200 Gbps/Lane IM/DD Technologies for Short Reach Optical Interconnects

Client-side optics are facing an ever-increasing upgrading pace, driven by upcoming 5G related services and datacenter applications. The demand for a single lane data rate is soon approaching 200 Gbps. To meet such high-speed requirement, all segments of traditional intensity modulation direct detection (IM/DD) technologies are being challenged. The characteristics of electrical and optoelectronic components and the performance of modulation, coding, and digital signal processing (DSP) techniques are being stretched to their limits. In this context, we witnessed technological breakthroughs in several aspects, including development of broadband devices, novel modulation formats and coding, and high-performance DSP algorithms for the past few years. A great momentum has been accumulated to overcome the aforementioned challenges. In this article, we focus on IM/DD transmissions, and provide an overview of recent research and development efforts on key enabling technologies for 200 Gbps per lane and beyond. Our recent demonstrations of 200 Gbps short-reach transmissions with 4-level pulse amplitude modulation (PAM) and discrete multitone signals are also presented as examples to show the system requirements in terms of device characteristics and DSP performance. Apart from digital coherent technologies and advanced direct detection systems, such as Stokes–vector and Kramers–Kronig schemes, we expect high-speed IM/DD systems will remain advantageous in terms of system cost, power consumption, and footprint for short reach applications in the short- to mid- term perspective.


I. INTRODUCTION
I NTENSITY modulation direct detection (IM/DD) technol- ogy had been the dominating solution in all segments of fiber-optic communication networks since its first generation.However, approximately one decade ago, transceivers based on the digital coherent optical technology were demonstrated and developed [1], and have quickly replaced the IM/DD solutions in many scenarios, in particular, the core/metro segments.To date, there has been a continuous process of upgrading fiber-optic networks from the legacy 10 Gbps systems to 100/200 Gbps, and soon 400 Gbps coherent based solutions on the line side for long-haul and metro networks, following the requirements of telecommunication carriers [2].Together with the demand of 5G related applications from the access networks, such upgrades push the data rate of IM/DD based client-side optics to match the line-side transponder speed by supporting 400 Gbps capacity and beyond [3].Besides such a continuous evolution of traditional telecom networks, a more demanding scenario for IM/DD interfaces with a higher date rate has recently been driven by the content providers, who often have their own network infrastructure in the form of datacenters [4].Nowadays, datacenters experience an enormous traffic growth due to the vast amount of data to be stored, transmitted and processed.These high-speed data links can be divided into two groups: intra-datacenter and inter-datacenter links.The first group consists short-reach data links covering distances ranging from a few meters to a few kilometers, connecting servers and racks inside the datacenters.The second group is also often referred as datacenter interconnects (DCI), which enable data exchange among multiple datacenters with much longer distances compared with the intra-datacenter links, normally in the scale of from a few kilometers up to a few hundred kilometers [5].Various industrial standards and multi-source agreement (MSA) groups are established to propose transceiver specifications for these application scenarios [6]- [8].For example, regarding the supported distance, the 100 Gigabit Ethernet (100 GbE) transceiver standards can be categorized into short reach (SR) segment, supporting up to 100 m OM4 multimode fiber (MMF) links; datacenter reach (DR), supporting up to 500 m single mode fiber (SMF) links; fiber reach (FR) for up to 2 km SMF, long reach (LR) for up to 10 km SMF; and extended reach (ER), offering up to 40 km SMF.For the upcoming 400 G era, IM/DD solutions are still prioritized for these specifications due to advantages in cost, power consumption, and footprint [9]- [11].
Fig. 1 shows a typical core/metro and access network scenario, covering most of the applications where high-speed IM/DD links are required.Currently, the boundary separating the digital coherent and the IM/DD solutions exists arguably between the metro edge and the access networks.However, this boundary is becoming less clear in the 400 G era and beyond, as both technical and economic challenges arise to keep up with the bandwidth growth for the IM/DD solution.Meanwhile, low-cost digital coherent solutions are proposed and being developed to cover medium to short reach applications, e.g., the 400 G ZR solution which is to be carried by the QSFP-DD-DCO (Quad Small Form Factor Pluggable Interface Double Density Digital Coherent Optics) or the OSFP (Octal SFP)-DCO form factor, to enable dense wavelength division multiplexing (DWDM) systems of up to 120 km [12].
Concerning post 400G-era solutions, enormous challenges for both the digital coherent and the IM/DD technologies are foreseen.Currently, research and development efforts are put in many aspects, including components, modulation formats and digital signal processing (DSP) techniques, to meet the speed, cost, power consumption and footprint requirements [13].On one hand, the digital coherent transceivers are mainly facing engineering challenges to simplify the high-performance versions, which are used for long-haul networks in order to adapt to short reach application requirements.On the other hand, conventional IM/DD technologies are facing fundamental challenges to meet the bitrate requirement.Unlike coherent technologies, which utilize all feasible dimensions of freedom of an optical carrier (amplitude, phase and polarization) to carry data, the IM/DD systems only use the amplitude dimension.Therefore, an underlying trade-off between the bitrate per lane and the lane counts becomes important to consider.Though utilizing more parallel data lanes increases the aggregated data rate linearly, there are always a limited number of lanes one can pack into an optical transceiver given the power and size constraints.Therefore, in the context of 400G (and beyond) client optics, many efforts are devoted on tackling the challenge to increase the data rate per lane.
Currently, broadband electrical and optoelectronic components are being designed and developed to facilitate high bandwidth modulations.Advanced modulation formats with powerful DSP and coding techniques are proposed and implemented to optimize the spectral efficiency and transmission performance.Combining these advanced techniques, many systemlevel demonstrations of over 100 Gbps per lane transmissions have been reported.It shows a promising progress towards maturity for industrial development, where 200 Gbps per lane is the nearest target.Besides the digital coherent and the traditional IM/DD solutions, there are many hybrid approaches, including Stokes-vector (SV)-DD receivers [14] and Kramers-Kronig (KK) receivers [15], which were recently proposed and studied extensively.These hybrid schemes can be categorized as advanced direct-detection solutions [16].These schemes effectively employ complex vector modulation and/or polarization division multiplexing in a self-coherent manner to improve the bitrate-distance product, and have shown potentials as a way to forward in certain application scenarios.These advanced DD schemes can be treated as a compromise between the digital coherent and IM/DD solutions, having a clear tradeoff between the system performance and complexity.As these self-coherent approaches do not strictly follow the conventional definition of IM/DD system, they are not elaborately studied in this review.Interested readers can refer to [14]- [18], where more details on the topics can be found.
In this paper, we extend our OFC contribution [19] to provide a more detailed overview and outlook of different aspects of current high-speed IM/DD technologies, which are of a great potential to conquer the 200 Gbps/lane milestone for the development of the next-generation client side optical transceivers.The rest of this paper is organized as follows: in Section II, we present a review of recent development of optoelectronic devices with a main focus on the broadband modulators to facilitate high-speed signal modulation.Section III describes the use of different modulation formats together with DSP algorithms for IM/DD system impairment mitigation.We also show our recent works on 200 Gbps/lane IM/DD transmissions as examples and discuss the system performance in that section.In Section IV, a summary of state-of-the-art IM/DD system demonstrations enabled by various key technologies are presented.Finally, we give our conclusions and future outlook in Section V.

II. BROADBAND OPTOELECTRONIC DEVICES FOR HIGH-SPEED MODULATION
For a long period of time, the components for electrical-tooptical and optical-to-electrical conversions, the optoelectronic modulators in particular, have been the bottleneck of end-to-end channel bandwidth in the fiber-optic communication links.This is mainly due to the fact that the design, fabrication and packaging process of optoelectronic components and devices to support broad bandwidth while keeping a low noise level are fundamentally challenging.It requires technology advances in different fields, including material process, design, fabrication and packaging.Recently, there has been a significant progress in the design and manufacturing of such broadband components, which greatly enhance the channel capacity of fiber-optic communication systems.In this section, we focus on four types of commonly used devices, i.e., vertical cavity surface emitting laser (VCSEL), direct modulated laser (DML), Mach-Zehnder modulator (MZM) and electro-absorption modulator integrated with a distributed feedback laser (EA-DFB).The advantages in these devices are among the key enablers for high-speed IM/DD transmissions to support future applications.
VCSEL: Most of today's commercial short-reach (<300 m) intra-datacenter optical links employ GaAs 850 nm multimode (MM) VCSELs combined with MMF.With MM VCSELs, over 30 GHz 3 dB modulation bandwidth was achieved with power consumption of less than hundred femtojoules per bit [20].Beyond 100 Gbps transmission over tens or hundreds of meters MMF were demonstrated using various modulation formats [21]- [26].The achievable data rate and distance of the VCSEL and MMF-based scheme are mainly limited by modal dispersion since different transverse modes travel at different propagation velocities in the MMF, resulting in severe intersymbol-interference (ISI) at the receiver side.The impact of the modal dispersion on the system performance is determined by the number of transverse modes emitted from the VCSEL source and the bandwidth-distance product of the optical fiber, and it can be reduced or eliminated with few-mode or single-mode operation [27].Therefore, to support emerging hyper-scale datacenters with optical interconnects of 500 m and longer, SMF should be deployed.Correspondingly, it is desirable to adopt single mode (SM) VCSEL, which may support spectrally efficient transmissions over a longer reach compared with the MM VCSEL-based technologies.However, a main drawback is that SM VCSELs often have limited output power and need more complex optical alignment.To rule out the lasing of high-order transvers modes and realize quasi/single mode lasing, one can shrink the aperture size via oxidation [28] or integrate a mode filter generated with a surface relief [29].Over 100 Gbps per lane transmissions with SM VCSELs operating in the telecom bands were recently reported [30]- [33].Moreover, within the upper limit of GaAs technology, VCSELs emitting at 980 nm, 1060 nm and 1110 nm have been explored with high modulation bandwidth and a negligible increase in fabrication complexity [34].
DML: As the modulated signal directly drives the laser bias current, DMLs normally emit high output power and is considered a more power-and cost-efficient solution than the external modulation solutions.Additionally, their compactness also facilitates integration with other devices.These merits make DMLs favorable for cost-sensitive datacenter and access networks.However, limited modulation bandwidth often appears as the constraint in extending its potential to provide high-speed data links.Lately, several novel techniques are reported for enhancing the modulation bandwidth of DMLs [35], including multiple quantum wells (MQWs) laser design [36], multi-section laser design [37], and injection locking [38].With the multi-section laser design, a state-of-the-art DML with 55 GHz modulation bandwidth was reported, which enables a single lane of 112 Gbps 4-level pulse amplitude modulation (PAM-4) transmission without any off-line equalization [39].Furthermore, by using advanced modulation formats combined with DSP, single channel 100 Gbps transmissions were demonstrated with commercial low-cost 10G-class DMLs [40].These results indicate the promising potential of DMLs for supporting beyond 200 Gbps/lane applications.Besides the bandwidth limitation, another well-known problem with DML-based system is the DML's inherent chirp effect that broadens the spectrum.Correspondingly, both optical [41], [42] and digital signal processing techniques [43] are proposed to tackle the chirp effect and make a full use of it to enhance transmission performance [44].
MZM: A commonly used external modulator type for IM/DD optical communications is the MZM, which achieves intensity modulation by combining two phase modulators with a Mach-Zehnder interferometer structure.In order to support high-speed transmissions with advanced modulation formats, there is a growing demand for high-performance and smallsize MZMs [45].Commercial lithium niobite (LiNbO 3 ) MZMs have been used to demonstrate 100 Gbps transmissions and beyond [46]- [48].However, these commercial components are normally packaged into large-size modules, which are expensive and power hungry, hindering their use for client-side optical interfaces such as the pluggable optical transceivers [49].Some recent works were reported on designing integrated nanophotonic LiNbO 3 MZMs with low voltage and high bandwidth (>100 GHz) [50], [51], and their mass production capabilities remain to be seen.Nowadays, indium phosphide (InP)-based MZMs can be fabricated at low cost and allow for monolithic integration with a small size.Recently, S. Lange et al. presented an InP-based DFB laser monolithically integrated with an MZM of 54 GHz bandwidth [52], [53], and Yamazaki et al. demonstrated an InP-based 80 GHz MZM with a capacitance-loaded traveling-wave electrode (CL-TWE) [54].Another attractive candidate is silicon photonics (SiP)-based MZMs, which can be fabricated using wafer-scale technology compatible with the semiconductor industry.Recently, SiP-based traveling-wave MZM (TW-MZM) [55], [56] and multi-electrode MZM (ME-MZM) [57] have been widely investigated.A detailed review of the development of silicon photonics-based modulator can be found in [58].Besides the InP and SiP-based MZMs that are already in industrial development stage, recent explorations of silicon-organic hybrid (SOH) MZMs have demonstrated promising properties to support modulation of high data rates [59], [60].
EA-DFB laser: Semiconductor lasers integrated with electroabsorption modulators (EAM) have been used in commercial transceivers for 10G and 25G applications.This type of laser has generally better performance in terms of modulation linearity, extinction ratio and bandwidth compared with the DML [61].On the other hand, compared with the external MZM, the EA-DFB laser usually has a smaller size, lower driving voltage at a potentially lower cost.Therefore, development of this type of laser should continue for future high-speed applications.Demonstrations of over 100 Gbps transmissions are reported using EAMs of around 25 GHz bandwidth [62], [63].An EA-DFB laser of >50 GHz bandwidth was designed and reported in [64].One state-of-the-art device of this type is a DFB laser monolithically integrated with a traveling-wave EAM (DFB-TWEAM) of >100 GHz bandwidth [65], with which several high-speed transmissions were demonstrated [66], [67].This device was designed by KTH Royal Institute of Technology, fabricated by KTH and Syntune, and packaged by u 2 t Photonics (currently II-VI/Finisar) [68].The absorber is based on the 12 strain compensated InGaAsP quantum wells/barriers (QWs) of around 9 nm thickness each.The gain section of the DFB is based on 7 QWs 7 nm thick grown by metalorganic vapor-phase epitaxy (MOVPE) coupled with butt-joint technique on n-doped InP substrate.Figs.2(a) and (b) show the P(I) and P(V) characteristics of the DFB-TWEAM [66].A picture of the packaged device is shown in Fig. 2(c).In Fig. 2(d), the small signal transfer response of the TWEAM is displayed, where 3 dB bandwidth beyond 100 GHz with less than 2 dB ripple in the passband can be observed [65].It is worth noting that this DFB-TWEAM device was used for a real-time transmission system demonstration of a 100 Gbps non-return to zero (NRZ) on-off-keying (OOK) signal without using any pre-or post-signal processing during the EU FP6 HECTO Project [69].In this demonstration, the modulated signal showed negligible distortion compared with the electrical driving signal, evincing both broad bandwidth and high linearity of the phase response of the DFB-TWEAM device.

III. MODULATION FORMATS, DSP AND PERFORMANCE
For IM/DD transceivers supporting up to 100 GbE traffic, the NRZ OOK modulation format has been employed with lane rate of up to 25 Gbps [6].For the forthcoming transition to the GbE, an upgrade from NRZ OOK to PAM-4 is adopted by the IEEE 400 GbE 802.3bs standard [70].For future applications beyond 400G, the options for the modulation format to support 100/200 Gbps per lane are still open for discussion.Among different options, the high-level PAM and discrete multitone (DMT) are the two main candidates, which attract much attention.Hence, they are selected for detailed discussions in the remaining of this section.We present our experimental investigations on using these two modulation formats to approach Gbps/lane transmissions [71]- [73].Besides the PAM and DMT, it is worth mentioning that other modulation formats, including carrier-less amplitude and phase (CAP) modulation [74], [75] and half-cycle subcarrier modulation (SCM) [76] (though not elaborately covered in this paper) have also been investigated and demonstrated experimentally with over 100 Gbps/lane data rates.

A. PAM
PAM is a modulation format that encodes binary data into multi-level signal pulses, and its simplest 2-level form is the NRZ OOK in the context of IM/DD communications.Employing PAM signal with a higher number of amplitude levels enables higher system spectral efficiency.On the other hand, higher-level PAM signal sets the stricter requirement for system signal-to-noise ratio (SNR).This is because for a N-level PAM signal (PAM-N, N > 2), its eye height is reduced by a factor of (N − 1) compared to that of NRZ OOK, given the same signal amplitude peak-to-peak [77], [78].To increase the system bitrate, one can either increase the signal bandwidth by using a higher symbol rate, or improve the spectral efficiency by using modulation formats with a higher order.The former approach requires an upgrade of the end-to-end system bandwidth, including all the bandwidth-limited optical and electrical components.This, in turn, requires an upgrade of design and fabrication technologies for devices and materials.The latter approach requires a system with a large end-to-end effective number of bits (ENoB), and often can be used to achieve a higher lane rate without replacing all the bandwidth-limiting components of a deployed IM/DD system.In addition, DSP algorithms and forward-error-correction (FEC) coding can be used to improve the system performance.
Feedforward equalizer (FFE) and decision-feedback equalizer (DFE) are two commonly used equalizer structures for PAM signals [79].FFE consists of a number of taps with impulse response determined by the tap weights.The tap weights can be adapted by several different algorithms, e.g., decision-directed least-mean-square (LMS) algorithm or recursive least squares (RLS) algorithm.In terms of implementation, one can use either symbol spaced or fractionally spaced FFE configuration, i.e., the FFE operates at 1 sample per symbol (SPS) or > 1 SPS.By minimizing the cost function defined for each algorithm, the FFE eventually converges to a state when the equalizer response represents the inverse of the channel frequency response.It is known that in a bandwidth-limited system the FFE boosts the high frequency signal components and minimizes the ISI.Nevertheless, the high-frequency noise can be enhanced, which may degrade the overall performance for the case of a limited modulation dynamic range.To address such drawback of FFE, DFE can be employed, which utilizes the post-decision symbols for cost function reduction.The implementation of DFE is often combined with the FFE by adding the decision feedback loop with symbol-spaced taps to suppress the high-frequency noise induced by the FFE, and to effectively compensate for both the pre-cursor and post-cursor ISI.Such a configuration shows superior equalization performance compared with that of solely using the FFE.However, it suffers from a decision delay and may also cause error propagation problem due to erroneous decisions.Therefore, when selecting an optimal equalizer type for a transceiver design, one needs to consider both the transceiver specifications and the aimed application scenario.
We have experimentally explored the potential of using FFE and DFE for 200 Gbps/lane IM/DD transmissions with PAM-4, and our experimental setup is shown in Fig. 3 [71].At the transmitter side, an electrical NRZ OOK signal of up to 100 Gbaud was generated by a multiplexer, and up to 100 Gbaud electrical PAM-4 signal was formed by combining two streams of decorrelated NRZ OOK signals.The DFB-TWEAM device reviewed in Section II was used for intensity modulation.The optical spectra of the DFB-TWEAM output for different modulations are shown in the inset of Fig. 3.At the receiver, the received signal is sampled at 200 GSa/s at the real-time digital storage oscilloscope (DSO), and then down sampled to 1 SPS after clock recovery.A combination of FFE and DFE is used to equalize the signal.Both the FFE and the DFE operate at 1 SPS in order to cover the sufficient memory length due to the pulse broadening induced by filtering and the fiber chromatic dispersion (CD).In Fig. 4, we show quantitative and qualitative measures for up to 100 Gbaud OOK and PAM-4 signals for back to back (b2b) and 400-m SMF transmission.The bit-error-rate (BER) results as a function of received power for different equalizer configurations are shown in Fig. 5(a).For the analysis purposes, hard-decision (HD)-FEC code with 7% and 20% overhead (OH) and soft decision FEC (SD-FEC) with 20% OH are considered  (pre-FEC BERs at 5 × 10 −3 [80], 1.1 × 10 −2 [81] and 2 × 10 −2 , respectively).The SD-FEC is considered due to poor electrical b2b signal quality suffered from the implementation penalty.The BER curves are obtained using a 9-feedforward (FF)-tap and 9-feedback (FB)-tap symbol-spaced DFE for the OOK, while 71-FF-tap and 15-FB-tap DFE are implemented for PAM-4.Fig. 5(b)-(d) show the eye diagrams and histograms of the OOK and PAM-4 signals with and without equalizations.In the case of the OOK, BER performance of below the 7% HD-FEC limit for both optical b2b and 400-meter transmissions was achieved.From the PAM-4 results, several messages can be extracted when compared with OOK: 1) a severer degradation due to high sensitivity requirements and poor electrical signal performance at the transmitter; 2) a significant increase of the equalizer tap number is needed to reduce the impact of the electrical components on the PAM-4 signals; and 3) a larger overhead of FEC is needed, in particular when increasing the baud rate.Improvement in the performance can be confidently expected with a better-quality electrical driving signal and more advanced DSP algorithms.For instance, the maximum likelihood sequence estimation (MLSE) function has been used to effectively enhance the performance of bandwidth-limited signals [82], [83], and partial-response PAM signals [84].Moreover, equalizers based on Volterra series can be used to compensate for system nonlinear impairments [85]- [87].Recently, machine learning (ML)-based equalization techniques, e.g., artificial neural network (ANN) [88]- [92] and support vector machine (SVM) [93], [94] are also proposed and investigated for short-reach IM/DD systems.These advanced equalizers have demonstrated improved performance compared with the conventional FFE/DFE equalizers.Interested readers can refer to [95] for a more detailed review of digital equalizers for PAM signals.

B. DMT
The DMT is a type of orthogonal frequency division multiplexing (OFDM) technique, where the input data sequence is encoded in parallel onto many subcarriers [96].It is the modulation format originally chosen for the first ITU-T asymmetric digital subscriber line (ADSL) standard [97] and later for the ITU-T very high speed digital subscriber line 2 (VDSL2) standard [98].Recently, DMT attracted attention for short reach IM/DD systems owing to its intrinsic flexibility to shape the frequency spectrum of the transmitted signal, which can be used to maximize the spectral efficiency through bandwidth-limited channels with bit-and power-loading schemes.For instance, over 10 Gbps real-time IM/DD DMT transmission was demonstrated over 25 km SMF with transmitter and receiver FPGAs operating only at 4 GS/s [99].Unlike single carrier modulation formats, DMT does not perform wideband pre-or post-equalization to flatten the received signal spectrum by suppressing the low-frequency components of the signal.Such an equalization approach sacrifices the overall channel capacity as the low-frequency regime of the modulated signal normally corresponds to a high SNR.Instead, the DMT can first estimate the channel response and calculate the in-band frequency-dependent SNR values with a probe signal through the channel, and then adaptively assign modulation orders and power levels respectively for each subcarrier [100].An effective and widely adopted bit-and power-loading solution is known as the Chow's algorithm [101].With the bitand power-loading, the subcarriers at low-frequency regime can benefit from the high SNR with assignment of higher modulation orders, while subcarriers on the high-frequency roll-off edge can still be used to carry data with lower modulation orders.In such a way, the overall channel bandwidth usage can be maximized.However, similar to other multi-carrier systems, the DMT has its drawbacks such as high peak-to-average power ratio (PAPR).Theoretically, the upper bound of PAPR in a DMT waveform is proportional to the number of subcarriers [102].This circumstance imposes a performance trade-off between spectral granularity of the DMT subcarriers and the required resolution of digital-to-analog converters (DAC) and analog-todigital converters (ADC).Additionally, DMT with a high PAPR are less tolerant to the relative intensity noise (RIN) compared with PAM [103].
In order to optimize the efficiency of the DAC and ADC with limited ENoB, the signal waveform is often deliberately clipped.It is also common to drive the optical modulators in the nonlinear region to guarantee a high modulation index, and, hence, to maximize the achievable system SNR.This is particularly important for high spectrally efficient signals when the SNR is a limiting factor.However, enhanced system nonlinearities occur when the transmitter-induced nonlinearities interplay with the fiber CD and the nonlinear square-law detection at the receiver.Therefore, nonlinear equalizers at the receiver can be used to mitigate such nonlinear distortions and improve the overall transmission performance, which was verified in our recent experiment with a high-speed C-band DMT transmission system [73].Fig. 5 shows a typical DSP routine of the DMT transmitter and receiver that we employed for our experiment.On the transmitter side, the bit-and power-loading technique was used in addition to conventional DSP blocks to encode the subcarriers and generate the DMT waveform.At the receiver, a time domain nonlinear equalizer (TD-NE) was used to mitigate system nonlinearities.To reduce complexity, a simplified nonlinear model was suggested which takes into consideration the 2nd-order and partially 3rd-order terms of the Volterra series model to mitigate the nonlinearity components with certain channel memories [104].The DSP structure of the employed TD-NE is shown in Fig. 6.The signal is first equalized by a frequency domain one-tap linear equalizer before feeding back to the TD-NE block.The input to the TD-NE is denoted as Y (N ) and its corresponding time domain sample is y(n).In the first iteration, Y (N ) is fed to a decision-feedback (DF) function followed by an inverse fast Fourier transform (IFFT) module to get the estimation of the transmitted temporal samples y(n), denoted as x(n).The nonlinear kernels are estimated by comparing y(n) with x(n).The estimation process is realized by a data-aided RLS algorithm.Upon its convergence, we can obtain the nonlinear kernels.The signal is then equalized by subtracting the reconstructed nonlinear noise components.After this iteration, the equalized signal y (n) is utilized as an input for the DF module for the next iteration.The performance is initially improved with a number of iterations until the improvement becomes saturated.In [73] it is observed that the improvement saturation occurs after the 3 rd iteration.
The experimental setup for our DMT transmission study is similar to the one shown in Fig. 3 except for the transmitter and fiber link configurations.At the transmitter, a 92 GSa/s arbitrary waveform generator (AWG) is used to generate the signal drive the DFB-TWEAM, and the fiber link consists of up to 1.6 km SMF.For each tested link distance, we set the received optical power to maximum (+7 dBm in all test cases), and then iteratively optimize the parameters of the bit-and power-loading algorithm and the corresponding parameters of the TD-NE to obtain a maximum achievable gross data rate with stable performance of a BER level below the 20% OH SD-FEC limit.In our experiments, the gross date rates after the bit-and power-loading are 215.4Gbit/s, 209 Gbit/s, and 200 Gbit/s for the three tested cases, respectively [73].Fig. 7(a) shows the measured BER curves of the received DMT signals of the maximum gross data rates with and without the TD-NE.Significant performance improvement can be observed after mitigating the nonlinear impairments compared with the cases without TD-NE shown in Fig. 7(a).A longer distance causes an obvious reduction on the maximum achievable data rate.
The achievable net data rates for different distances with and without bit-and power-loading are shown in Fig. 7(b).It can be seen that bit-power loading can significantly improve the net rate (increase by 30-40%).With the work shown in Fig. 7 as a benchmark, further performance improvement in the achievable transmission distance and data rate can be expected.In terms of the transmission distance, the SMF attenuation in the C-band is at a minimum, whereas the fiber CD induces the well-known small-signal transfer function, i.e., the power fading notches at certain frequencies for double sideband (DSB) modulated signals.This results in the end-to-end channel bandwidth limit, and consequently generates penalty.One straightforward approach is to shift the transmission window to the O-band where dispersion is minimized [40].Single sideband (SSB) or vestigial side-band (VSB) DMT configurations in this case can be used to overcome such limitations and is demonstrated to improve the transmission distance [105]- [107].With respect to improving achievable data rate, the bottleneck lies within the electrical signal source, as the 3 dB analog bandwidth of the DAC is limited, which is much smaller compared with the bandwidth of the electro-optic modulator (e.g., DFB-TWEAM in [65]).Therefore, an improved electrical signal source, e.g., a DAC with high resolution and/or broader analog bandwidth can potentially unlock the bandwidth bottleneck and improve the system capacity.To date, there have been a number of record demonstrations exceeding 200 Gbps/lane milestone with different key technologies from different aspects, which are summarized in the following section.

IV. RECENT ADVANCES FOR BEYOND 200 GBPS PER LANE IM/DD TRANSMISSIONS
There are technical challenges in almost every part of the IM/DD system when single lane bitrate goes beyond 200 Gbps.Meanwhile, we have witnessed tremendous research efforts and significant technical breakthroughs in high-speed DAC/ADC technologies, optoelectronic devices and advanced modulation, coding and DSP techniques, since only a few years back.Along with these advances, there have already been a number of system-level demonstrations reporting line rates of beyond 200 Gbps per lane IM/DD transmissions.In this section, we summarize the state-of-the-art IM/DD transmission works and review the key enabling technologies, aiming to provide an overall picture of the frontline in this research direction.
Table I summarizes and compares various IM/DD transmission demonstrations with line rates of 200 Gbps per lane and beyond.To the best of our knowledge, the first demonstration breaking this borderline was reported back in 2016, where Kanazawa et al. achieved a 214 Gbps PAM-4 transmission by using an O-band lumped-electrode electro-absorption modulator integrated with a distributed feedback laser (LE-EADFB) with a 3 dB bandwidth over 59 GHz [108], [109].With the same laser module, a 300 Gbps DMT transmission was reported by further extending the driving signal bandwidth with a digital-preprocessed analog multiplexed DAC (DP-AM-DAC) and an analog multiplexer (AMUX) [110], [111].In the latest achievement from the same group, a line rate of 400 Gbps DMT transmission was demonstrated, by using the 80-GHz MZM with CL-TWE, as mentioned in Section II, which was wire-bonded to the AMUX [54].In [112] and [113], an in-house fabricated selector power digital-to-analog converter (SP-DAC) was used to demonstrate up to 214 Gbps generation and 200 Gbps transmission of PAM-4 over 0.5 km SMF in the C-band, with a maximum likelihood sequence detection (MLSD) at the receiver.Besides the aforementioned works, the DFB-TWEAM reported in [65] was also employed for a 204 Gbaud OOK transmission, where two 2:1 InP DHBT multiplexing selector was used to generate the high-baud rate signal [114], [115].A maximum a posteriori (MAP) symbol detector with a look-up-table (LUT) at the receiver was used to detect the received symbols.On the receiver hardware, a 1-to-4 SiGe HBT BiCMOS ADC was reported and used for a 200 Gbps PAM-4 transmission over 40 km SMF with only 14 GHz of ADC bandwidth [116].In a very recent report, a capacity coupled SOH modulator with 3 dB bandwidth above 65 GHz was employed for an optical b2b 200 Gbps PAM-4 demonstration [117].It is worth noting that this work only has pre-compensation at the transmitter but no post-processing, indicating the end-to-end system with a broad bandwidth and high linearity.
Besides the novel design and development in the device domain, advanced modulation, coding and DSP schemes represent another way to further push forward the line rate even with limited system bandwidth.With respect to modulation formats, the Trellis coding modulation (TCM)-assisted DMT [118] and PAM [119]- [121] signals were used to optimize the Euclidean distances between bits or symbols, thus to improve the overall system performance.Besides, probabilistic-shaped (PS) PAM-8 signal was employed for a demonstration of up to 260 Gbps C-band transmission [122].In terms of novel equalization techniques, transmitter-side pre-equalization using the Tomlinson-Harashima precoding (THP) can effectively avoid the error propagation problem in the DFE, and it can be combined with the other post-equalization schemes to maximize channel efficiency [123].Novel post-compensation algorithms including nonlinear (NL)-MLSE [124], and adaptive notch filter (ANF) [125] were also proposed and employed to demonstrate beyond 200 Gbps IM/DD transmissions, showing a great potential to overcome the system bandwidth limitations.Interested readers can refer to the original papers listed in Table I for technical details.

V. CONCLUSIONS AND OUTLOOK
We made an overview the state-of-the-art technologies in devices, modulation formats, and DSP algorithms that can potentially enable 200 Gbps per lane IM/DD system development.Novel broadband electronics and optoelectronic devices can considerably relax the system bandwidth limitations.Meanwhile, advanced modulation formats, coding, and DSP schemes can improve the system efficiency and transmission performance, to further push forward the single lane rate with given system bandwidth.With such a significant progress during the past years, it is expected that novel technology candidates will converge into feasible solutions to fulfill the requirements for future high-speed client-side optics, where IM/DD transmissions are still dominating.On the other hand, the fast development of digital coherent technologies has been pushing towards shortreach scenarios and closing the gap of cost, power consumption, and a packaging size.Advanced direct detection systems with SV-DD or KK schemes are also pushing the limitations ahead to find their way for industrial development.Nevertheless, fundamental research in continuing driving the traditional IM/DD technologies to the higher speed will remain an important task in the field of fiber-optic communications, and such research efforts will also eventually benefit other advanced technology alternatives in the long-term perspective.

Fig. 1 .
Fig. 1.A typical core, metro and access network scenario, including the metroedge and intra-/inter-datacenter links, where the high-speed IM/DD links are required.CO: center office; RN: remote node; DCI: datacenter interconnects.

Fig. 4 .
Fig. 4. (a) BER curves for the OOK and PAM-4 signals for b2b and 400 meters, (b)-(c) eye diagram and histograms at optical b2b without and with DFE at 8 dBm, and (d)-(e) eye diagrams and histograms with DFE for electrical b2b (top) and optical b2b (bottom) at 8 dBm.

Fig. 6 .
Fig. 6.The structure of the time-domain nonlinear equalizer.

Fig. 7 .
Fig. 7. (a) BER performance of the DMT signals for b2b and up to 1.6-km transmissions with and without the time-domain nonlinear equalizer, (b) achieved net data rate as a function of transmission distance.

TABLE I SUMMARY
OF RECENT IM/DD ACHIEVEMENTS WITH LINE RATE OF 200 Gbps PER LANE AND ABOVE a Results are not included in the paper but presented at the conference.