1. Introduction

Datacenter (DC) traffic contributes a major portion of nowadays global Internet Protocol (IP) traffic due to the limitless spectrum of cloud-centric services and applications. Nearly 77% of the overall DC traffic stays inside the local network of the source DC accounting for the so-called intra-datacenter traffic [1]. It is produced from storage, production, development and authentication of DC data. Due to the incessant growth in DC sizes and capacity demand, single mode fiber (SMF-) based interconnects are being heavily deployed by DC operators over distances ranging from 500 m to 10 km. Meanwhile, switching still takes place in the electrical domain, hence many optical to electrical (O/E) and electrical to optical (E/O) conversion processes are required before and after electrical switches, respectively. These E/O and O/E conversions are performed by optical transceiver pluggables that are typically attached directly to the switch faceplate. Current commercial 100 G transceivers utilize multiple channels (wavelengths or fiber lanes) multiplexed together to realize the aggregate 100 G throughput per pluggable. Each one of the multiplexed channels carries binary On-Off keying (OOK) modulated data that is direct detected using a single photodetector [2, 3].

According to the Ethernet roadmap [4], 800 Gb/s and 1.6 Tb/s will be the future switch interface speeds after 2020. Hence, aggregate throughputs of optical pluggables need to be increased by 8- or 16-fold from the current commercially available speed of 100 Gb/s. For this goal, multiplexing more channels (wavelengths or lanes) will become obviously more challenging to scale due to the large number of lasers required which adds to the overall module cost and packaging complexity. In addition, the laser frequency stability requirements will become more stringent as more dense optical spectra will be required to multiplex more wavelengths.

Because of aforementioned drivers and present challenges, a plethora of transmission schemes have been proposed in the recent literature that bridge the gap between legacy On-Off keying (OOK)-based intensity modulation and direct detection (IM-DD) and spectrally efficient but rather expensive coherent detection by combining the advantages of both worlds [5–17]. These low-cost schemes aim at increasing the spectral efficiency by using an M-ary modulation alphabet instead of a binary alphabet and by modulating more dimensions of the field emitted from the CW laser (e.g. phase and/or polarization) while still avoiding classic expensive and power consuming coherent detection. These spectrally efficient transmission schemes enable using fewer channels to realize the future aggregate Ethernet speeds by achieving higher capacity per channel. On the other hand, coherent detection used in metro and long haul enables high spectral efficiency by giving access to all four dimensions on the optical field while providing superior performance to all aforementioned schemes. Nonetheless, it is still widely regarded as too expensive and power consuming for short reach intra-DC optics. Recently in [18], a DSP-free coherent system has been proposed and verified via simulation for intra-DC applications. However the proposed architecture, which uses a simplified analog carrier recovery (CR) with multiplier-free phase detectors, is only suitable for QPSK modulation. Also, the analog CR still requires analog mixers, XORs, voltage controlled oscillator (VCO), etc., which at the end, is estimated to consume ~4 W while delivering 200 Gb/s [18].

In this work, we propose a power efficient optical intra-DC ‘coherent-lite’ transceiver that requires neither high-speed DSP nor high resolution data converters. It also avoids the use of high-speed analog circuitry to perform carrier recovery. The proposed scheme is in principle suitable for any higher order QAM, e.g. 16QAM. We validate the proposed DSP-free coherent-lite transceiver in a proof-of-concept experiment achieving 50 Gbaud DP-16QAM yielding 400 Gb/s over 10 km below the 2.2 × 10⁻⁴ KP4 FEC threshold. To the best of our knowledge, this is the first experimental demonstration of a single channel 400 Gb/s transmission over 10 km at the KP4 FEC using the proposed coherent-lite scheme. In addition, we perform a detailed experimental parametric study of the coherent-lite system in which various system parameters are swept such as baud rate, reach, laser power and laser linewidth. This study confirms that the proposed transceiver can be implemented using low cost DFB lasers. Moreover, we experimentally compared the performance of the coherent-lite system to that of a coherent system employing the full receiver DSP stack to quantify the performance penalty resulting from the DSP removal. Finally, we estimate the savings achieved in power consumption by adopting the proposed coherent-lite scheme relative to: i) the classic DSP-based coherent system, ii) the recently proposed DSP-free coherent system in [18], and iii) the standard IM-DD system employing parallel eight or four WDM lanes.

The remainder of the manuscript is organized as follows. Section 2 explains the architecture and principles of the coherent-lite system. Section 3 depicts the experimental setup used to verify the viability of the proposed scheme followed by the experimental results. Section 4 is dedicated for the power consumption comparison. We finally conclude in section 5.

2. Principle of proposed coherent-lite system

The main idea of the proposed coherent-lite system is eliminating transmitter and receiver DSP conventionally required in a coherent transceiver. Hence, power consuming, high-speed, high resolution digital-to-analog and analog-to-digital converters (DACs and ADCs) are no longer required along with application specific integrated circuits (ASIC) that perform DSP. In the proposed coherent-lite system depicted in Fig. 1(a), legacy coherent DSP functions in Fig. 1(b) are either discarded, possibly resulting in performance penalty, or replaced by an optical component with an equivalent function. Henceforth, we explain the proposed coherent-lite system and how the impairments corrected by DSP in a classic coherent transceiver are dealt with in the proposed system.

 

Fig. 1 (a) Full-duplex transmission using the proposed coherent-lite transceivers (low-speed electrical signals represented by green lines, high-speed electrical signals by magenta lines, optical signals by black, red, blue lines), (b) Legacy coherent DSP that are omitted in coherent-lite (left: transmitter DSP, right: receiver DSP excluding fiber nonlinearity mitigation), (c) Schematics of the polarization controller in the signal path (top) and tone path (bottom) which can be integrated on a PIC comprising polarization splitter rotator (PSR), phase shifters and couplers.

Download Full Size | PDF

Figure 1(a) shows the architecture of the proposed transceivers at both ends of the link to realize full-duplex communication. Similar to a conventional coherent transceiver, it includes a continuous wave (CW) laser, dual-polarization in-phase and quadrature modulator (DP-IQM), and a DP coherent receiver (DP-CRx) realized from a 2 × 8 90° optical hybrid and four balanced photodiodes. The DP-IQM is driven by an electrical signal generator providing four high-speed RF signals depending on the modulation format employed. In principle, the electrical signal generator does not need to employ expensive power consuming DACs with large bit resolution (e.g. 6 or 8) since no DSP is applied on the transmitted waveforms. Producing four multilevel RF signals each with √M levels is possible by active RF combining of log2√M binary signals to generate DP M-QAM formats [19, 20]. Alternatively, M-QAM signals can be generated by a segmented DP-IQM with each segment driven by a binary RF signal forming optical DAC [21]. Since the transmitter DSP in Fig. 1(b) is omitted, all system components must have adequate RF bandwidth to avoid large penalties from uncompensated inter-symbol interference (ISI). Finally, in absence of Nyquist pulse shaping, we have no control on the pulse shape of the RF signals which will generally be non-bandlimited.

At the receiver, some DSP blocks in Fig. 1(b) are omitted namely, optical front-end correction and ISI post-compensation. Penalty due to residual ISI should be small provided sufficient component bandwidth is sustained. In addition, the absence of front-end correction imposes strict requirements on the coherent front-end specifications (e.g. skew, power and phase imbalance). Finally, the remaining blocks in Fig. 1(b) are indispensable and cannot be ignored; chromatic dispersion (CD) compensation, polarization demultiplexing, carrier recovery and clock and data recovery (CDR).

Firstly, CD can be avoided by operating the system in the O-band near the SMF zero dispersion wavelength. In fact, O-band is the standard operating window for most current commercial intra-datacenter 100GBASE-LR4 optical pluggables that operate over 10 km to avoid CD [2]. If C-band operation is favoured for the coherent-lite system due to the readily available library of C-band components of a coherent system (e.g. DP-IQM, 90° hybrid), optical CD compensators can replace the omitted DSP. Though optical CD compensators amenable for photonic integration are available using all-pass structures (e.g. rings [22]), their additional insertion loss is larger than the additional propagation loss (~1.5 dB) incurred after10 km in the O-band relative to the C-band [23]. Hence, the coherent-lite system is better suited for O-band operation. Moreover, researchers in [24] have already demonstrated that Silicon photonic coherent transceivers can be designed to operate in C- and O-bands.

Next, removal of the carrier recovery (CR) DSP is accounted for by using a self-homodyne approach similar to our scheme in [25]. The main idea is to make use of the full-duplex fiber that is readily deployed in intra-DC links (10 km and less) to transmit the modulated signal and a copy of the transmit laser on the fiber pair of a full-duplex fiber. This is achieved via the four circulators C1 to C4 and the two couplers in Fig. 1(a). For example, transceiver 1 uses a laser at λ₁₂ that is first split; one part is modulated then passed by C1 and transmitted along the red fiber, while the other portion is passed unmodulated along the blue fiber by C2. At transceiver 2, C3 and C4 pass the received signal and tone to the coherent front-end. Likewise, communication in opposite direction at λ₂₁ is allowed by the four circulators and second coupler. Bidirectional transmission results in no penalty if λ₁₂ and λ₂₁ are sufficiently spaced to avoid nonlinear back scattering effects due to Brillouin scattering [26]. If the signal and tone paths are almost matched, near perfect homodyne detection is achieved after beating in the coherent front-end and therefore DSP carrier recovery can be omitted without having to replace it with an analog phase locking. Although the above self-homodyne scheme results in CW tone power lost in propagation compared to if a local oscillator laser is used, power consumption savings by the removal of converters and ASICs are more significant which will be quantified in detail in the forthcoming section.

Next, the polarization demultiplexing DSP is replaced by optical polarization controllers (PCs) in both the tone and signal paths driven by low speed polarization tracking circuitry (<1MHz). Since photonic integrated electrically controllable PCs have been realized by couplers and phase shifters [27], inserting them before a coherent front-end will achieve the necessary polarization demultiplexing. The top schematic of Fig. 1(c) shows a possible realization of the polarization controller that is implementable on a photonic integrated circuit [27, 28] which can be used in the received signal path prior to entering the hybrid. It comprises a polarization splitter/rotator (PSR) that accepts an incoming DP signal with an arbitrary state of polarization, splits it into orthogonal modes, and rotates one by 90° yielding two TE polarized modes that are supported by subsequent integrated waveguides. These two polarization components are passed through three phase shifting sections sandwiching two 50/50 couplers. By controlling the three phase shifters any complex unitary polarization rotation can be induced to the incoming DP signal. It can in principle invert any random polarization rotation incurred along transmission since PMD is negligible over short-reach transmission. To control the PC in the signal path, a possible polarization tracking algorithm based on marker (pilot) tone detection was presented in [18] which requires transmitting a low frequency (~50 kHz) tone multiplexed with the transmitted signal on one quadrature on one polarization tributary (e.g. in-phase component on X polarization, XI). Since the pilot tone is continuously transmitted, there is no need to retrieve the clock or symbol timing to perform the polarization tracking algorithm. Based on the detection of the received tone, a low speed circuitry can control the phase shifters of the signal PC such that it maintains the received pilot power maximized in the received XI quadrature while being minimum in all other three quadratures (XQ, YI, and YQ). On the other hand, the PC in the tone path can be simpler than the one in the signal path because the received tone is a single polarization CW field with an arbitrary state of polarization unlike the signal field which is dual polarized. A possible implementation of the PC in the tone path is shown in the bottom schematic of Fig. 1(c); it can be made of a PSR followed by one stage phase shifter and a 3 dB coupler. This one stage phase shifter is controlled depending on the state of polarization of the received tone to simply produce equal tone powers at the two outputs of the 3 dB coupler. The two outputs of the coupler are then fed to the two polarization sections of the DP hybridwhere each tone beats with the corresponding polarization tributary of the signal. Hence the tone polarization tracking algorithm is simply designed to control a single phase shift to continuously try to split the tone power equally between the X and Y sections of the DP hybrid. For both the tone and signal paths, a polarization independent coupling scheme from each fiber into the receiver chip is required before the arbitrarily polarized signal is fed to the PSR. Polarization independent low loss edge couplers have been demonstrated for coupling from a standard single mode fiber into integrated silicon-on-insulator SOI waveguides [29]. Finally, a CDR circuitry accepts the outputs from balanced photodiodes, extracts the clock and performs data decisions. Hard decision FEC decoding, that is not shown in Fig. 1(a), operates on the hard values following decisions.

3. Experimental setup and results

3.1. Experimental setup

Figure 2(a) shows the experimental setup in which we demonstrate unidirectional transmission, i.e. one transmitter and receiver of the full coherent-lite system in Fig. 1(a) were realized. Due to unavailability of certain O-band parts, the coherent-lite system was verified in the C-band. CD was compensated by an optical tunable dispersion compensator module (TDCM) based on tunable chirped fiber Bragg grating to demonstrate a DSP free system and was later compared in Fig. 3 to CD compensation in DSP. The TDCM has an optical bandwidth of ~80 GHz when set to compensate for 170 ps/nm of dispersion (equivalent to 10 km of SMF) and its insertion loss is around 5 dB. Also, instead of using a full duplex fiber we used a pair of similar fiber spools with the desired lengths (2 or 10 km). An optical delay line was inserted in the signal path to ensure near matching with the tone path similar to an actual full-duplex fiber. We used a 15.5 dBm external cavity laser (ECL) with linewidth <100 kHz in all measurements except for the last part where we also used two different distributed feedback (DFB) lasers to examine the linewidth tolerance of the system in Fig. 3. The DP-IQM is a commercially available InP modulator with 3dB bandwidth of 35 GHz. The electrical signal generator comprised four 43 GHz RF amplifiers which generated 4-level RF signals yielding DP-16QAM. The 50 Gbaud 4-level electrical eyediagram is shown in the top left inset of Fig. 2(a). At the receiver, four 40 GHz balanced PDs follow the DP hybrid. Due to unavailability of balanced PDs with TIAs, we used an EDFA in the signal path prior to the CRx however the tone path was unamplified. Manually controlled PCs were used in both signal and tone paths because the polarization state was stable in our lab environment hence polarization tracking was not necessary for a proof-of-concept experiment. Two 63 GHz real-time scopes (RTOs) sample the outputs of balanced photodiodes at 160 GSa/s for offline error counting. Resampling and timing recovery were performed offline on the sampled waveforms before error counting. In a real time coherent-lite receiver, timing recovery is performed by a CDR circuit. In addition to timing recovery, a slowly varying phase rotation angle had to be estimated from the sampled waveforms, and the I and Q waveforms per polarization had to be derotated offline using the estimated angle. This inevitable slowly varying rotation is due to any small path length mismatches between the signal and tone paths which will vary slowly in time due to ambient temperature and pressure changes around the two fibers spools. Also any sight fluctuations in the propagation constants of the two optical paths can lead to a slowly varying phase rotation after beating the signal with the tone even if the paths have exactly equal lengths. However, this phase rotation is slowly varying; it is found to be constant during each entire captured waveform which lasts ~12.5 μs (each trace captured at 160 GS/s contains 2 million points) but varies from one capture to the next. The entire 50 Gbaud 16QAM waveform from one full captured trace is depicted in the constellations of Fig. 2(b) before and after common phase derotation. Other than the common phase derotation, no DSP is applied at the receiver to obtain these constellations after 2km. In a real time coherent-lite receiver, the slowly varying phase derotation can be performed by the electrically controlled PCs depicted in Fig. 1(c) since the two phase shifters in the last stage of the signal PC can introduce slowly varying independent phase rotations on the X and Y polarizations which can counteract the slowly varying phase rotations from the channel. This is only possible because the time varying phase is slowly changing (found almost constant over 12.5 μs traces) and hence thermal phase shifters, such as the design in [30] with 3-dB modulation bandwidth > 100 kHz, should in principle be capable of producing time varying phases with enough speed to counteract the slowly varying rotation from the channel. The real-time received raw unequalized 4-level eyediagram of one quadrature obtained on the RTO by the Keysight PAM4 analysis tool is also shown in Fig. 2(c) which depicts fairly open eyes without DSP. The slow phase derotation was performed on the RTO prior to plotting the real-time eyes. These raw open eyes will be confirmed by BER results in the following subsection.

 

Fig. 2 (a) Experimental setup, (b) Raw constellations of the 50 Gbaud 16QAM signal after only timing recovery to find the optimum sampling location. The constellations show the entire length of the captured frame from the RTO lasting 12.5 μs before and after the common phase de-rotation, (c) Raw unequalized 50 Gbaud 4-level eyediagram of one received quadrature from Keysight PAM-4 analysis tool.

Download Full Size | PDF

 

Fig. 3 Experimental results showing (a) BER versus baud rate for DP-16QAM in back-to-back (B2B) for DP and SP transmission, (b) BER versus baud rate for DP-16QAM after 2 and 10 km comparing DSP-free, DSP CD compensation and full DSP performance, (c) BER of 50 Gbaud 16QAM in B2B versus the laser power (note that decreasing laser power decreases both signal and tone powers), (d) BER of 50 Gbaud 16QAM in B2B versus mismatch between tone and signal paths for three different lasers.

Download Full Size | PDF

3.2. Experimental results

Figure 3(a) shows the bit error rate (BER) versus baud rate in back-to-back for 16QAM in two scenarios: single and dual polarization (SP and DP) transmission where the BER is counted per polarization. This figure shows no penalty arising from DP transmission compared to SP, i.e. polarization demultiplexing is properly achieved optically. For later results, average BER of the two polarizations is reported for the DP case. Next, Fig. 3(b) compares the DSP-free performance obtained with the TDCM at 2 and 10 km with two cases in which TDCM was removed: i) only CD is compensated digitally in DSP, ii) the full DSP stack of Fig. 1(b) is applied including ISI compensation. No significant difference between compensating CD optically or digitally is noticed. Also, results confirm DSP-free 400 Gb/s 10 km transmission below the KP4 FEC threshold of 2.2 × 10⁻⁴. Furthermore, compensating residual ISI by leveraging the full DSP stack improves the BER by an order of magnitude and pushes the capacity that can be achieved below KP4 FEC to 512 Gb/s (64 Gbaud) at the expense of added DSP power consumption that will be quantified in the next subsection. In Fig. 3(c), we study the BER of 50 Gbaud 16QAM in B2B with decreasing laser power which simultaneously decreases signal and LO powers. We notice the system still operates at the KP4 FEC when the laser power is lowered to 12 dBm (by 3.5 dB). Finally, Fig. 3(d) shows the effect of mismatch between tone and signal paths on the BER of 50 Gbaud 16QAM in B2B. Three 15.5 dBm lasers with different linewidths were compared. With the narrow linewidth (< 100 kHz) ECL, the system tolerates large mismatches up to 6 m. With 350 kHz DFB laser, the system still tolerates 25 cm of mismatch without any DSP. The mismatch tolerance of 25 cm is more than sufficient for an actual full-duplex fiber where both the signal and tone fiber pair are in one cable.

4. Power consumption analysis and comparison to other systems

This section is devoted to analyzing the power consumption of the DSP-free coherent-lite scheme and comparing it with two solutions being discussed in the literature. The section is divided into two subsections; the first subsection quantifies the power consumption savings achieved by the coherent-lite approach relative to a conventional DSP-based coherent system, and the second subsection compares the coherent-lite scheme with the standardized intensity modulation and direct detection (IM-DD) scheme utilizing WDM multiplexed wavelengths (e.g. eight or four wavelengths) while making the aggregate capacity delivered by both systems equal.

4.1 Comparison with classic DSP-based coherent system

In this subsection, we quantify the power consumption savings that can be achieved by using the coherent-lite approach relative to a classic DSP-based coherent system. In addition, we highlight the savings our system realizes compared to the recently proposed DSP-free system in [18] that utilizes an analog phase de-rotation stage and a multiplier-free phase estimator to replace the DSP based carrier recovery. It is noteworthy here that the multiplier-free phase estimator in [18] works only for QPSK signals, and a traditional Costas loop must be used for higher order QAMs such as 16QAM. We restrict our power consumption comparison to the receiver side since a classic DSP-based coherent system can, in principle, also operate without DSP and DACs at the transmitter side at the expense of some performance penalty. Moreover, we disregard any power consumption analysis of the FEC encoding/decoding as well as timing recovery since they are common to all systems.

We begin by estimating the power consumption for the classic DSP-based coherent receiver. We take into account the power consumed by ADCs and the following DSP subsystems: CD compensation, 2 × 2 MIMO polarization demultiplexing and carrier recovery. Essentially, these are the subsystems which the proposed coherent-lite approach aims at removing. We use the model in [31] with the parameters listed in Table 1 to evaluate the power consumption of these individual subsystems. Any other parameter in the model of [31] not listed in Table 1 is assumed to take the same value as in [31]. We assume CD compensation is implemented in the frequency domain via an overlap-save method. Assuming a symbol rate of 50 Gbaud and an oversampling factor of 2, the number of taps required for CD compensation after 10 km of SMF is 13 [32], and hence we assume a 64-point FFT/IFFT implementation of the overlap-save frequency domain equalizer. Since there is negligible PMD after 10 km, the number of taps of the time domain adaptive MIMO is chosen to be only 5. Using these parameters plugged into the model in [31], it basically calculates the number of basic operations (e.g. multipliers, adders, register read / write) for each subsystem, and multiplies each number by the power required to implement each individual operation in a specific target CMOS node. The outcome of the model is as follows: the combination of ADCs, CD compensation, MIMO and CR subsystems consume about 18.5 W assuming 28 CMOS node for implementation. This estimate goes down to around 9.25 or 4.5 W in smaller 14 or 7 nm CMOS nodes.

Table 1. Parameters used for power consumption of ADCs and DSP in a classic 400 Gb/s DP-16QAM coherent system.

View Table | View all tables in this article

Next, we turn our attention to the proposed coherent-lite system. We now quantify the additional power consumption incurred in the coherent-lite scheme to enable removing the ADCs, CD compensation, MIMO and CR subsystems. Since a classic coherent system uses a local oscillator laser, the additional optical laser power in the coherent-lite scheme that is required to account for the transmission of a copy of the transmit laser is equal to the fiber propagation loss in addition to the insertion losses (ILs) of the two circulators and the coupler. In addition, the removal of DSP CD compensation is enabled by operation in the O-band which introduces a higher attenuation parameter in SMF. Assuming an attenuation parameter of 0.35 dB/km in the O-band [23] leading to propagation loss of 3.5 dB over 10 km, and ILs of 0.5 and 1 dB for the coupler and circulator, respectively, an additional 6 dB of optical laser power is required in the coherent-lite transceiver to maintain the same LO power of an equivalent classic coherent system. Assuming the laser of a classic coherent transceiver has 13 dBm of optical power that is split into transmit and LO portions, a threshold current of 20 mA and slope efficiency of 0.3 mW/mA [33], the laser drive current needs to be increased from 86 to 286 mA. If we assume a laser forward voltage of 2.5 V [33], this translates into 0.5W of additional electric power required to drive the laser emitting at higher optical power. Finally, the polarization demultiplexing in the coherent-lite system as well as the removal of the slowly varying phase rotation is achieved via the signal and tone polarization controllers in Fig. 1(c). According to [30], each thermo-optic phase shifter fabricated on a Silicon on insulator (SOI) standard platform consumes about 25 mW resulting in 200 mW of power consumption for the signal and tone PCs. If we add to this the power consumption of the low speed circuitry that executes the polarization tracking and controls the phase shifters, we can claim conservatively that the total power consumption due to additional laser power, polarization and phase control circuit in the coherent-lite scheme is ~1 W.

Hence, we conclude this analysis that in the coherent-lite system, the power consumption of the ADCs, CD compensation, MIMO and CR subsystems, which was estimated to be ~18.5 W assuming 28 nm CMOS, is replaced by only ~1 W consumed by the more powerful laser and the actively controlled polarization / phase controller. This represents 95% relative saving in the power consumption compared to a classic DSP-based coherent implemented in 28 nm CMOS. In case of 14 and 7 nm CMOS, the power consumption saving decreases to 89 and 78%, respectively. This enormous power consumption reduction comes at the price of performance degradation due to the limited component bandwidth leading to residual uncompensated ISI, which is quantified in Fig. 3(b) of our experimental results (one order of magnitude in BER degradation compared to a full DSP-based system). Nonetheless, the performance of the coherent-lite system was adequate to achieve 400 Gb/s over 10 km with BER below 2.2 × 10⁻⁴.

Finally, we also highlight the savings achieved in our proposed coherent-lite system compared to the recently proposed DSP-free coherent system in [18]. The system in [18] already saves some power compared to a classic DSP-based coherent due to the removal of ADCs and DSP. However, DSP-based carrier recovery is replaced by an analog circuitry which requires the use of analog mixers, multipliers, VCO etc. According to [18], their proposed analog CR which uses a multiplier free phase detection consumes ~4 W. Also, this estimate is expected to increase when 16QAM is used since the multiplier-free phase detector works only for QPSK and a conventional Costas loop needs to be used for higher order modulations. Hence, it becomes readily apparent that our coherent-lite system achieves major power consumption saving relative to the system in [18].

4.2 Comparison with wavelength-division-multiplexed based direct detection systems

In this subsection, we compare our proposed DSP-free coherent-lite system with an IM-DD system utilizing parallel wavelength-division-multiplexed (WDM) channels to achieve the same target aggregate bit rate. We begin the comparison by listing the building components of each system as well as their respective count. Then we pinpoint the major differences in power consumption between both systems in order to assess where the proposed coherent-lite system stands relative to the IM-DD-WDM systems in terms of power consumption. We finally conclude and provide a qualitative discussion highlighting other merits of the coherent-lite system relative to the IM-DD-WDM system in terms of packaging complexity and scalability to future aggregate capacities.

Throughout the discussion, we fix the following system parameters for all systems for fair comparison: the aggregate bit rate is fixed at 400 Gb/s, the reach is assumed to be 10 km over SMF, and the operating wavelength window is in the O band. To achieve the target aggregate bit rate of 400 Gb/s over the intended 10 km reach, the coherent-lite scheme requires a single optical carrier generated by one laser source, whereas the IM-DD-WDM typically requires eight optical carriers, each carrying a net data rate of 50 Gb/s using 25 Gbaud PAM4, that are WDM multiplexed in the O band. The 8 × 50 Gb/s IM-DD-WDM system is the standard solution as defined by the 400GBASE-LR8 physical medium dependent (PMD) sublayer specifications in the recently approved IEEE standard for 400G Ethernet [34]. It uses eight closely spaced (800 GHz channel spacing) optical carriers in the O band. Since using eight carriers in 400GBASE-LR8 is challenging from cost, form factor and power consumption perspectives, there exists non-IEEE multi source agreements (MSAs) such as the 100G Lambda MSA [35] whose initial release has already moved to 100 Gb/s per wavelength over 10 km SMF (100 GBASE-LR) with a pathway to transmitting 400 Gb/s on four multiplexed wavelengths over 10 km in future phases of the MSA. Hence, we also opted to include a 4 lambda IM-DD-WDM system in our comparison of 400 Gb/s solutions. This 4 × 100 Gb/s IM-DD-WDM system operates at the same baud rate as the coherent-lite system, i.e. 50 Gbaud, and hence uses the same class of 50 G electronics.

In terms of component requirement for each transmission direction over the duplex fiber, Table 2 compares the coherent-lite system with the two IM-DD-WDM variants using either 8 or 4 wavelengths. Depending on the number of required lanes to achieve the target aggregate bit rate, the three systems require either one, eight or four distributed feedback (DFB) lasers centered at the designated wavelengths across the O band. For the IM-DD-WDM systems, thermo-electric coolers (TECs) are typically required because of the relatively densely spaced wavelengths (800 GHz spacing or ~4.5 nm) whereas it may be omitted for the single carrier coherent-lite system because both the signal and tone will drift similarly on the duplex fiber pair preserving homodyne detection. The coherent-lite system requires four RF drivers connected to the four inputs of a single DP-IQM whereas the two IM-DD-WDM systems need either eight or four RF amplifiers driving either electro-absorption modulators (EAMs) or Mach-Zhender modulators (MZMs). In the table, we also list the bandwidth and output swing requirements of the RF drivers in each system. For the coherent-lite system, the RF driver is required to produce an output swing close to (but less than) twice the half-wave voltage (~2V_π) of the DP-IQM because the child MZMs inside the DP-IQM are biased at the null transmission point of the field transfer characteristics. InP DP-IQMs, including the unit we used in our experiment, customarily have V_π close to 2 V meaning RF drivers need to produce peak-to-peak swings in the order of 4 V. On the other hand, IM-DD-WDM systems using external MZMs need RF drivers with output swings roughly half that of a coherent system, i.e. ~2V, since MZMs are biased at the quadrature transmission point to be used for intensity modulation. Also, more commonly used IM-DD systems employing EAMs have peak-to-peak swing requirements around 2 V from the RF drivers. In addition, Table 2 lists the passive optics required in each of the three systems where the IM-DD-WDM systems require WDM multiplexer and demultiplexer as opposed to the 90° hybrid and circulators required in the coherent-lite system. Finally, the coherent-lite system needs four balanced photodetectors (BPDs) followed by TIAs whereas the IM-DD systems require either eight or four single-ended PDs (SE-PDs) equipped with TIAs.

Table 2. Comparison between various systems in terms of component requirements per transmission direction on the duplex fiber; all systems deliver 400 Gb/s aggregate capacity and have 10 km target reach over SMF.

View Table | View all tables in this article

Next, we move on to the power consumption comparison between the three systems. Although the IM-DD-WDM systems typically use some DSP at the receiver for equalization, we will ignore the power consumption of this DSP subsystem since there is no particular equalizer implementation that is standardized (e.g. the IEEE 400G Ethernet standard specifies a 5-tap T-spaced FFE as the reference equalizer for the TDECQ measurement but clearly states there is no particular equalizer length or design for normal operation [34]). We will compare the power consumption of the three systems taking into account the building components listed in Table 3 along with their power consumption values per unit. These values, taken based on [36], are not necessarily exact and may vary from one vendor to another but they will give us estimates precise enough to help us conclude where the three systems stand relative to each other in terms of module power consumption. Each CW laser is assumed to consume 0.3 W to emit 30 mW (~15 dBm) of CW optical power assuming threshold current of 20 mA, slope efficiency of 0.3 mW/mA and forward voltage of 2.5 V [33]. Without getting into the optical link budget details of the three systems, this CW power should be sufficient for the three systems under study. For the coherent-lite system, this is confirmed by our results in Fig. 3(c) if we assume that the EDFA we used for our proof-of-concept demonstration can be replaced with TIAs following the BPDs without compromising the optical link budget. For IM-DD systems which usually employ EAMs, DFB lasers are integrated with EAMs and hence CW power is typically defined at the output of the EAM under no bias (in transparency) and not after the laser itself. Nonetheless, assuming a power consumption of ~0.3 W for the DFB is also a reasonable value for DFB-EAM (or EML) modules (e.g. the datasheet in [37] specifies laser driving current of 150 mA and forward voltage of 1.8 V to produce 1 dBm average modulated optical power). Considering the 6.3 dB optical channel insertion loss defined in the IEEE 400GBASE-LR8 standard, a launch power of 1 dBm per lane is sufficient to provide received power within the tolerable range of average received optical power defined in the 400GBASE-LR8 receiver specifications [34]. Next, the power consumption of the modulator drivers depends on the bandwidth class and the output swing requirements. For the 25G class drivers, 28 nm CMOS is assumed whereas 16 nm CMOS drivers are assumed for 50G operation and all values are taken similar to [36]. We also assume that doubling the output swing requirement should double the power consumption of the driver (see Eq. (4) in [31]). Next, the TIA power consumption is also taken from [36] for both 25G and 50G classes. Finally, the outcome of this comparison is as follows: the coherent-lite, 4 × 100G IM-DD, and 8 × 50G IM-DD systems consume 3.75, 4.4, and 7.2 W, respectively taking into account the power consumption of lasers, TECs, modulator drivers, and TIAs.

Table 3. Power consumption of individual components.

View Table | View all tables in this article

As a conclusion to this section, we infer that the DSP-free coherent-lite system achieved major power consumption savings relative to a classic DSP-based coherent system. These major savings discussed in subsection 4.1 brought the power consumption of the coherent-lite system very close and even slightly lower than that of a classic IM-DD system with four parallel wavelength lanes according to our findings in subsection 4.2. In other words, the coherent-lite system has the potential to bring the well-known benefits of coherent (spectral efficiency and improved sensitivity) into the intra-datacenter world while maintaining roughly the module power consumption of a corresponding 4-lambda IM-DD solution using the same class of electronics. In addition, the coherent-lite system requires fewer light sources which translates to simpler thermal budget, reduced packaging cost, smaller form-factor packages leading to lower overall module cost. Furthermore, the improved spectral efficiency of the coherent-lite system and the fewer number of light sources makes it better scalable to future aggregate bit rates of 800 G and 1.6 T by only using 2 or 4 wavelengths if we assume nowadays class of 50 G electronics.

5. Conclusion

We propose a DSP-free coherent-lite transceiver that circumvents high-speed, high-resolution DACs and ADCs for 10 km intra-DC optics. The proposed transceiver architecture can be fully integrated on chip and can exploit low cost DFB lasers with linewidths of few hundreds of kHz. The coherent-lite scheme makes use of the readily available full duplex SMFs to transmit the signal and a copy of the transmit laser tone on a fiber pair to enable coherent detection at the receiver. Hence it does not require high-speed analog carrier recovery to replace its DSP counterpart. Our proof-of-concept experimental result confirms the viability of the coherent-lite scheme while demonstrating 400 Gb/s over 10 km with BER below 2.2 × 10⁻⁴. This is potentially scalable to 1.6 Tb/s aggregate capacity by multiplexing four wavelengths or fiber lanes. We finally estimate the savings achieved by the proposed coherent-lite approach compared to both a standard DSP-based coherent system and IM-DD systems employing parallel WDM lanes. Our calculations estimate that the power consumption of the coherent-lite system is much less than a DSP-based coherent system and is very close and even slightly lower than a standard IM-DD system with four parallel wavelength lanes using the same class of electronics. The DSP-free coherent-lite is a potentially attractive solution for 10km intra-datacenter interconnects due to superior spectral efficiency compared to direct-detection schemes and hence fewer lasers to be packaged per transceiver module making it a better scalable solution to future 800G and 1.6T speeds.