Demonstration of a 17 × 25 Gb/s Heterogeneous III-V/Si DWDM Transmitter Based on (De-) Interleaved Quantum Dot Optical Frequency Combs

We discuss the design and demonstration of a space and dense wavelength division multiplexed heterogeneous III-V/Si transmitter based on a single multi-wavelength quantum dot laser source and ultra-power-efficient metal-oxide-semiconductor capacitor (MOSCAP) (de-)interleaver. This paper begins by introducing a transceiver architecture capable of > 1 Tb/s transmission with < 1.5 pJ/bit power consumption, followed by a detailed discussion of the heterogeneous laser source and (de-)interleaver. The O-band quantum dot laser, based on a compound cavity design, has a FSR ∼ 64 GHz with a 1σ variation of ∼ 1.08 GHz and a measured relative intensity noise (RIN) of ∼ −144 dB/Hz for the largest comb peak. The single-ring-assisted asymmetric Mach-Zehnder interferometer (1-RAMZI) MOSCAP (de-)interleaver exhibit cross-talk (XT) levels down to −27 dB for tuning powers of 10.0 nW. Finally, to the best of our knowledge, we have demonstrated for the first time, a simultaneous wavelength and space division multiplexed transmitter fabricated on a heterogeneous III-V-on-silicon platform. Experiments show (de-)interleaved 17 optical comb lines, each modulated at 25 Gb/s non-return-to-zero (NRZ) for an aggregate bandwidth of 425 Gb/s.


I. INTRODUCTION
T HERE continues to be an increased demand for high bandwidth density optical interconnects for growing mega data centers, long-haul telecommunications, and peta/exa-scale high-performance computing. A study by Cisco suggests 66% of the global population will have internet access by 2023 and the number of devices connected to IP networks will be 3× the total population [1]. In addition, machine-to-machine (M2M) connections will grow by 50% with the arrival of edge computing, machine learning, artificial intelligence, etc. [1]- [3]. Manuscript  To satisfy this exponential growth in data traffic, the trend has been to scale mega-data centers with hundreds of thousands of servers, which underlies the concern for data center power consumption given increased ecological concerns and future environmental impact [2], [4]. It is estimated by 2030 that 8 -21% of global power consumption will be directly attributed to data centers [5]. As a result, silicon photonics has catapulted to the fore-front as a vanguard technology that promises to reduces system-level power consumption (sub-picojoule/bit), increase aggregate bandwidth (Tb/s), and lower manufacturing costs by leveraging an already well-established complementary-metaloxide-semiconductor (CMOS) technology industry. To address the aforementioned issues, at Hewlett Packard Labs, we are developing a novel heterogeneous III-V/Si dense wavelength division multiplexing (DWDM) architecture to address chip power consumption (< 1.5 pJ/bit) and increased transmission bandwidth (> 1 Tb/s) [6]- [8]. The heterogeneous platform described in this work has shown the technical capability and fabrication compatibility to integrate all building blocks, such as heterogeneous quantum dot (QD) optical frequency comb (OFC) laser sources [6], [9]- [14], wavelength (de-)interleavers [15], [16], micro-ring modulators (MRRs) [8], [17]- [19], photodetectors (PDs) [20]- [23], and semiconductor optical amplifiers (SOAs) [24] to form a space division multiplexing (SDM)-DWDM transceiver as shown in Fig. 1 [6], [7]. Within this architecture, there is a practical limit to the number of MRRs that can be cascaded onto a single bus waveguide due to crosstalk (XT), off resonance insertion loss (IL), and free spectral range (FSR) limitation due to finite bend radius, [24]- [27]. On-chip wavelength (de-)interleavers solve this problem by spatially dividing even and odd numbered OFC frequencies onto separate waveguides [16], [27], [28]. For our architecture, each waveguide will have a half number (10) of the total MRR count (20), but with the channel spacing doubled. By cascading more low-loss (de-)interleavers in series, more spatial channels and wider channel spacing is possible, allowing to use a comb source with tens or hundreds of comb lines but with smaller channel spacing [8]. For demonstration purpose, the channel spacing is chosen to be 65 GHz due to moderate power penalty for 10 Gb/s transmission in DWDM applications [29]. It has also been shown that optimal energy-cost [fJ/bit] lies in low data rate per channel where serializer-deserializer (SERDES) power consumption is lower [30], [31]. In this paper, we demonstrate preliminary link performance based on discrete components (QD OFC + (de-)interleaver) and then modulating each (de-)interleaved comb line. The QD OFC [6], [9]- [14] and MOSCAP (de-)interleaver [15], [16] are constructed on the same platform and can be seamlessly integrated in the future along with energy efficient MOSCAP MRRs [8], [17]- [19]. The paper will first discuss individual component design and performance of the heterogeneous OFC laser source and MOSCAP (de-)interleaver followed by an experimental demonstration linking these two together.

A. QD Optical Frequency Comb (OFC) Source
On this platform, heterogeneous QD comb lasers are used because homogeneous broadening and low partition noise of QDs offer a broad gain profile and quiet wavelength stream, thus a large number of comb lines can be realized. In addition, QDbased lasers work very efficiently at the elevated temperatures of a typical HPC system [14]. The comb laser used here is based on a coupled cavity defined by L1 (2.6 mm) and L2 (650 μm) as illustrated in Fig. 2(a) and has a center comb wavelength of ∼ 1324 nm. L1 includes a 1.2 mm long SOA section, passive silicon waveguides, and two MMI based loop mirrors designed for a reflectivity of 50%. A 180 μm long saturable absorber (SA) is centered within the SOA region. The external cavity (L2) is defined between the GC and the MMI loop mirror. The FSR of the comb laser is defined by Δλ = |(FSR 1 × FSR 2 ) / (FSR 1 -FSR 2 )| where the subscripts denote cavities L1 and L2. L2 was designed such that FSR 2 was 4·FSR 1 of cavity L1. This allows us to use a relatively long cavity (thus increasing laser output power when compared to a shorter cavity) while maintaining a relatively large comb channel spacing. A less ideal comb operation whose lasing spectra from the right-sided GC ( Fig. 2(b)) was chosen to study the impact of stronger and weaker comb lines to transmitter signal integrity. The measured mean channel spacing of Δλ is 63.93 GHz under a current bias of I bias = 270 mA and SA bias of V bias = −5.6 V. The laser testing stage temperature is set to T = 24.6°C. Fig. 2(c) shows the measured comb separation and associated power levels of each comb after measuring through a circulator and 1% directional coupler tap. The shaded green areas indicate the 17 comb lines that will be fed into the (de-)interleaver in the experimental part of Section III. 12 of the 17 combs exhibit relative 3 dB power variation. If we take into account measured grating coupler (GC) loss of ∼ 10 dB, circulator loss of 1dB and a 1% power tap (20 dB) for optical spectrum analyzer (OSA) measurements, the OFC is capable of a 17-wavelength stream (34 channels in total), where 12 out of the 17 combs should have ∼ -7 dBm/line within the output waveguide. We can further improve channel power by minimizing optical waveguide loss, improving loop mirror reflectivity (currently ∼ 50% by design), and integrating SOAs. Recently, we have developed integrated straight Sections III-V/Si QD SOAs capable of 18, 15, and 12 dB small signal gain at 1300 nm for operating temperature of T = 20, 50, and 80°C. No optical power penalty was observed for a modulated signal or comb laser [24]. The comb line with the largest power (1322.02 nm) had a measured RIN of ∼ −144 dB/Hz. The adjacent combs at 1321.65 and 1322.39 nm had a measured RIN of −135.1 and −138.9 dB/Hz respectively. We have also demonstrated error free operation of the comb lines with a FEC-free bit-error-rate (BER) of 10 −12 for 5 measured comb lines [14].
The temperature sensitivity of the comb laser was measured to be ∼ 0.1 nm/°C and can lead to a spectral shift of 2.5 nm for temperature ranges from T = 25 to 50°C. However, we have shown MOSCAP micro-ring devices that can tune up to 2.0 nm at a bias of 6 V with nW power consumption [19].

B. Heterogeneous Platform for Ultra-Energy-Efficient MOSCAP Phase Tuners
Current state-of-the-art silicon photonic MZIs and (de-) interleavers use either power inefficient thermal or current injection phase shifters to compensate for either waveguide phase  errors, power splitting errors, or temperature drift [32], [33]. In this work, we employ the use of a III-V/Si metal-oxidesemiconductor capacitor (MOSCAP) structure to efficiently tune the phase errors in the (de-)interleaver structure such that channel XT is reduced. These ultra-power-efficient MOSCAP phase tuners are based on a GaAs/dielectric/Si heterogeneous structure as shown in Fig. 3(a) and (b). There have been many demonstrations of energy-efficient InGaAsP/InP/Al 2 O 3 MOSCAP phase tuners in the form of MZI structures [34]- [40], however, we believe we are the first to implement GaAs/Al 2 O 3 and GaAs/HfO 2 based MOSCAPs for (de-)interleaver structures at 1310 nm. In addition, we have demonstrated these MOSCAP phase shifters in lasers [12], [41], [42] and MRRs [17], [18] individually, thus enabling the possibility of a complete heterogeneously integrated transceiver system based on ultra-low power consumption. The single mode condition for this heterogeneous structure is defined by a width, height, and etch depth of 500, 300, and 170 nm respectively. The wafer-bonded III-V consists of a 190 nm-thick n-GaAs doped at 3×10 18 cm -3 . Fig. 4(a) and (b) illustrates the calculated transverse electric (TE) effective index change (Δn TE00 ) and associated free carrier absorption (FCA) optical losses for both n-GaAs/Al 2 O 3 /p-Si and n-GaAs/HfO 2 /p-Si structures vs. forward bias for various dielectric thicknesses. A 5 nm-thick Al 2 O 3 with a refractive index of n Al2O3 = 1.75, will yield calculated optical confinement factors of Γ Al2O3 = 1.154% and Γ III-V = 28.27% with an overall effective index of n eff = 3.1144. A 5 nm-thick HfO 2 with a refractive index of n HfO2 = 1.88 has optical confinements of Γ HfO2 = 1.159% and Γ III-V = 28.33% with an overall effective index of n eff = 3.1154. In forward bias, the MOSCAP structure operates in carrier accumulation mode for efficient phase change. Experimentally, we explored 2 different MOSCAP gate oxide designs with varying degrees of Si doping and a dielectric selection of Al 2 O 3 and/or HfO 2 . These designs are shown in Table I.
The bulk HfO 2 dielectric has a higher dielectric constant (k ∼ 25) compared to Al 2 O 3 (k ∼ 9), therefore an HfO 2 -based capacitor should have ∼3× the capacitance for the same unit area. However, the measured thickness of HfO 2 in Design 2 is 10 nm with an additional 3 nm of Al 2 O 3 which indicates the capacitance is only ∼ 1.7× or less compared to Design 1.
Initial phase tuning measurements were performed on a 350 μm-long p-Si/Al 2 O 3 (6 nm)/n-GaAs MOSCAP MZI structure and spectral responses indicated ∼ 1.69 nm of tuning at a 2 V bias while maintaining an extinction ratio of ∼ 24 dB. The measured FSR = 17.9 nm and the calculated V π L = 0.370 V-cm which is 4× smaller than typical values seen in PN junction-based phase tuners. The observed leakage current appears to be smaller than the current meter limit of sub-nA, indicating negligible power consumption. The MOSCAP MZI is capable of achieving RC constant-limited 4 Gb/s eye diagrams and an f 3dB ∼ 1.5 GHz. We also measured a MZI with HfO 2 (10 nm)/Al 2 O 3 (3 nm) dielectric as shown in Fig. 4(b). The measured FSR = 19.0 nm with 2.26 nm of tuning at a 2V bias while maintaining an extinction ratio of ∼ 30 dB. The V π L was slightly lower (V π L = 0.294 V-cm) than the Al 2 O 3 counterpart even with double the dielectric thickness. TE waveguide losses using the cut-back method were measured for pre-bonded 0.5 and 0.8 μm wide waveguides and exhibited optical losses of ∼ 9.2 and 9.8 dB/cm, respectively, at 1310 nm. After III-V removal, TE waveguide losses were measured to be ∼ 21.1 and 42.2 dB/cm for the 0.8 and 0.5 μm wide waveguides respectively. TE waveguide losses with the wafer-bonded III-V regions were undetermined due to the absence of these particular test structures. Root causes of high waveguide losses are still under investigation.

C. MOSCAP (de-)Interleaver
MOSCAP phase tuners provide power efficient phase error correction for the heterogeneous (de-)interleavers needed to separate the OFC lines. Various (de-)interleaver designs such as 2nd/3rd order asymmetric MZIs (AMZIs), single, double, triple-ring assisted AMZIs and cascaded 2 nd order micro-ring filters were explored. Unlike designs based on power inefficient thermal tuners, each of our designs includes MOSCAP phase tuners integrated onto the delay paths such that nay phase errors can be corrected to achieve high extinction ratio. For this work, we chose to work with a one-ring-assisted AMZI (1-RAMZI), shown in Fig. 5(a), because of decent insertion loss and XT performance. Ring-assisted AMZIs with N number of rings have shown to exhibit wider flat-top response with improved channel XT [43]. There have been a number of demonstrations on silicon with channel spacing ranging from 120-1250 GHz and channel XT ∼ −22 dB [27], [33], [43]- [46]. The 1-ring assisted AMZI (1-ring RAMZI) is modeled by [47]: The through and cross port transmission are respectively defined as c 0,1 = 1 − κ 0,1 (λ) and −js 0,1 = −j κ 0,1 (λ), where κ 0,1 is the power coupling coefficient for each coupler and κ r is the ring coupling coefficient. The FSR is defined by the ring circumference such that the FSR = c/n g /L ring . Therefore, a channel spacing of 65 GHz for the 1-ring AMZI requires L ring = 1200 μm for a calculated group index of n g = 3.78. The ideal ring resonator coupling for a 1-RAMZI occurs when κ r = 0.89, and c 0,1 = 0.50. The spectral characteristics of the MOSCAP (de-)interleavers are measured by coupling a Thorlabs superluminescent diode with 40 nm of bandwidth (1290-1330 nm) and a launch power of ∼ 12 dBm. These devices are fabricated on a 100 mm SOI wafer and is vacuum mounted onto a sample stage of a semi-automatic probe station for measurements at room temperature. Fig. 5(a) and (b) illustrates the device schematic of the 1-RAMZI and calculated transmission spectrum for the bar and cross state. The solid lines in Fig. 5(b) represent the condition for perfect phase matching whereas the dashed lines indicate the case for a π phase offset for the ring resonator. The measured bar and cross channels for the 1-RAMZI before phase tuning show at XT ∼ −7.3 dB and −10.7 dB. By applying a −2 V bias on the delay length (V delay ), the XT of the bar channel was improved from -7.3 dB to −16.4 dB while the cross channel XT improved from −10.7 dB to −26.6 dB. For a bias of V delay = -2 V, approximately 5.0 nA was drawn resulting in a tuning power consumption of 10.0 nW [16]. In addition to ∼ 5 dB lower than the reported best XT for a 1-RAMZI design, there is a 2.5 million reduction in power consumption [33]. In parallel, a similar 1-RAMZI based on p-doped silicon heaters was measured within the same die. These p-doped (1e20 cm -3 ) heaters have a width = 2 μm and are placed adjacent to the ring and delay waveguides by an offset of 2 μm. Initial phase error correction experiments show a power consumption of 27.59 mW for similar XT levels compared to the MOSCAP devices. It should be noted that optical IL will have the effect of reducing passband flatness, therefore, waveguide scattering should be kept to a minimum. Simulations indicate ∼ 14% reduction of the 0.5 dB bandwidth for every 10 dB/cm loss incurred in the device. The measured spectral responses shown in Fig. 5(c) and (d) obviously indicate passband shapes that are far from theoretically calculated flat-top response shown in Fig. 5(b).
This indicates a combination of waveguide loss and errors in power coupling coefficients. It was determined that coupling coefficients of c 0 = 0.46, c 1 = 0.43, κ r = 0.88, and a waveguide loss ∼ 40 dB/cm resulted in a similar spectrum as shown in Fig. 5(d). None flat-top response is mainly attributed to waveguide loss, whereas increased channel XT indicates non-ideal power coupling coefficients. Experimentally, it is difficult to determine how much the coupling coefficients are off from ideal, however, 3D-FDTD simulations indicate a +/-20 nm etch depth variation from the nominal etch (170 nm), will produce a -/+ 10% change in the coupling coefficient. We plan to fabricate an improved design which consists of robust 50% MMI power couplers for c 0 and c 1 as well as tunable MOSCAP directional couplers for κ r . Initial MMI designs show ∼ +/-1% variation from a nominal coupling coefficient of 50%, whereas a typical directional coupler show ∼ +/-10% variation. Another loss mechanism to consider are the III-V/Si interface transitions. The (de-)interleavers employ an angled III-V interface of 45°with respect to the silicon waveguide. Numerical three-dimensional finite-difference-time-domain (3D-FDTD) simulations indicate a theoretical insertion loss of 0.36 dB/facet with a reflection < -60 dB. Experimentally determined III-V/Si interface losses were evaluated by cutback loss structures which indicate a loss of 1.08, 0.69, and 0.29 dB/facet for interface angles of 0°, 45°, and 72°. We believe these transition losses can be further lowered since Ohno, et al., have demonstrated promising transition losses of ∼ 0.055 dB [37].

D. MOSCAP Micro-Ring Modulator
In this section, we discuss a heterogeneous III-V/Si MOSCAP micro-ring modulator capable of transmitting 25 Gb/s OOK data at temperatures up to 80°C. The modulators are fabricated using the exact same process as that of the QD comb laser and (de-)interleaver. Fig. 6(a) shows a perspective schematic of the modulator as well as the cross-section and a TEM image of the Al 2 O 3 dielectric bonding interface. This particular ring has a radius = 10 μm with a 0°coupling angle. Fig. 6(b) shows the measured normalized transmission spectra for various bias voltages from 0 to 7 V, indicating a phase tuning efficiency of ∼ 1.1 V-cm. As bias voltage increases, the resonance blue shifts with a reduction in quality factor (Q-factor) due to plasma dispersion and increased loss from free carrier absorption (FCA). This also indicates the resonator is initially under-coupled. We have extracted the roundtrip loss, power coupling coefficient, and Q factor as a function of wavelength using the methods identified in [48], and are 13 cm -1 , 0.0494, and 8250 respectively.
Next, we measured the small signal response (S 21 ) for several ring diameters of 10, 20, and 30 μm as shown in Fig. 6(c). The S 21 curves are highly dependent on detuning between the laser wavelength relative to the micro-ring resonance. In this case, the detuning is such that optical modulation amplitude (OMA) is maximized which leads to a f 3dB ∼ 15 GHz for the 10 μm radius ring. The resistance-capacitance (RC) time constant bandwidth does not vary significantly with radius mainly because as capacitance increases with a larger radius, the resistance decreases proportionally due to increased contact area. Next, we evaluated large signal performance. Fig. 6(d) shows the measured eye diagrams at 25 Gb/s with a PRBS15 pattern and a peak-to-peak swing of V pp = 4 V for three temperatures of 20, 50, and 80°C. The extinction ratio (ER) is ∼ 5.7 dB at all temperatures. The change in resonance wavelength with stage temperature, without any self-heating effect from two-photon absorption (TPA)-induced free-carrier absorption (FCA) is ∼ 73.6 pm/°C. As mentioned in Section II A., the comb laser has a temperature shift of ∼ 100.0 pm/°C, thus indicating ∼ 0.66 nm drift from a comb line relative to a ring resonance for a temperature increase of 25 to 50°C. Our previous work on HfO 2 based MOSCAP micro-ring [19] has a tuning range of ∼ 2.0 nm and should be able to accommodate temperature drift offset.

III. DWDM TRANSCEIVER EXPERIMENT
In this section, we discuss the experimental demonstration of the heterogeneous III-V/Si transmitter based on (de-)interleaved frequency combs without the MOSCAP MRR [17], [19]. However, these MRRs have been independently demonstrated on the same platform but in different chips and will be fully integrated in future versions. The experimental setup is illustrated in Fig. 7. The OFC QD laser is mounted on a temperature controlled stainless steel and biased with the conditions mentioned in Section II A. This operation condition enables the generation of comb lines where 12 out of the 17 combs exhibit a 3 dB power variation. The total power consumption of the OFC QD laser is 1309.08 mW, however, if we consider symmetric dual outputs, only half (654.54 mW) should be considered. The combs from the right GC are coupled via a cleaved angled fiber (26°) with a GC loss of ∼ 10 dB. The signal then goes into a circulator to prevent feedback instability into the laser, and then tapped off at 1% for both power and spectrum monitoring respectively. This is then amplified with a PDFA with a gain of ∼ 15 dB and then coupled in/out of the 1-RAMZI (de-)interleaver mounted on a separate stage with a polished angled (7°) fiber array. GC loss were estimated to be ∼ 22 dB. The 1-RAMZI (de-)interleaver spectrum is normalized to a 0.5 μm wide straight waveguide with a length of ∼ 850 μm as shown in Fig. 8(a) along with the associated extinction ration (ER) between bar and cross ports in Fig. 8(b). The mean ER for both bar and cross ports across the 17 (de-)interleaved channels are ∼ 14.3 and 16.0 dB respectively. The non-uniformity of the (de-)interleaved combs is due to both OFC comb source non-uniformity as well as imperfect comb alignment with the peak passband of the (de-)interleaver (Δλ comb = 63.93 GHz, Δλ (de-)interleaver = 65.78 GHz). Comb laser alignment with the (de-)interleaver was optimal at 1322.02 nm and sequential (de-)interleaved combs become misaligned for longer wavelengths. As a result, the (de-)interleaved comb near 1326 nm will see an extra 3 dB loss. We expect any misalignments to be remedied in the future with MOSCAP tuning and improved waveguide losses which have a significant effect on passband flatness as discussed in Section II C. Since the comb laser is based on a coupled cavity design, channel spacing can also be tuned via MOSCAP phase shifters or thermal tuners.
Next, the signal goes through SOA1 and SOA2 designed for a center wavelength of 1310 and 1350 nm respectively. The first tunable filter prevents gain saturation into SOA2 and the second tunable filter selects the desired de-interleaved comb line with bandpass of 80 pm. The selected comb line is fed into a 65 GHz LiNbO 3 modulator and then into a 40 GHz photodetector with a 0.7 A/W responsivity. Eye diagrams are then monitored on a 60 GHz DCA and we were able to (de-)interleave 8 and 9 comb lines onto the bar and cross waveguide respectively for a total of 17 combs lines. Each line was modulated at 25 Gb/s PRBS9 NRZ OOK for a total of 425 Gb/s as shown in Fig. 9. 25 Gb/s was chosen due to project goals, however, higher data rates are indeed possible. All 17 comb lines show open eye diagrams without equalization which correspond to signal-to-noise ratios (SNRs) from ∼ 5.3 to ∼ 8.6 dB. The eye noise should mainly come from the amplified spontaneous emission (ASE) noise generated by multiple optical amplifiers. ASE noise is proportional to optical power, therefore, the "1" level exhibits more noise than the "0" level and is usually more significant for larger eye diagrams. This can be alleviated by eliminating the GC coupling loss after integrating them together. The eye amplitude variance at different wavelengths is mainly due to the optical power and RIN differences per comb line as well as the wavelength offset between the (de-)interleaver and OFC laser. It should be noted that the OFC source center wavelength is ∼ 1324 nm, whereas the PDFA gain profile decreases significantly at these longer wavelengths. We believe more combs could have been measured if we were not limited by the PDFA bandwidth. In the future, either the OFC laser will be centered at 1310 nm, or the (de-)interleaver will use broadband couplers to accommodate improved XT performance in excess of 50 nm bandwidth. Table II itemizes measured performance numbers of the (de-)interleaved comb lines. Eye signals with the largest SNR are typically associated with high ER and associated comb powers.
Inter-channel XT with multiple modulated signals can be of a concern. The inter-channel XT can be determined by 10log 10 (V xt /V sig ) where V xt and V sig is the mean and sigma of level 1 respectively [49]- [51]. We have simulated the case where two lasers spaced 121.0 GHz apart are each modulated at 25 Gb/s. For the two laser case, the XT = −16.06 dB, whereas for only 1 laser, the XT = −16.70 dB.

IV. TRANCEIVER ARCHITECTURE SPECIFICATIONS
In this section, we discuss details of the transceiver link budget and projected energy efficiency. Table III lists the optical losses of each device and the power consumption in the proposed DWDM link operating at T = 50°C. All loss numbers are based on past in-house fabrication and can be improved in an established foundry. We assume a worse case comb laser power/line to be ∼ -12.2 dBm with a power consumption of ∼ 500 mW. The booster SOAs that follow the comb laser have an experimentally determined gain of 15 dB with ∼ 300 mW power consumption. Two (de-)interleavers are employed to spatially separate the front and back comb laser signal into 4 spatial channels that consists of 10 MRR each. This results in an aggregate of 40 MRR operating at 25 Gb/s for a total of 1 Tb/s. Each MRR on a particular spatial channel is designed to have 128 GHz separation. The power consumption of each MRR was determined to be ∼ 0.58 mW and will be ∼ 23.2 mW with all 40 MRR being operated. The power consumption of the (de-)interleavers are negligible (nWs) and assumed to be ∼ 0 mWs. The total calculated loss per channel is −14 dBm which becomes the required APD sensitivity. Assuming a reasonable OSNR at 40 dB, a bit-error-rate (BER) of < 1e-9 should be attainable according on our past calculations on in-house designed APDs [52]. Assuming the same OSNR, if the comb laser power/line can be ∼ -7 dB as demonstrated in Section II A., there is reason to believe a BER < 1e-12 is achievable. In terms of total power consumption, this architecture will consume ∼ 862.3 mW, which yields a 0.86 pJ/bit energy-cost number if we assume aggregate 1.0 Tb/s from 40 MRRs operating at 25 Gb/s.

V. CONCLUSION
This work, for the first time, demonstrates a simultaneous wavelength and space division multiplexed transmitter fabricated on a heterogeneous III-V-on-silicon platform. The QD OFC source yields 12 out of 17 comb lines within a 3 dB power variation for a single facet output along with a FSR ∼ 63.9 GHz. Assuming symmetric dual outputs, the total power consumption of the OFC QD laser is estimated to be ∼ 654.54 mW. The MOSCAP based 1-RAMZI (de-)interleaver has improved XT levels down from −11 dB to −26 dB for tuning powers of only 10.0 nW. With both building blocks, we have demonstrated wavelength (de-)interleaving of a QD OFC source with a total of 17 comb lines each modulated at 25 Gb/s NRZ for an aggregate bandwidth of 425 Gb/s. Future demonstrations will include full integration of a 40 channel QD OFC laser source, MOSCAP (de-)interleavers, and 40 MRR modulators capable of > 1 Tb/s transmission.