations in the separation between the segmented electrodes. By setting a parameter scan for distances between the electrode segments, we can obtain the PAM-2 modulated optical eye diagrams shown in Figure 5. The bandwidth narrowing due to deviations in inter-electrode distances is clear: the higher the deviation, the worse the eye opening.

Impact of driver time-delay deviations
In SE-MZM transmitters, phase shifters are driven by electrical driving circuits. There are several manufacturing process variations and packaging constraints that can affect the device’s properties. For example, slight variations in the chip placement in a die can lead to differing lengths of the copper traces to electrical pads. These variations can result in electrical timing deviations in driving electrodes, thereby affecting the mismatch between RF and optical group velocities, and ultimately the bandwidth of the transmitter PIC.

We can study the impact of these driver time-delay deviations using a design for an SE-MZM PAM-2 transmitter similar to the one in the previous section. The OptSim schematic is shown in Figure 6. OptSim’s bidirectional waveguide model connects phase shifters and implements separations between the segmented electrodes.

The electrical time delay component models driver timing delay deviations as integer multiples of the sampling interval. For the design under study, this is 0.78125 picoseconds, corresponding to a data rate of 20 GBaud sampled at 64 samples per symbol. The lower arm of the MZM uses a discrete 90-degree phase-shifter model from OptSim to bias the MZM at quadrature.

Figure 6. Schematic of a PAM-2 transmitter using SE-MZM.

As with the schematic in Figure 1, the CW light is modulated by an electrical signal derived from PRBS data followed by an electrical driver, and the inverter model provides a push-pull electrical bias to one of the electrodes relative to the other. At the output of the SE-MZM PIC, an eye diagram analyser and a signal analyser are connected to observe the transmitter signal. A back-to-back receiver at the SE-MZM PIC’s output detects the received signal, and the bit error rate (BER) tester is used to estimate the penalties induced on the BER due to driver time delay.

A parameter scan is set for driving delays at segment electrodes as integer multiples of the sampling interval. Figure 7 shows the PAM-2 modulated optical eye diagrams obtained for different values of drive delays. The resulting bandwidth penalties can also be visualised as back-to-back BER at the transmitter chip as shown in Figure 8. Optimal performance is achieved when the optical and RF group velocities match. This analysis can help designers find the optimal trace lengths to obtain performance within acceptable bounds.

Figure 7. Optical eye diagrams at the transmitter output.

Capacitive charge and R-C response times
Another important design consideration in SE-MZMs is the R-C response time of phase-shifter electrodes, which directly impacts the bandwidth of the transmitter PIC. The junction capacitance of the p-n diode and parasitics from packaging both contribute to the capacitive charge and R-C time limitations in a phase shifter. As shown in the schematic in Figure 9, we lump both effects into the overall R-C response time of the phase shifter.

After setting the electrical response model of the phase shifter to “RC,” as shown in Figure 10, a parameter scan can be set for capacitance to model different amounts of capacitive charge and R-C response times. Figure 11 shows the PAM-2 modulated optical eye diagrams resulting from different values of R-C response times. As expected, higher response times adversely affect modulation speed.

Despite the three challenges outlined above, SE-MZM-based DAC-less transmitter PICs are nonetheless preferable in many contexts, due to their small footprint and high energy efficiency. However, to fulfil their potential, it is essential to consider the impacts of deviations in some of the key design parameters arising from packaging, technology, and manufacturing process variations. OptSim is a powerful tool for understanding these effects and thus achieving design for manufacturing and creating high-performance, reliable products.

Figure 8. Bit error rate (BER) as a function of driving time delay.


Although this article has described optical performance in terms of eye diagrams at the transmitter, it is also possible to measure it in terms of performance penalty. Two of the most common performance metrics in NRZ transmitter design are extinction ratio and the dispersion penalty at maximum dispersion. In the case of PAM-4 transmitters, the equivalent quantities are optical modulation amplitude (OMA) and dispersion eye closure penalty quaternary (TDECQ) [10].

Figure 9. OptSim schematic of a PAM-2 transmitter using SE-MZM.

For the specified value of symbol error rate (SER), the TDECQ algorithm seeks to determine the maximum amount of noise that can be added to the input signal while still meeting the target SER, and then compare this to the amount of noise that could be added if the signal were ideal.

Figure 10. Choice of electrical response model via the parameter window of the phase shifter model.

The TDECQ measurement functionality in Synopsys OptSim also produces plots of several quantities of interest, in addition to the TDECQ, to help with a detailed assessment of the transmitter compliance and equaliser performance. These additional transmitter performance parameters include optimised decision shift, reference-equaliser noise enhancement, OMA, and noise characteristics of the incoming and ideal signals. A detailed discussion of TDECQ measures is beyond the scope of this article, but more information is available in reference [11].



Figure 11. Optical eye diagrams at the transmitter output.