High-capacity data transmission for optical I/O driven by efficient microcomb and microring modulator

High-capacity data transmission for optical I/O driven by efficient microcomb and microring modulator | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article High-capacity data transmission for optical I/O driven by efficient microcomb and microring modulator Liangjun Lu, Hongyi Zhang, Shihuan Ran, Yuanbin Liu, Shuxiao Wang, and 5 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5365298/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract The escalating demand for high-speed, low-power data transmission between processing units (XPUs) has underscored the limitations of traditional electrical input/output (I/O) technologies. Silicon photonics emerges as a promising solution for chip-level optical I/O by integrating Kerr microcombs, microring-based modulators, and photodetectors. In this study, we demonstrate a record-breaking error-free optical I/O transmission achieving 2.3 Tbit/s per fiber port. This feat is enabled by dark soliton microcombs generated in a 400-nm-thick Si 3 N 4 microring, exhibiting a high conversion efficiency of 49% and an on-chip spectral bandwidth of 28 nm at -5 dBm, achieved through precise coupling and dispersion engineering. Utilizing a silicon microring modulator with an electro-optic bandwidth of 61.7 GHz, 36 comb lines are encoded with PCIe6.0-compatible 64 Gbit/s on-off keying (OOK) signals. Additionally, these comb lines support 100 Gbit/s OOK per channel with a bit error rate (BER) of 10 − 10 . The successful integration of these foundry-compatible platforms confirms the viability of microcomb-based optical I/O, paving the way for the next generation of high-speed, energy-efficient data communication systems. Physical sciences/Optics and photonics/Other photonics/Solitons Physical sciences/Optics and photonics/Applied optics/Integrated optics Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction As the accuracy and comprehensiveness demands of artificial intelligence (AI) applications escalate, the size of clusters for large language models correspondingly expands [ 1 ]. For instance, for contemporary large model training, such as those conducted by OpenAI, Meta, and Inflection AI, etc, clusters comprising over 10,000 GPUs have been utilized to train and process models with more than a trillion parameters [ 2 – 6 ]. Although server hardware floating-point operations per second (Flops) have exhibited a tripling approximately every two years over the past two decades, the growth rates of dynamic random-access memory (DRAM) and interconnect bandwidth are only 1.6× and 1.4× every two years, respectively [ 7 ]. It is anticipated that as cluster sizes continue to grow, their performance will increasingly be constrained by interconnect bandwidth, making enhancements in this area both critical and urgent [ 7 , 8 ]. Furthermore, the power consumption of interconnects within GPU clusters is a crucial consideration. As cluster size expands, the energy consumption of network devices and interconnections facilitating GPU communication also escalates, potentially affecting not only energy efficiency but also overall operational costs. High-capacity inter-chip communication with more than TB/s has emerged as a critical requirement in technological advancements. Traditional XPU-to-XPU electronic input/output (I/O) is increasingly proving inadequate in meeting the heightened requirements for high-speed interconnects and reduced power consumption [ 9 – 11 ]. Optical I/O, possessing the capability for terahertz bandwidth with minimal loss and crosstalk, has been advanced as a superior alternative to traditional electrical interconnects for XPU-to-XPU communication [ 10 , 12 , 13 ]. This innovation not only supports substantial increases in bandwidth and reductions in latency but also achieves remarkable improvements in energy efficiency [ 13 , 14 ]. Compared to traditional intra- and inter-datacenter optical interconnects, optical I/O between XPUs requires significantly lower bit error rates (BERs), higher I/O bandwidth density, lower power consumption, even to the scale of a chip's dimensions, and little or no equalization [ 10 ]. Intensity-modulation, direct-detection (IM/DD) is favored for the optical interconnects between the XPUs, due to its simplicity, power efficiency, and cost-effectiveness. Error-free transmission is critical to ensure high data integrity, which can otherwise degrade performance and efficiency. By leveraging IM/DD, the system can achieve this reliability without the need for complex Forward Error Correction (FEC) algorithms that rely heavily on Digital Signal Processing (DSP). This reduction in FEC requirements directly lowers both power consumption and latency. Error-free transmission also enhances system performance and simplifies network design, especially for low-latency, high-throughput communication. Since the single lane data rate of the IM/DD scheme is much lower than the coherent scheme, wavelength division multiplexing (WDM) technology is commonly employed to further expand the transmission capacity. Therefore, optical I/O architectures typically incorporate multi-wavelength laser sources, modulators, (de)multiplexers, and photodetectors (PDs) to facilitate data transmission and reception [ 15 ]. These devices have been implemented on a silicon photonics platform either monolithically or as part of the system due to the inherent advantages of complementary metal oxide semiconductor (CMOS) compatibility, high integration density, low loss, and high bandwidth [ 15 , 16 ]. Therefore, we believe the integration of optical I/O on the silicon photonics platform has the potential to unlock unprecedented performance enhancements. For the multi-wavelength laser source, several solutions have been proposed for WDM parallel data transmission. Laser arrays, such as distributed feedback laser (DFB laser) [ 17 ] and vertical-cavity surface-emitting laser (VCSEL) [ 18 ] arrays, etc, have already been commercially implemented in the coherent or IMDD transceiver for data center interconnects. However, for the WDM transmission, the employment of multiple carriers requires independent control of each laser, leading to increased system complexity and footprint as the number of lasers scales up. Based on the electro-optic (EO) modulation, the EO comb is constructed with versatility, excellent comb stability, and phase coherence through cascading intensity and phase modulators [ 19 ]. On-chip EO comb based on thin-film lithium niobate (TFLN) features a conversion efficiency of 30% and an optical bandwidth of 132 nm [ 20 ]. Harnessing the full potential of EO combs has enabled remarkable advancements in data transmission, achieving rates exceeding several Tb/s and extending the reach to hundreds of kilometers [ 21 , 22 ]. Nevertheless, the EO combs are constrained by the bandwidth of the modulators, making it challenging to achieve larger comb line spacing, which in turn further restricts the line rate of the single carrier. Quantum dot-mode locked laser (QD-MLL) combs have demonstrated their advantages as WDM sources due to their compact, efficient, robust, and dispersion-insensitive characteristics. Previous works have reported more than 10 Tb/s WDM transmission based on the QD-MLL comb [ 23 , 24 ]. Soliton microcombs, generated in the high-nonlinearity microring, can easily realize dozens of phase-coherent and evenly spaced optical frequency lines with low noise and 10’s GHz ~ 1 THz comb spacings [ 25 – 33 ]. It is highly desired to expand the communication capacity of optical I/O [ 34 – 37 ]. Through dispersion engineering, soliton microcombs are capable of generating significantly broad spectral bandwidths that encompass the entire O, C, and L bands [ 38 , 39 ], greatly exceeding the spectral bandwidth of the QD-MLL comb [ 39 , 40 ]. In the previous research, dissipative Kerr soliton (DKS) generated in the anomalous dispersion regime has been extensively studied. The single soliton, a popular type of DKSs, exhibits a smooth Sech2-shaped spectrum but has a low pump-to-comb conversion efficiency of only ~ 1% [ 41 , 42 ]. Therefore, the comb lines of the single soliton microcomb exhibit very low line power, which will reduce the signal-to-noise ratio (SNR) and lower the data rate for IM/DD data transmission. In contrast, dark solitons generated in the normal dispersion regime, show the merits of high conversion efficiency and robustness to pump laser tuning. High-power comb lines supported by the efficient dark soliton can ensure high-performance data transmission, maintaining low BER or error-free operation when carrying high-rate signals, which is crucial for optical I/O interconnects [ 15 ]. Silicon microring modulators (MRMs) are favored for high-density optical I/O due to their compactness, high modulation efficiency, low power consumption, and wavelength selectivity, which are suitable for comb-based parallel data transmission [ 43 , 44 ]. In this work, we propose and demonstrate a high-capacity error-free data transmission driven by an efficient dark soliton microcomb and silicon MRMs for the optical I/O application. The conceptual diagram is shown in Fig. 1 A. Microring-based comb source and modulators are implemented for compact chip size and low power consumption. The dark soliton comb source is constructed using a 400-nm-thick Si 3 N 4 microring with the assistance of avoided mode crosses (AMXs). With optimized power coupling ratio and group velocity dispersion (GVD) of the Si 3 N 4 microring, we generate a dark soliton with a record high conversion efficiency of close to 50% and an on-chip − 5-dBm bandwidth of 28 nm. We further use the 36 comb lines in the − 5-dBm bandwidth as the optical carriers for data transmission, which are modulated by a silicon MRM with an electro-optic bandwidth of 61.7 GHz. The line rate per channel reaches 64 Gb/s with on-off keying (OOK) modulation, meeting the data rate requirements in the next-generation I/O standard PCIe 6.0 [ 45 ]. A total 2.3 Tbit/s OOK signal error-free (BER < 10–12) transmission is successfully realized. Among the 36 comb lines, 32 comb lines’ BER is measured with the original signal, and only 4 comb lines’ BER is measured after the use of feed-forward equalization (FFE). As shown in Fig. 1 B, in comparison with state-of-the-art error-free optical interconnects [ 15 , 46 – 53 ], our work achieves remarkable advancement in both the total capacity per fiber port of 2.3 Tb/s and the lane rate per wavelength channel of 64 Gb/s, benefiting from our highly-efficient multi-wavelength laser source and high-speed silicon MRMs. Our proposed error-free parallel data transmission scheme is highly promising for high-capacity low-latency and power-efficient optical I/O applications. Results Dark soliton generation To create a high-performance WDM comb source, we focus on two aspects. Firstly, the microcomb is expected to have a high conversion efficiency for supporting high comb line power. Secondly, it is essential to have a relatively flat and wide spectrum so that more high-power comb lines are available. Figure 2 A shows the dispersion and dissipation engineering for constructing the microcomb source with both broad bandwidth and high optical power. We used the Lugiato-Lefever equation (LLE) for the theoretical simulation of the dark soliton comb source [ 46 ] and the waveguide structure design is based on the simulation from Lumerical FDTD and Mode solution. The details about the engineering are discussed in the supplementary information. According to the simulation, the waveguide dimension of the Si 3 N 4 microring is designed to be 0.4 µm (height) × 2 µm (width) with a simulated 𝛽2 (GVD) of 560 ps2/km, supporting multiple modes for AMX and low transmission loss simultaneously. The width of the straight bus waveguide is designed to be 1.5 µm, so that both the fundamental (TE0) and the first-order transverse electric (TE1) modes can be exited in the microring in order to trigger mode coupling for the dark soliton generation. The gap between the straight bus waveguide and the microring is 0.5 µm. The radius of the microring is 246 µm, corresponding to an FSR of 96.2 GHz. The Si 3 N 4 microring was fabricated on an 8-inch wafer using 193-nm photolithography by Shanghai Industrial µTechnology Research Institute (SITRI). Figure 2 B shows the packaged chip with a pair of fiber arrays, facilitating stable optical coupling between the fiber and the chip in the experiments. The right side of Fig. 2 B shows the microscope image of the fabricated microring resonator. Figure 2 C shows the measured and fitted spectra at the pump resonance. The extracted power coupling ratio is 0.011, and the loss is about 0.04 dB/cm. The loaded quality (Q) factor is approximately 1×106. By extracting the resonance frequency of each longitudinal mode (𝜔𝜇), we calculated the integrated dispersion curve \(\:{D}_{int}\:=\:{\omega\:}_{\mu\:}\:-\:{\omega\:}_{o}\:-\:{D}_{1}\mu\:\) , where \(\:{\omega\:}_{o}\) is the resonance frequency of the center mode, \(\:\mu\:\) is defined as the mode index in terms of the center mode ( \(\:\mu\:=0\) ), and \(\:{D}_{1}/2\pi\:\) is the equidistant FSR, as shown in Fig. 2 D. The second-order dispersion ( \(\:{D}_{2}/2\pi\:\) ) is -5.2 MHz, corresponding to 𝛽2 of 608.6 ps2 /km, which is in the optimal region for large − 5-dBm optical bandwidth. We used a continuous-wave (CW) laser to pump the Si 3 N 4 microring, which was amplified to 29 dBm by an erbium-doped fiber amplifier (EDFA). The spectrum of the generated dark soliton microcomb is shown in Fig. 2 E. The measured conversion efficiency and on-chip − 5-dBm (subtracting 5 dB coupling loss from the optical fiber) bandwidth of the microcomb are 49% and 28 nm, respectively, which are consistent with the simulation. There are 36 comb lines in the on-chip − 5-dBm bandwidth and nearly 88.9% of them have a power variation of less than 8 dB (-10 dBm to -2 dBm), indicating the dark soliton features a flat spectrum. When the pump power is set at various powers, the formation of dark soliton is different, which is discussed in supplementary information. We also explore the stability and frequency noise of dark soliton in the supplementary information. Large-bandwidth silicon microring modulator Figure 3 A illustrates the schematic diagram of the designed MRM. The microring has a radius of 8 µm, a rib waveguide height of 220 nm, a width of 500 nm, and a slab thickness of 90 nm. A horizontal PN junction is embedded in the microring waveguide for modulation. The gap between the microring and the bus waveguide is 250 nm, with a coupling region length of 1.8 µm. The doping concentrations of both P-type and N-type are 2×1018 cm − 3 in the light doping regions and 1×1020 cm − 3 in the heavy doping regions. The chip was fabricated on the SOI platform using a standard foundry process. The microscope image of the chip is shown in Fig. 3 B. To enable flexible adjustment of the MRM's operating wavelength, a titanium nitride-based thermal phase shifter is integrated on top of the microring modulator. As shown in Fig. 3 C, for the MRM with a measured FSR of 11.39 nm, a thermo-optic tuning efficiency of 0.22 nm/mW (Pπ = 26.46 mW) is achieved. Specifically, at the thermal tuning power of 55.13 mW, the resonance wavelength redshifts by 11.86 nm, surpassing one FSR. Thus, the MRM's operating wavelength can be tailored to any desired wavelength by precise control of the heating power. Figure 3 D shows the transmission spectra of the MRR at various bias voltages. At 0 V bias, the MRM has a resonance wavelength of 1545.1 nm with a Q-factor of approximately 2870. The resonance wavelength redshifts upon increased reverse bias voltage, with a modulation efficiency of 25 pm/V. Figure 3 E displays the EO response of the microring modulator. Upon − 3 V bias, the EO bandwidth at the 3 dB IL operation point is 61.7 GHz and 38.6 GHz at the 6 dB IL operation point. Figure 3 F presents the real and imaginary parts of the S11 parameters under the − 3 V bias, with the PN junction capacitance extracted as 21.85 fF based on the model fitting in [ 54 ]. For high-speed modulation, we implemented 50 Gbit/s, 80 Gbit/s, and 100 Gbit/s OOK signal on the MRR with drive voltages of 2.1 Vpp, 2.8 Vpp, and 2.0 Vpp, respectively, at the 6 dB IL operation point for larger extinction ratio. A commercial CW laser was employed as the light source with an output power of 13 dBm. Figure 3 G exhibits the measured eye diagrams. The 50 Gbit/s and 80 Gbit/s OOK modulation achieved BERs of less than 1×10 − 18 for error-free transmission without any DSP, while the 100 Gbit/s OOK modulation attained a BER of 2×10 − 13 with 7-tap feed-forward equalization. The power consumption for OOK at the three data rates was calculated to be 17.8 fJ/bit, 31.6 fJ/bit, and 16.1 fJ/bit, respectively, using the formulae in [ 55 ]. These results underscore the capability of the MRM to achieve error-free high-speed OOK signal transmission with low power consumption, positioning it as an excellent candidate for optical I/O applications. Error-free parallel optical data transmission We implemented the error-free transmission based on the efficient dark soliton microcomb and the silicon MRM. The experimental setup is shown in Fig. 4 A. An amplified CW laser was coupled to the Si 3 N 4 chip before being adjusted to the fundamental TE mode by a polarization controller. An additional bandpass filter with a 3-dB bandwidth of 75 GHz was employed to eliminate the amplifier ASE noise from the EDFA before injecting the pump light into the Si 3 N 4 microring for comb generation. Then, we filtered out each of the comb lines within the on-chip − 5-dBm bandwidth and sequentially encoded them with 64 Gbit/s OOK by the silicon MRM. Note that within on-chip − 5-dBm bandwidth, a total of 36 comb lines are present. We selected the first 35 comb lines and excluded the last one (1572.8 nm), limited by the effective amplification bandwidth of the EDFA. Instead, we opted for an adjacent comb line located at the shorter wavelength end of the on-chip − 5-dBm wavelength range, which has a power slightly below − 10 dBm. As a result, the total number of carriers remains at 36 and the combined modulated optical spectra are shown in Fig. 4 B. Figure 4 C reveals the recorded BER for each channel upon a received optical power of 2 dBm. Owning to the limitations of the DCA, the minimum detectable BER is constrained to 10–18. Consequently, all BER below this threshold are recorded as 1×10–18 in the results. As evident from the graph, in Ch. 9 to Ch. 25, the BERs reach 1×10–18 except for the pump channel (Ch. 17). To mitigate the residual ASE noise from the EDFA, a 7-tap FFE is applied to reduce the BER of the pump channel from 10 − 8 to 2×10–13 for error-free operation. As noted, the channels distant from the pump mode are subjected to increased noise due to thermal noise-induced repetition rate instability [ 56 ]. Since the effective amplification band of EDFA is in the C-band, more ASE noise was introduced at both edges of the comb spectrum. Consequently, the channels on the two sides exhibited larger BER. For Ch. 1, Ch. 2, and Ch. 36, the 64-tap, 7-tap, and 15-tap FFE were implemented to minimize the corresponding BER into the error-free regime. We achieve 2.3 Tbit/s (36× 64 Gbit/s) error-free transmission and 2 Tbit/s (32×64 Gbit/s) with both error-free and FFE-free operation. Figure 4 D shows the measured eye diagram for all 36 channels, the SNR for nearly all channels exceeds 6.5 dB which exhibits good and consistent results. Furthermore, Ch. 4, Ch. 9, and Ch. 21 were selected to measure the BER versus received optical power. For comparative analysis, we utilize CW lasers with identical wavelengths and power levels as these chosen comb lines for OOK modulation. At each received optical power, the CW laser-based transmission link exhibits slightly lower BER compared to the comb channel. The average power penalty is less than 1 dB at the BER of 10–14, demonstrating a comparable performance between these two sources. Besides, the measured results demonstrate the BER can be reduced to less than 10–12 when the received power exceeds − 1 dBm. Consequently, we can decrease the previously mentioned received power from 2 dBm while still maintaining error-free operation. Furthermore, to investigate the capability of the efficient dark microcomb for carrying higher-speed signals, three comb lines were selected and modulated with 80 Gbit/s and 100 Gbit/s OOK signals. Figures 4 F-H show the BER curves of 80 Gbit/s and 100 Gbit/s OOK modulation versus the received optical power for Ch. 9, Ch.18, and Ch. 27. For 80 Gbit/s OOK modulation, after the 11-tap FFE, the BER of all the chosen three channels can achieve error-free operation, even reaching the lower boundary of 1×10–18. For 100 Gbit/s OOK modulation, the lowest BER is around 1×10–10 after the application of 64-tap FFE. When loading higher-rate signals, there are increased requirements on the carrier itself, such as relative intensity noise (RIN). By employing a pump source with lower RIN and mitigating the link loss to reduce the use of EDFA, error-free transmission of the high-speed signal can be maintained. Discussion In this proof-of-concept experimental demonstration, mounts of microcomb lines were loaded with a modulation rate of 64 Gbit/s to achieve DSP-free error-free transmission for high-capacity optical I/O. Both the single line rate and the total line rate significantly exceeded the previously demonstrated error-free transmission results. The comb spacing of the microcomb is ~ 96 GHz, and the spectral efficiency yields 0.66 bit/s/Hz which is close to the limit of 0.75 bit/s/Hz for the comb-based OOK error-free transmission due to the crosstalk between adjacent channels [ 57 ]. Compared with Mach-Zehnder modulators (MZMs), MRMs with their low power consumption and compactness, are highly suitable for optical I/O applications. In our demonstration, we have utilized silicon MRMs, fabricated at the wafer scale by commercial foundries, facilitating the requirements of a compact size and extremely low driving power. A modulation density of 1.06 Tbps/mm2 and an energy efficiency of 17.8 fJ/bit was achieved with our silicon MRMs. For the microcombs, with the optimized coupling and dispersion, we have developed a comb source with both high conversion efficiency and broad bandwidth, satisfying the power requirement for parallel data transmission. The entire design and optimization flow show a simple and feasible path to creating the high-performance on-chip WDM comb source for optical I/O. Additionally, it is worth noting that our microcomb source is based on a standard low-loss low-pressure chemical vapor deposition (LPCVD) silicon nitride platform with a thickness of 400 nm, a specification that has been standardized across various commercial silicon photonic foundries for wafer-level fabrication. Compared to the fabrication process of low-loss thick (> 600 nm) silicon nitride films, its fabrication process is simpler, with higher yields and lower costs [ 58 , 59 ]. Therefore, the hybrid integration of these two foundry-compatible on-chip devices can promote the large-scale manufacturing of microcomb-driven optical I/O chips [ 36 , 37 ], resulting in reduced size, weight, power, and cost. Moreover, with the development of heterogenous or hybrid integration technologies, tunable lasers and optical amplifiers are possible to be integrated and make it possible for the manufacturing of the chip-scale microcomb source [ 60 , 61 ]. Conclusion We have successfully implemented an error-free WDM transmission for optical I/O systems, utilizing an efficient dark soliton microcomb and a high-speed MRM. Through meticulous optimization of the coupling and dispersion properties of a Si 3 N 4 microring, we effectively increased the number of high-power comb lines while maintaining a high conversion efficiency. This led to the generation of a flat-power dark soliton microcomb with a conversion efficiency nearing 50% and an on-chip optical bandwidth of 28 nm at -5 dBm, serving as a high-performance multi-wavelength light source. The compact silicon MRM exhibited a high EO bandwidth of 61.7 GHz and operated with a low drive voltage of 2.1 V, enabling high-speed signal modulation while maintaining low power consumption. Capitalizing on these advancements, we demonstrated a record total capacity of 2.3 Tbps per fiber port by modulating 36 comb lines with 64 Gbit/s OOK signals—comprising 32 channels operating without FFE and 4 channels utilizing 7-tap FFE. Additionally, we showed that a single comb line from the dark soliton microcomb is capable of supporting error-free transmission at 80 Gbit/s and low-error transmission at 100 Gbit/s. Our findings highlight the tremendous potential of microcomb-based optical I/O in achieving high-capacity inter-chip interconnects. The integration of efficient dark soliton microcombs with high-speed silicon MRMs not only demonstrates the feasibility of scaling data transmission rates but also paves the way for energy-efficient, high-bandwidth optical communication systems. This work lays a solid foundation for the development of next-generation optical interconnects, addressing the escalating demands of data centers and high-performance computing environments. Methods Numerical simulation We used the LLE equations to simulate the dark soliton microcomb source for the design of broad high-power bandwidth and the equations are expressed as follows: $$\:{T}_{R}\frac{\partial\:{E}_{p}\left(t,\tau\:\right)}{\partial\:t}=\left(-{\alpha\:}_{p}-i{\delta\:}_{p}-iL\frac{{\beta\:}_{2}}{2}\left(\frac{\partial\:}{\partial\:\tau\:}\right)+i\gamma\:L{\left|{E}_{p}\right|}^{2}\right){E}_{p}\left(t,\tau\:\right)+\sqrt{\theta\:}{E}_{p}^{in}$$ $$\:{E}_{out}={E}_{in}-\sqrt{\theta\:}{E}_{p}$$ where \(\:{E}_{p}\) is the intracavity optical field, \(\:{E}_{in}\) is the input optical field and \(\:{E}_{out}\) is the output optical field at the through port while \(\:t\) and \(\:\tau\:\) represent the fast time and slow time. In the simulation, the pump power ( \(\:{\left|{E}_{p}\right|}^{2}\) ) is set to be 200 mW, \(\:{\alpha\:}_{p}\) =0.00615 is the cavity loss, \(\:{\delta\:}_{p}\) is the normalized absolute detuning and \(\:\theta\:\) =0.011 is the power coupling ratio between the bus and microring waveguides. \(\:L\) = 246×2π µm is the cavity length which determines the comb spacing of the microring and \(\:\gamma\:\) = 1.22 m2W-1 is the nonlinear refractive index. We added an AMX-induced resonance shift of 200 MHz at the pump mode. Meanwhile, a weak noise is induced as the initial field to seed the intracavity field. The detuning is increased gradually as the intracavity field goes through the roundtrips while the total number of roundtrips is 200000. Chip fabrication The Si 3 N 4 chip was fabricated by Shanghai Industrial µTechnology Research Institute (SITRI) using an 8-inch Si wafer. The Si 3 N 4 thin film with a thickness of 400 nm was deposited based on the LPCVD process. The waveguides are patterned using 193-nm photolithography. The thicknesses of the bottom and top SiO2 claddings are 3.18 µm and 3.7 µm. The MRM was fabricated on an 8-inch SOI wafer with a top silicon layer thickness of 220 nm by Advanced Micro Foundry (AMF) using the standard fabrication processes. After the waveguide etching process, the modulation region was doped, with boron as the p-type dopant and phosphorus as the n-type dopant. Subsequently, silicon dioxide was grown, followed by the formation of contact openings and the deposition of metal electrodes and the heater. Method for dark soliton generation We put the Si 3 N 4 chip on a thermo-electric cooler (TEC) and used a commercial wavelength-tunable laser (Santec TSL-710) as the pump laser source. The pump laser wavelength was initially set at a fixed wavelength on the blue-detuned side of the resonance and tuned with a step of 0.05 pm. At the output, we used a notch filter to filter out the pump for measuring the comb power. The comb power was recorded after each pump tuning, and the spectrum was recorded by an optical spectrum analyzer (Yokogawa, AQ6370D) every ten detunings. For the measurement of the low-frequency noise, the comb line that locates 2-comb spacing away from the pump on the left side was filtered out. We used a variable optical attenuator (VOA) to adjust the line power to around 1.8 dBm and sent it to the PD with a bandwidth of 50 GHz. The RF spectrum between 0 and 2 GHz was recorded by an electrical signal analyzer (Keysight, N9000B), indicating the results of low-frequency noise. Method for error-free parallel data transmission For the error-free parallel data transmission, we used a bandpass filter (Santec, OTF-350) to filter out individual comb lines and encoded each carrier with 64 Gbit/s OOK by the MRM, respectively. The high-speed OOK was generated by an arbitrary waveform generator (AWG, Keysight M8199A). At the receiver, due to the loss of the MRM, another EDFA (Amonics AEDFA-23-B-FA) was used before being sent to a digital communications analyzer (DCA, Keysight DCA-X N1000A) to measure the eye diagram and BER. Another bandpass filter with a 3-dB bandwidth of 85 GHz (II-VI, WS-04000B) was used before the PD for simulating demultiplexing. A VOA was used before the PD to adjust the received optical power. For the MRM, a thermal phase shifter was employed to shift the micro-ring resonance which is closest to the selected comb line wavelength. This adjustment ensures that the comb line wavelength is positioned at the 6-dB IL operation point of the resonance. Declarations Acknowledgments We thank SJTU-Pinghu Institute of Intelligent Optoelectronics for their assistance with the high-speed transmission experiments, and chip packaging. Besides, we thank SITRI for the fabrication of the Si 3 N 4 chip and AMF for the fabrication of the SOI chip. Funding: We acknowledge the funding from the National Natural Science Foundation of China (62120106010, 62090052, 62135010, 62075128), and the Shanghai Science and Technology Committee Rising-Star Program (23QA1404500). Author contributions: H.Z. designed, simulated, and characterized the efficient dark soliton microcomb source. S.R. designed, simulated, and characterized the microring modulator. H.Z. and S.R. conceived the link architecture and performed the high-speed comb-based data transmission experiments. S.W. and Y.C. supported the fabrication of the Si 3 N 4 chip. All authors helped analyze the data. L.L., Y.Li, Y.Liu, H.Z., and R.S. prepared the manuscript. H.Z., R.S., L.L., Y.Li, Y.G., and L.Z. revised the manuscript. J.C., Y.Li., L.L., and L.Z. provided suggestions and feedback during the revisions. L.L. and Y.L. co-supervised the research. Competing interests: The authors declare that they have no competing interests. Data and materials availability: All data needed to evaluate the conclusions in the paper are present in the paper and/or the Supplementary Materials. Additional data related to this paper may be requested from the authors. References Naffziger S (2023) Innovations For Energy Efficient Generative AI, in 2023 International Electron Devices Meeting (IEDM) , 1–4 Openai Update: ChatGPT runs 10K Nvidia training GPUs with potential for thousands more https://www.fierceelectronics.com/sensors/chatgpt-runs-10k-nvidia-training-gpus-potential-thousands-more Meta Building Meta’s GenAI Infrastructure https ://engineering.fb.com/2024/03/12/data-center-engineering/building-metas-genai-infrastructure/ I. AI. Announcing our collaboration with NVIDIA and CoreWeave on MLPerf https://inflection.ai/blog/nvidia-coreweave-mlperf Utils G https://gpus.llm-utils.org/nvidia-h100-gpus-supply-and-demand/#fn:23 Chung J-W, Gu Y, Jang I et al (2024) Reducing Energy Bloat Large Model Train Gholami A, Yao Z, Kim S et al (2024) AI and Memory Wall. IEEE Micro 44:33–39 Sahni S, Abhyankar A, Aggarwal A et al (2024) Beyond the Beachfront: Integration of Silicon Photonic I/Os under a High-Power ASIC . in. Optical Fiber Communications Conference and Exhibition (OFC) . 2024. IEEE Rizzo A, Daudlin S, Novick A et al (2023) Petabit-Scale Silicon Photonic Interconnects With Integrated Kerr Frequency Combs. IEEE J Sel Top Quantum Electron 29:1–20 Young IA, Mohammed E, Liao JTS et al (2010) Optical I/O Technology for Tera-Scale Computing. IEEE J Solid-State Circuits 45:235–248 Yuan S, Chen L, Wang Z et al (2021) On-chip terahertz isolator with ultrahigh isolation ratios. Nat Commun 12 Hosseini K, Kok E, Shumarayev SY et al (2022) 5.12 Tbps Co-Packaged FPGA and Silicon Photonics Interconnect I/O , in 2022 IEEE Symposium on VLSI Technology and Circuits (VLSI Technology and Circuits) . pp. 260–261 Feng L, Ayache M, Huang J et al (2011) Nonreciprocal light propagation in a silicon photonic circuit. Science 333, 729 – 33 Yu Z, Fan S (2009) Complete optical isolation created by indirect interband photonic transitions. Nat Photonics 3:91–94 Rizzo A, Novick A, Gopal V et al (2023) Massively scalable Kerr comb-driven silicon photonic link. Nat Photonics 14 Khilwani D, Lee S, Ou C et al (2024) 3D-Integrated, Low Power, High Bandwidth Density Opto-Electronic Transceiver , in 2024 IEEE International Symposium on Circuits and Systems (ISCAS) . pp. 1–5 Lu J, Liu S, Tang Q et al (2015) Multi-wavelength distributed feedback laser array with very high wavelength-spacing precision. Opt Lett 40 Zhang J, Zheng W, Liu A (2023) Polarized Vertical-Cavity Surface-Emitting Laser Arrays With Asymmetrical Structures for Three-Dimensional Sensing With Polarization Imaging. IEE J Quantum Electron 59:1–8 Zhuang R, Ni K, Wu G et al (2023) Electro-Optic Frequency Combs: Theory, Characteristics, and Applications. Laser Photonics Rev 17 Hu Y, Yu M, Buscaino B et al (2022) High-efficiency and broadband on-chip electro-optic frequency comb generators. Nat Photonics 16:679–685 Pfeifle J, Vujicic V, Watts RT et al (2015) Flexible terabit/s Nyquist-WDM super-channels using a gain-switched comb source. Opt Express 23 Weimann C, Schindler PC, Palmer R et al (2014) Silicon-organic hybrid (SOH) frequency comb sources for terabit/s data transmission. Opt Express 22 Lu Z, Liu J, Mao L et al (2019) 12.032 Tbit/s coherent transmission using an ultra-narrow linewidth qantum dot 34.46-GHz C-Band coherent comb laser , in Next-Generation Optical Communication: Components, Sub-Systems, and Systems VIII . p. 109470J Bernal S, Dumont M, Berikaa E et al (2024) 12.1 terabit/second data center interconnects using O-band coherent transmission with QD-MLL frequency combs. Nat Commun 15 Kippenberg TJ, Gaeta AL, Lipson M et al (2018) Dissipative Kerr solitons in optical microresonators. Science 361 Shu H, Chang L, Tao Y et al (2022) Microcomb-driven silicon photonic systems. Nature 605:457–463 Shen B, Chang L, Liu J et al (2020) Integrated turnkey soliton microcombs. Nature 582:365–369 Geng Y, Zhou H, Han X et al (2022) Coherent optical communications using coherence-cloned Kerr soliton microcombs. Nat Commun 13:1070 Picqué N, Hänsch TW (2019) Frequency comb spectroscopy. Nat Photonics 13:146–157 Lukashchuk A, Riemensberger J, Tusnin A et al (2023) Chaotic microcomb-based parallel ranging. Nat Photonics Zhou H, Geng Y, Cui W et al (2019) Soliton bursts and deterministic dissipative Kerr soliton generation in auxiliary-assisted microcavities. Light Sci Appl 8:50 Mazur M, Suh M-G, Fulop A et al (2021) High Spectral Efficiency Coherent Superchannel Transmission With Soliton Microcombs. J Lightwave Technol 39:4367–4373 Corcoran B, Tan M, Xu X et al (2020) Ultra-dense optical data transmission over standard fibre with a single chip source. Nat Commun 11:2568 Zhou X, Urata R, Liu H (2020) Beyond 1 Tb/s Intra-Data Center Interconnect Technology: IM-DD OR Coherent? J Lightwave Technol 38:475–484 Alam MS, Li X, Jacques M et al (2021) Net 300 Gbps/λ Transmission Over 2 km of SMF With a Silicon Photonic Mach-Zehnder Modulator. IEEE Photonics Technol Lett 33:1391–1394 Berikaa E, Alam MS, Li W et al (2023) TFLN MZMs and Next-Gen DACs: Enabling Beyond 400 Gbps IMDD O-Band and C-Band Transmission. IEEE Photonics Technol Lett 35:850–853 Liu Y, Zhang H, Liu J et al (2024) Parallel wavelength-division-multiplexed signal transmission and dispersion compensation enabled by soliton microcombs and microrings. Nat Commun 15 Liu X, Gong Z, Bruch AW et al (2021) Aluminum nitride nanophotonics for beyond-octave soliton microcomb generation and self-referencing. Nat Commun 12 Gu J, Li X, Qi K et al (2023) Octave-spanning soliton microcomb in silica microdisk resonators. Opt Lett 48 Dong B, Dumont M, Terra O et al (2023) Broadband quantum-dot frequency-modulated comb laser. Light: Sci Appl 12 Helgason ÓB, Girardi M, Ye Z et al (2023) Surpassing the nonlinear conversion efficiency of soliton microcombs. Nat Photonics Jang JK, Okawachi Y, Zhao Y et al (2021) Conversion efficiency of soliton Kerr combs. Opt Lett 46:3657–3660 Rahim A, Hermans A, Wohlfeil B et al (2021) Taking silicon photonics modulators to a higher performance level: state-of-the-art and a review of new technologies. Adv Photonics 3 Reed GT, Mashanovich G, Gardes FY et al (2010) Silicon optical modulators. Nat Photonics 4:518–526 PCI-SIG https://pcisig.com/pci-express-6.0-specification Parsons R (2024) Silicon Photonic Integration of DWDM and Mode-Division Multiplexing for Advancing Multi-Dimensional Data Transmission. CLEO Europe and EQEC 2023 Intel I-S-OCI -Optical-I-O-Chiplet-Co-packaged-with-CPU. OFC 2024 Sun C (2024) Photonics for Die-to-Die Interconnects: Links and Optical I/O Chiplets. IEEE International Solid-State Circuits Conference (ISSCC) (2024) Raj M, Silicon Photonics-Based High-Throughput Optical Transceivers (2024). IEEE International Solid-State Circuits Conference (ISSCC)- Forum 6.5 (2024) Wade M, Anderson E, Ardalan S et al (2021) An Error-free 1 Tbps WDM Optical I/O Chiplet and Multi-wavelength Multi-port Laser. Optical Fiber Communication (OFC) Conference 2021 Hosseini K, Kok E, Shumarayev SY et al 8 Tbps Co-Packaged FPGA and Silicon Photonics Optical IO. Optical Fiber Communication (OFC) Conference 2021 Ong KYK, Chia XX, Rahim AA et al (2024) IMDD Data Transmission with Microresonator Soliton Crystals. Optical Fiber Communication (OFC) Conference 2024 Tanikawa K, Tasaka S, Fujii S et al (2022) FEC-free dense WDM transmission with Kerr soliton microcombs in crystalline MgF2 microresonators. Pacific Rim Conference on Lasers and Electro-Optics (CLEO/Pacific Rim) (2022) Ran S, Zhou G, Li Y et al (2024) Micro-Ring Modulators With Integrated Inductor to Mitigate Bandwidth and Extinction Ratio Trade-Off. IEEE Photonics Technol Lett 36:231–234 Dubé-Demers R, LaRochelle S, Shi W (2016) Ultrafast pulse-amplitude modulation with a femtojoule silicon photonic modulator. Optica 3 Shu H, Chang L, Lao C et al (2023) Submilliwatt, widely tunable coherent microcomb generation with feedback-free operation. Adv Photonics 5 Lin Q, Painter OJ, Agrawal GP (2007) Nonlinear optical phenomena in silicon waveguides: modeling and applications. Opt Express 15:16604–16644 Zhang Y, Zhang S, Bi T et al (2023) Geometry optimization for dark soliton combs in thin multimode silicon nitride microresonators. Opt Express 31 Liu J, Huang G, Wang RN et al (2021) High-yield, wafer-scale fabrication of ultralow-loss, dispersion-engineered silicon nitride photonic circuits. Nat Commun 12 Guo Y, Li X, Jin M et al (2022) Hybrid integrated external cavity laser with a 172-nm tuning range. APL Photonics 7:066101 Xiang C, Liu J, Guo J et al (2021) Laser soliton microcombs heterogeneously integrated on silicon. Science 373:99–103 Additional Declarations There is NO Competing Interest. Supplementary Files NCOIOsupp.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-5365298","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":374883035,"identity":"c6c32e4b-ecd5-40fa-b7d2-246df868a8fa","order_by":0,"name":"Liangjun Lu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA8ElEQVRIiWNgGAWjYBACPgbmhgNAmhmIDjBIgMUS8GthY2CEaWFLYJBIIFILlMljAFVNSItEYuOBnztq2Q1u93z+YPnjMAM/e44Bw88deLU0HOw9c5zZ4M7ZbRISCYcZJHveGDD2nsGv5QBv2zFmgxu52xhAWgxu5BgwM7YRsOUvWEvO4w8gLfbEaDnM21YD0sIAdpiBBCEtPA8bDsu2HWCWvJFmJiGRls4jceZZwcFePFr42ZMPf3zbVpfMdyP58WcJG2s5/vbkjQ9+4tECBYeTQSQzMPZ5QIwDBDUwMNTZgUjGD0QoHQWjYBSMgpEHAE7DUhRkjWDKAAAAAElFTkSuQmCC","orcid":"","institution":"Shanghai Jiao Tong University","correspondingAuthor":true,"prefix":"","firstName":"Liangjun","middleName":"","lastName":"Lu","suffix":""},{"id":374883036,"identity":"6f0c5d9a-ee75-4931-9c91-b182b2a145f7","order_by":1,"name":"Hongyi Zhang","email":"","orcid":"","institution":"Shanghai Jiao Tong University","correspondingAuthor":false,"prefix":"","firstName":"Hongyi","middleName":"","lastName":"Zhang","suffix":""},{"id":374883037,"identity":"f4f1b89e-a890-4737-80da-92be39d16a0b","order_by":2,"name":"Shihuan Ran","email":"","orcid":"https://orcid.org/0000-0003-4119-3438","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Shihuan","middleName":"","lastName":"Ran","suffix":""},{"id":374883038,"identity":"38916187-21a4-48ab-9d85-73411f48dba1","order_by":3,"name":"Yuanbin Liu","email":"","orcid":"https://orcid.org/0000-0002-9928-2591","institution":"Shanghai Jiao Tong University","correspondingAuthor":false,"prefix":"","firstName":"Yuanbin","middleName":"","lastName":"Liu","suffix":""},{"id":374883039,"identity":"98d01b34-adec-4403-9364-af0f7272038f","order_by":4,"name":"Shuxiao Wang","email":"","orcid":"","institution":"Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Shuxiao","middleName":"","lastName":"Wang","suffix":""},{"id":374883040,"identity":"59d2f5fb-fd42-49c9-a465-068ec83c9089","order_by":5,"name":"Yan Cai","email":"","orcid":"","institution":"Shanghai Institute of Microsystem and Information Technology","correspondingAuthor":false,"prefix":"","firstName":"Yan","middleName":"","lastName":"Cai","suffix":""},{"id":374883041,"identity":"e4ec0a8a-42cc-4cb2-a261-8b3af093c27b","order_by":6,"name":"Yuyao Guo","email":"","orcid":"","institution":"Shanghai Jiao Tong University","correspondingAuthor":false,"prefix":"","firstName":"Yuyao","middleName":"","lastName":"Guo","suffix":""},{"id":374883042,"identity":"d6d35848-fabe-4606-b114-3eb01ab5da70","order_by":7,"name":"Yu Li","email":"","orcid":"","institution":"Shanghai Jiao Tong University","correspondingAuthor":false,"prefix":"","firstName":"Yu","middleName":"","lastName":"Li","suffix":""},{"id":374883043,"identity":"3335cc21-58a4-49fc-af44-f7f5161c97e4","order_by":8,"name":"Jianping Chen","email":"","orcid":"","institution":"Shanghai Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Jianping","middleName":"","lastName":"Chen","suffix":""},{"id":374883044,"identity":"de09be2d-1866-4462-b5c6-2539cc0f5ef5","order_by":9,"name":"Linjie Zhou","email":"","orcid":"https://orcid.org/0000-0002-2792-2959","institution":"Shanghai Jiao Tong University","correspondingAuthor":false,"prefix":"","firstName":"Linjie","middleName":"","lastName":"Zhou","suffix":""}],"badges":[],"createdAt":"2024-10-31 06:45:17","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-5365298/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-5365298/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":68423037,"identity":"20972420-cb6a-46ee-996d-f94a11d669fc","added_by":"auto","created_at":"2024-11-07 06:29:02","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":140140,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eHigh-capacity optical I/O driven by efficient dark soliton microcomb.\u003c/strong\u003e (A) Optical I/O is implemented for the high-speed interconnects between the XPUs with large bandwidths, low latency, and low consumption. The SOI platform drives the optical I/O system. The microring for the efficient comb source is based on the foundry-compatible 400-nm-thick Si\u003csub\u003e3\u003c/sub\u003eN\u003csub\u003e4\u003c/sub\u003e platform The ring modulator and photodetector (PD) array are based on the commercial silicon photonic platform. The comb line is modulated with an OOK signal from XPU for low-error transmission and the PD at the receiver transmits the signal to another XPU, thereby achieving interconnection. The interconnect between the XPUs is through optical fiber. (B) Comparison of error-free optical interconnects. 'free error' is characterized by a bit error rate (BER) less than 10-12. It is noted that C.C. refers to Columbia University; Keio refers to Keio University; SUTD refers to Singapore University of Technology and Design; SJTU refers to Shanghai Jiao Tong University.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-5365298/v1/9b89fbbe1e3a782d937f9652.png"},{"id":68423384,"identity":"cc592a8f-dcb4-4789-a5b6-6ae70ceba565","added_by":"auto","created_at":"2024-11-07 06:37:03","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":256904,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSi\u003c/strong\u003e\u003csub\u003e\u003cstrong\u003e3\u003c/strong\u003e\u003c/sub\u003e\u003cstrong\u003eN\u003c/strong\u003e\u003csub\u003e\u003cstrong\u003e4\u003c/strong\u003e\u003c/sub\u003e\u003cstrong\u003e microring and soliton characterization.\u003c/strong\u003e (A) Dispersion and dissipation engineering for obtaining the microcomb source with broad bandwidth and high optical power. (B) Left: photograph of the packaged Si\u003csub\u003e3\u003c/sub\u003eN\u003csub\u003e4\u003c/sub\u003e microchip. Right: microscope image of the fabricated microring. (C) Measured and fitted resonance trace of the pump resonance. (D) Integrated dispersion of the TE0 mode. Inset: the pump mode on the AMXs. (E) Recorded spectrum of the generated dark soliton microcomb. CE: conversion efficiency.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-5365298/v1/5b09c08839dba2961963f7a0.png"},{"id":68423039,"identity":"5a644ada-d93d-48da-a8b1-b6bacab1b942","added_by":"auto","created_at":"2024-11-07 06:29:03","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":265777,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSilicon MRM and experimental characterization.\u003c/strong\u003e (A) Schematic structure of the MRM. (B) Microscope image of the MRM. (C) Thermal-optical response versus the heater power. The Pπ is 26 mW. Inset: transmission spectra under various heater power. (D) Measured transmission spectra under various reverse bias voltages for the MRM. (E) Normalized EO response of the MRM upon 3 dB and 6 dB insertion loss (IL) operation points. (F) Measured real and imaginary parts of the S11 parameters. (G) Measured eye diagram of the MRM when the CW laser is modulated with 50 Gbit/s, 80 Gbit/s, and 100 Gbit/s OOK signal, respectively.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-5365298/v1/94b87d456368f1fd389754d7.png"},{"id":68423038,"identity":"f01abe7a-01f3-4dcb-bdec-fe4e4150b317","added_by":"auto","created_at":"2024-11-07 06:29:03","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":207649,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eError-free WDM parallel data transmission driven by the microcomb and the silicon MRM.\u003c/strong\u003e (A) Experimental setup for the dark soliton microcomb-based parallel data transmission. DCA: digital communication analyzer; BF: bandpass filter; AWG: Arbitrary Waveform Generator; EDFA: Erbium-Doped Fiber Amplifier. (B) Superimposed optical spectrum after each line is filtered out and modulated. (C) Measured BERs of 36 channels with a received optical power of 2 dBm, each line is modulated with 64 Gbit/s OOK signal. Four of the comb lines are added with FFE for free error transmission. (D) The measured 64 Gbit/s OOK eye diagrams of the 36 channels with 2 dBm received optical power. (E) BER curves versus received optical power using Ch.4, Ch.9, and Ch.21 of the dark soliton and the CW laser as the sources with 64 Gbit/s OOK signal. (F)-(H) BER curves versus received optical power for (F) Ch. 9, (G) Ch. 18, and (H) Ch. 27 when the line is modulated with 80 Gbit/s and 100 Gbit/s OOK signal. FFE was added to lower the BER.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-5365298/v1/23d28635938330c340ead720.png"},{"id":71171802,"identity":"330865bd-8d59-459d-aa9c-ae4e2d4bc679","added_by":"auto","created_at":"2024-12-11 20:03:21","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1336422,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5365298/v1/783a5dbc-ec71-400d-a22a-5d9a04f9aa1e.pdf"},{"id":68423040,"identity":"642796a7-5ddc-44db-af21-79bc398ebe3c","added_by":"auto","created_at":"2024-11-07 06:29:03","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":1669272,"visible":true,"origin":"","legend":"","description":"","filename":"NCOIOsupp.docx","url":"https://assets-eu.researchsquare.com/files/rs-5365298/v1/ce45a6853686f313998509c7.docx"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.","formattedTitle":"High-capacity data transmission for optical I/O driven by efficient microcomb and microring modulator","fulltext":[{"header":"Introduction","content":"\u003cp\u003eAs the accuracy and comprehensiveness demands of artificial intelligence (AI) applications escalate, the size of clusters for large language models correspondingly expands [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. For instance, for contemporary large model training, such as those conducted by OpenAI, Meta, and Inflection AI, etc, clusters comprising over 10,000 GPUs have been utilized to train and process models with more than a trillion parameters [\u003cspan additionalcitationids=\"CR3 CR4 CR5\" citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. Although server hardware floating-point operations per second (Flops) have exhibited a tripling approximately every two years over the past two decades, the growth rates of dynamic random-access memory (DRAM) and interconnect bandwidth are only 1.6\u0026times; and 1.4\u0026times; every two years, respectively [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. It is anticipated that as cluster sizes continue to grow, their performance will increasingly be constrained by interconnect bandwidth, making enhancements in this area both critical and urgent [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Furthermore, the power consumption of interconnects within GPU clusters is a crucial consideration. As cluster size expands, the energy consumption of network devices and interconnections facilitating GPU communication also escalates, potentially affecting not only energy efficiency but also overall operational costs.\u003c/p\u003e \u003cp\u003eHigh-capacity inter-chip communication with more than TB/s has emerged as a critical requirement in technological advancements. Traditional XPU-to-XPU electronic input/output (I/O) is increasingly proving inadequate in meeting the heightened requirements for high-speed interconnects and reduced power consumption [\u003cspan additionalcitationids=\"CR10\" citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Optical I/O, possessing the capability for terahertz bandwidth with minimal loss and crosstalk, has been advanced as a superior alternative to traditional electrical interconnects for XPU-to-XPU communication [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. This innovation not only supports substantial increases in bandwidth and reductions in latency but also achieves remarkable improvements in energy efficiency [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. Compared to traditional intra- and inter-datacenter optical interconnects, optical I/O between XPUs requires significantly lower bit error rates (BERs), higher I/O bandwidth density, lower power consumption, even to the scale of a chip's dimensions, and little or no equalization [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIntensity-modulation, direct-detection (IM/DD) is favored for the optical interconnects between the XPUs, due to its simplicity, power efficiency, and cost-effectiveness. Error-free transmission is critical to ensure high data integrity, which can otherwise degrade performance and efficiency. By leveraging IM/DD, the system can achieve this reliability without the need for complex Forward Error Correction (FEC) algorithms that rely heavily on Digital Signal Processing (DSP). This reduction in FEC requirements directly lowers both power consumption and latency. Error-free transmission also enhances system performance and simplifies network design, especially for low-latency, high-throughput communication. Since the single lane data rate of the IM/DD scheme is much lower than the coherent scheme, wavelength division multiplexing (WDM) technology is commonly employed to further expand the transmission capacity. Therefore, optical I/O architectures typically incorporate multi-wavelength laser sources, modulators, (de)multiplexers, and photodetectors (PDs) to facilitate data transmission and reception [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. These devices have been implemented on a silicon photonics platform either monolithically or as part of the system due to the inherent advantages of complementary metal oxide semiconductor (CMOS) compatibility, high integration density, low loss, and high bandwidth [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Therefore, we believe the integration of optical I/O on the silicon photonics platform has the potential to unlock unprecedented performance enhancements.\u003c/p\u003e \u003cp\u003eFor the multi-wavelength laser source, several solutions have been proposed for WDM parallel data transmission. Laser arrays, such as distributed feedback laser (DFB laser) [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] and vertical-cavity surface-emitting laser (VCSEL) [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e] arrays, etc, have already been commercially implemented in the coherent or IMDD transceiver for data center interconnects. However, for the WDM transmission, the employment of multiple carriers requires independent control of each laser, leading to increased system complexity and footprint as the number of lasers scales up. Based on the electro-optic (EO) modulation, the EO comb is constructed with versatility, excellent comb stability, and phase coherence through cascading intensity and phase modulators [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. On-chip EO comb based on thin-film lithium niobate (TFLN) features a conversion efficiency of 30% and an optical bandwidth of 132 nm [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. Harnessing the full potential of EO combs has enabled remarkable advancements in data transmission, achieving rates exceeding several Tb/s and extending the reach to hundreds of kilometers [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. Nevertheless, the EO combs are constrained by the bandwidth of the modulators, making it challenging to achieve larger comb line spacing, which in turn further restricts the line rate of the single carrier. Quantum dot-mode locked laser (QD-MLL) combs have demonstrated their advantages as WDM sources due to their compact, efficient, robust, and dispersion-insensitive characteristics. Previous works have reported more than 10 Tb/s WDM transmission based on the QD-MLL comb [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. Soliton microcombs, generated in the high-nonlinearity microring, can easily realize dozens of phase-coherent and evenly spaced optical frequency lines with low noise and 10\u0026rsquo;s GHz\u0026thinsp;~\u0026thinsp;1 THz comb spacings [\u003cspan additionalcitationids=\"CR26 CR27 CR28 CR29 CR30 CR31 CR32\" citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]. It is highly desired to expand the communication capacity of optical I/O [\u003cspan additionalcitationids=\"CR35 CR36\" citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e]. Through dispersion engineering, soliton microcombs are capable of generating significantly broad spectral bandwidths that encompass the entire O, C, and L bands [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e, \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e], greatly exceeding the spectral bandwidth of the QD-MLL comb [\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e, \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e]. In the previous research, dissipative Kerr soliton (DKS) generated in the anomalous dispersion regime has been extensively studied. The single soliton, a popular type of DKSs, exhibits a smooth Sech2-shaped spectrum but has a low pump-to-comb conversion efficiency of only\u0026thinsp;~\u0026thinsp;1% [\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e]. Therefore, the comb lines of the single soliton microcomb exhibit very low line power, which will reduce the signal-to-noise ratio (SNR) and lower the data rate for IM/DD data transmission. In contrast, dark solitons generated in the normal dispersion regime, show the merits of high conversion efficiency and robustness to pump laser tuning. High-power comb lines supported by the efficient dark soliton can ensure high-performance data transmission, maintaining low BER or error-free operation when carrying high-rate signals, which is crucial for optical I/O interconnects [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e].\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eSilicon microring modulators (MRMs) are favored for high-density optical I/O due to their compactness, high modulation efficiency, low power consumption, and wavelength selectivity, which are suitable for comb-based parallel data transmission [\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e, \u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e]. In this work, we propose and demonstrate a high-capacity error-free data transmission driven by an efficient dark soliton microcomb and silicon MRMs for the optical I/O application. The conceptual diagram is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eA. Microring-based comb source and modulators are implemented for compact chip size and low power consumption. The dark soliton comb source is constructed using a 400-nm-thick Si\u003csub\u003e3\u003c/sub\u003eN\u003csub\u003e4\u003c/sub\u003e microring with the assistance of avoided mode crosses (AMXs). With optimized power coupling ratio and group velocity dispersion (GVD) of the Si\u003csub\u003e3\u003c/sub\u003eN\u003csub\u003e4\u003c/sub\u003e microring, we generate a dark soliton with a record high conversion efficiency of close to 50% and an on-chip \u0026minus;\u0026thinsp;5-dBm bandwidth of 28 nm. We further use the 36 comb lines in the \u0026minus;\u0026thinsp;5-dBm bandwidth as the optical carriers for data transmission, which are modulated by a silicon MRM with an electro-optic bandwidth of 61.7 GHz. The line rate per channel reaches 64 Gb/s with on-off keying (OOK) modulation, meeting the data rate requirements in the next-generation I/O standard PCIe 6.0 [\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e]. A total 2.3 Tbit/s OOK signal error-free (BER\u0026thinsp;\u0026lt;\u0026thinsp;10\u0026ndash;12) transmission is successfully realized. Among the 36 comb lines, 32 comb lines\u0026rsquo; BER is measured with the original signal, and only 4 comb lines\u0026rsquo; BER is measured after the use of feed-forward equalization (FFE). As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eB, in comparison with state-of-the-art error-free optical interconnects [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan additionalcitationids=\"CR47 CR48 CR49 CR50 CR51 CR52\" citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e], our work achieves remarkable advancement in both the total capacity per fiber port of 2.3 Tb/s and the lane rate per wavelength channel of 64 Gb/s, benefiting from our highly-efficient multi-wavelength laser source and high-speed silicon MRMs. Our proposed error-free parallel data transmission scheme is highly promising for high-capacity low-latency and power-efficient optical I/O applications.\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eDark soliton generation\u003c/h2\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTo create a high-performance WDM comb source, we focus on two aspects. Firstly, the microcomb is expected to have a high conversion efficiency for supporting high comb line power. Secondly, it is essential to have a relatively flat and wide spectrum so that more high-power comb lines are available. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eA shows the dispersion and dissipation engineering for constructing the microcomb source with both broad bandwidth and high optical power. We used the Lugiato-Lefever equation (LLE) for the theoretical simulation of the dark soliton comb source [\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e] and the waveguide structure design is based on the simulation from Lumerical FDTD and Mode solution. The details about the engineering are discussed in the supplementary information. According to the simulation, the waveguide dimension of the Si\u003csub\u003e3\u003c/sub\u003eN\u003csub\u003e4\u003c/sub\u003e microring is designed to be 0.4 \u0026micro;m (height) \u0026times; 2 \u0026micro;m (width) with a simulated \u0026#120573;2 (GVD) of 560 ps2/km, supporting multiple modes for AMX and low transmission loss simultaneously. The width of the straight bus waveguide is designed to be 1.5 \u0026micro;m, so that both the fundamental (TE0) and the first-order transverse electric (TE1) modes can be exited in the microring in order to trigger mode coupling for the dark soliton generation. The gap between the straight bus waveguide and the microring is 0.5 \u0026micro;m. The radius of the microring is 246 \u0026micro;m, corresponding to an FSR of 96.2 GHz. The Si\u003csub\u003e3\u003c/sub\u003eN\u003csub\u003e4\u003c/sub\u003e microring was fabricated on an 8-inch wafer using 193-nm photolithography by Shanghai Industrial \u0026micro;Technology Research Institute (SITRI). Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eB shows the packaged chip with a pair of fiber arrays, facilitating stable optical coupling between the fiber and the chip in the experiments. The right side of Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eB shows the microscope image of the fabricated microring resonator.\u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eC shows the measured and fitted spectra at the pump resonance. The extracted power coupling ratio is 0.011, and the loss is about 0.04 dB/cm. The loaded quality (Q) factor is approximately 1\u0026times;106. By extracting the resonance frequency of each longitudinal mode (\u0026#120596;\u0026#120583;), we calculated the integrated dispersion curve \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{D}_{int}\\:=\\:{\\omega\\:}_{\\mu\\:}\\:-\\:{\\omega\\:}_{o}\\:-\\:{D}_{1}\\mu\\:\\)\u003c/span\u003e\u003c/span\u003e, where \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\omega\\:}_{o}\\)\u003c/span\u003e\u003c/span\u003e is the resonance frequency of the center mode, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\mu\\:\\)\u003c/span\u003e\u003c/span\u003e is defined as the mode index in terms of the center mode (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\mu\\:=0\\)\u003c/span\u003e\u003c/span\u003e), and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{D}_{1}/2\\pi\\:\\)\u003c/span\u003e\u003c/span\u003e is the equidistant FSR, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eD. The second-order dispersion (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{D}_{2}/2\\pi\\:\\)\u003c/span\u003e\u003c/span\u003e) is -5.2 MHz, corresponding to \u0026#120573;2 of 608.6 ps2 /km, which is in the optimal region for large \u0026minus;\u0026thinsp;5-dBm optical bandwidth. We used a continuous-wave (CW) laser to pump the Si\u003csub\u003e3\u003c/sub\u003eN\u003csub\u003e4\u003c/sub\u003e microring, which was amplified to 29 dBm by an erbium-doped fiber amplifier (EDFA). The spectrum of the generated dark soliton microcomb is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eE. The measured conversion efficiency and on-chip \u0026minus;\u0026thinsp;5-dBm (subtracting 5 dB coupling loss from the optical fiber) bandwidth of the microcomb are 49% and 28 nm, respectively, which are consistent with the simulation. There are 36 comb lines in the on-chip \u0026minus;\u0026thinsp;5-dBm bandwidth and nearly 88.9% of them have a power variation of less than 8 dB (-10 dBm to -2 dBm), indicating the dark soliton features a flat spectrum. When the pump power is set at various powers, the formation of dark soliton is different, which is discussed in supplementary information. We also explore the stability and frequency noise of dark soliton in the supplementary information.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eLarge-bandwidth silicon microring modulator\u003c/h3\u003e\n\u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA illustrates the schematic diagram of the designed MRM. The microring has a radius of 8 \u0026micro;m, a rib waveguide height of 220 nm, a width of 500 nm, and a slab thickness of 90 nm. A horizontal PN junction is embedded in the microring waveguide for modulation. The gap between the microring and the bus waveguide is 250 nm, with a coupling region length of 1.8 \u0026micro;m. The doping concentrations of both P-type and N-type are 2\u0026times;1018 cm\u0026thinsp;\u0026minus;\u0026thinsp;3 in the light doping regions and 1\u0026times;1020 cm\u0026thinsp;\u0026minus;\u0026thinsp;3 in the heavy doping regions. The chip was fabricated on the SOI platform using a standard foundry process. The microscope image of the chip is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eB.\u003c/p\u003e \u003cp\u003eTo enable flexible adjustment of the MRM's operating wavelength, a titanium nitride-based thermal phase shifter is integrated on top of the microring modulator. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eC, for the MRM with a measured FSR of 11.39 nm, a thermo-optic tuning efficiency of 0.22 nm/mW (Pπ\u0026thinsp;=\u0026thinsp;26.46 mW) is achieved. Specifically, at the thermal tuning power of 55.13 mW, the resonance wavelength redshifts by 11.86 nm, surpassing one FSR. Thus, the MRM's operating wavelength can be tailored to any desired wavelength by precise control of the heating power. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eD shows the transmission spectra of the MRR at various bias voltages. At 0 V bias, the MRM has a resonance wavelength of 1545.1 nm with a Q-factor of approximately 2870. The resonance wavelength redshifts upon increased reverse bias voltage, with a modulation efficiency of 25 pm/V. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eE displays the EO response of the microring modulator. Upon \u0026minus;\u0026thinsp;3 V bias, the EO bandwidth at the 3 dB IL operation point is 61.7 GHz and 38.6 GHz at the 6 dB IL operation point. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eF presents the real and imaginary parts of the S11 parameters under the \u0026minus;\u0026thinsp;3 V bias, with the PN junction capacitance extracted as 21.85 fF based on the model fitting in [\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eFor high-speed modulation, we implemented 50 Gbit/s, 80 Gbit/s, and 100 Gbit/s OOK signal on the MRR with drive voltages of 2.1 Vpp, 2.8 Vpp, and 2.0 Vpp, respectively, at the 6 dB IL operation point for larger extinction ratio. A commercial CW laser was employed as the light source with an output power of 13 dBm. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eG exhibits the measured eye diagrams. The 50 Gbit/s and 80 Gbit/s OOK modulation achieved BERs of less than 1\u0026times;10\u0026thinsp;\u0026minus;\u0026thinsp;18 for error-free transmission without any DSP, while the 100 Gbit/s OOK modulation attained a BER of 2\u0026times;10\u0026thinsp;\u0026minus;\u0026thinsp;13 with 7-tap feed-forward equalization. The power consumption for OOK at the three data rates was calculated to be 17.8 fJ/bit, 31.6 fJ/bit, and 16.1 fJ/bit, respectively, using the formulae in [\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e]. These results underscore the capability of the MRM to achieve error-free high-speed OOK signal transmission with low power consumption, positioning it as an excellent candidate for optical I/O applications.\u003c/p\u003e\n\u003ch3\u003eError-free parallel optical data transmission\u003c/h3\u003e\n\u003cp\u003eWe implemented the error-free transmission based on the efficient dark soliton microcomb and the silicon MRM. The experimental setup is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eA. An amplified CW laser was coupled to the Si\u003csub\u003e3\u003c/sub\u003eN\u003csub\u003e4\u003c/sub\u003e chip before being adjusted to the fundamental TE mode by a polarization controller. An additional bandpass filter with a 3-dB bandwidth of 75 GHz was employed to eliminate the amplifier ASE noise from the EDFA before injecting the pump light into the Si\u003csub\u003e3\u003c/sub\u003eN\u003csub\u003e4\u003c/sub\u003e microring for comb generation. Then, we filtered out each of the comb lines within the on-chip \u0026minus;\u0026thinsp;5-dBm bandwidth and sequentially encoded them with 64 Gbit/s OOK by the silicon MRM. Note that within on-chip \u0026minus;\u0026thinsp;5-dBm bandwidth, a total of 36 comb lines are present. We selected the first 35 comb lines and excluded the last one (1572.8 nm), limited by the effective amplification bandwidth of the EDFA. Instead, we opted for an adjacent comb line located at the shorter wavelength end of the on-chip \u0026minus;\u0026thinsp;5-dBm wavelength range, which has a power slightly below \u0026minus;\u0026thinsp;10 dBm. As a result, the total number of carriers remains at 36 and the combined modulated optical spectra are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eB. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eC reveals the recorded BER for each channel upon a received optical power of 2 dBm. Owning to the limitations of the DCA, the minimum detectable BER is constrained to 10\u0026ndash;18. Consequently, all BER below this threshold are recorded as 1\u0026times;10\u0026ndash;18 in the results. As evident from the graph, in Ch. 9 to Ch. 25, the BERs reach 1\u0026times;10\u0026ndash;18 except for the pump channel (Ch. 17). To mitigate the residual ASE noise from the EDFA, a 7-tap FFE is applied to reduce the BER of the pump channel from 10\u0026thinsp;\u0026minus;\u0026thinsp;8 to 2\u0026times;10\u0026ndash;13 for error-free operation.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAs noted, the channels distant from the pump mode are subjected to increased noise due to thermal noise-induced repetition rate instability [\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e]. Since the effective amplification band of EDFA is in the C-band, more ASE noise was introduced at both edges of the comb spectrum. Consequently, the channels on the two sides exhibited larger BER. For Ch. 1, Ch. 2, and Ch. 36, the 64-tap, 7-tap, and 15-tap FFE were implemented to minimize the corresponding BER into the error-free regime. We achieve 2.3 Tbit/s (36\u0026times; 64 Gbit/s) error-free transmission and 2 Tbit/s (32\u0026times;64 Gbit/s) with both error-free and FFE-free operation. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eD shows the measured eye diagram for all 36 channels, the SNR for nearly all channels exceeds 6.5 dB which exhibits good and consistent results. Furthermore, Ch. 4, Ch. 9, and Ch. 21 were selected to measure the BER versus received optical power. For comparative analysis, we utilize CW lasers with identical wavelengths and power levels as these chosen comb lines for OOK modulation. At each received optical power, the CW laser-based transmission link exhibits slightly lower BER compared to the comb channel. The average power penalty is less than 1 dB at the BER of 10\u0026ndash;14, demonstrating a comparable performance between these two sources. Besides, the measured results demonstrate the BER can be reduced to less than 10\u0026ndash;12 when the received power exceeds \u0026minus;\u0026thinsp;1 dBm. Consequently, we can decrease the previously mentioned received power from 2 dBm while still maintaining error-free operation. Furthermore, to investigate the capability of the efficient dark microcomb for carrying higher-speed signals, three comb lines were selected and modulated with 80 Gbit/s and 100 Gbit/s OOK signals. Figures\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eF-H show the BER curves of 80 Gbit/s and 100 Gbit/s OOK modulation versus the received optical power for Ch. 9, Ch.18, and Ch. 27. For 80 Gbit/s OOK modulation, after the 11-tap FFE, the BER of all the chosen three channels can achieve error-free operation, even reaching the lower boundary of 1\u0026times;10\u0026ndash;18. For 100 Gbit/s OOK modulation, the lowest BER is around 1\u0026times;10\u0026ndash;10 after the application of 64-tap FFE. When loading higher-rate signals, there are increased requirements on the carrier itself, such as relative intensity noise (RIN). By employing a pump source with lower RIN and mitigating the link loss to reduce the use of EDFA, error-free transmission of the high-speed signal can be maintained.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this proof-of-concept experimental demonstration, mounts of microcomb lines were loaded with a modulation rate of 64 Gbit/s to achieve DSP-free error-free transmission for high-capacity optical I/O. Both the single line rate and the total line rate significantly exceeded the previously demonstrated error-free transmission results. The comb spacing of the microcomb is ~\u0026thinsp;96 GHz, and the spectral efficiency yields 0.66 bit/s/Hz which is close to the limit of 0.75 bit/s/Hz for the comb-based OOK error-free transmission due to the crosstalk between adjacent channels [\u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e57\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eCompared with Mach-Zehnder modulators (MZMs), MRMs with their low power consumption and compactness, are highly suitable for optical I/O applications. In our demonstration, we have utilized silicon MRMs, fabricated at the wafer scale by commercial foundries, facilitating the requirements of a compact size and extremely low driving power. A modulation density of 1.06 Tbps/mm2 and an energy efficiency of 17.8 fJ/bit was achieved with our silicon MRMs. For the microcombs, with the optimized coupling and dispersion, we have developed a comb source with both high conversion efficiency and broad bandwidth, satisfying the power requirement for parallel data transmission. The entire design and optimization flow show a simple and feasible path to creating the high-performance on-chip WDM comb source for optical I/O.\u003c/p\u003e \u003cp\u003eAdditionally, it is worth noting that our microcomb source is based on a standard low-loss low-pressure chemical vapor deposition (LPCVD) silicon nitride platform with a thickness of 400 nm, a specification that has been standardized across various commercial silicon photonic foundries for wafer-level fabrication. Compared to the fabrication process of low-loss thick (\u0026gt;\u0026thinsp;600 nm) silicon nitride films, its fabrication process is simpler, with higher yields and lower costs [\u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e58\u003c/span\u003e, \u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e59\u003c/span\u003e]. Therefore, the hybrid integration of these two foundry-compatible on-chip devices can promote the large-scale manufacturing of microcomb-driven optical I/O chips [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e, \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e], resulting in reduced size, weight, power, and cost. Moreover, with the development of heterogenous or hybrid integration technologies, tunable lasers and optical amplifiers are possible to be integrated and make it possible for the manufacturing of the chip-scale microcomb source [\u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e60\u003c/span\u003e, \u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e61\u003c/span\u003e].\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eWe have successfully implemented an error-free WDM transmission for optical I/O systems, utilizing an efficient dark soliton microcomb and a high-speed MRM. Through meticulous optimization of the coupling and dispersion properties of a Si\u003csub\u003e3\u003c/sub\u003eN\u003csub\u003e4\u003c/sub\u003e microring, we effectively increased the number of high-power comb lines while maintaining a high conversion efficiency. This led to the generation of a flat-power dark soliton microcomb with a conversion efficiency nearing 50% and an on-chip optical bandwidth of 28 nm at -5 dBm, serving as a high-performance multi-wavelength light source. The compact silicon MRM exhibited a high EO bandwidth of 61.7 GHz and operated with a low drive voltage of 2.1 V, enabling high-speed signal modulation while maintaining low power consumption. Capitalizing on these advancements, we demonstrated a record total capacity of 2.3 Tbps per fiber port by modulating 36 comb lines with 64 Gbit/s OOK signals\u0026mdash;comprising 32 channels operating without FFE and 4 channels utilizing 7-tap FFE. Additionally, we showed that a single comb line from the dark soliton microcomb is capable of supporting error-free transmission at 80 Gbit/s and low-error transmission at 100 Gbit/s. Our findings highlight the tremendous potential of microcomb-based optical I/O in achieving high-capacity inter-chip interconnects. The integration of efficient dark soliton microcombs with high-speed silicon MRMs not only demonstrates the feasibility of scaling data transmission rates but also paves the way for energy-efficient, high-bandwidth optical communication systems. This work lays a solid foundation for the development of next-generation optical interconnects, addressing the escalating demands of data centers and high-performance computing environments.\u003c/p\u003e "},{"header":"Methods","content":"\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e \u003ch2\u003eNumerical simulation\u003c/h2\u003e \u003cp\u003eWe used the LLE equations to simulate the dark soliton microcomb source for the design of broad high-power bandwidth and the equations are expressed as follows:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:{T}_{R}\\frac{\\partial\\:{E}_{p}\\left(t,\\tau\\:\\right)}{\\partial\\:t}=\\left(-{\\alpha\\:}_{p}-i{\\delta\\:}_{p}-iL\\frac{{\\beta\\:}_{2}}{2}\\left(\\frac{\\partial\\:}{\\partial\\:\\tau\\:}\\right)+i\\gamma\\:L{\\left|{E}_{p}\\right|}^{2}\\right){E}_{p}\\left(t,\\tau\\:\\right)+\\sqrt{\\theta\\:}{E}_{p}^{in}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\:{E}_{out}={E}_{in}-\\sqrt{\\theta\\:}{E}_{p}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{E}_{p}\\)\u003c/span\u003e\u003c/span\u003e is the intracavity optical field, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{E}_{in}\\)\u003c/span\u003e\u003c/span\u003e is the input optical field and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{E}_{out}\\)\u003c/span\u003e\u003c/span\u003e is the output optical field at the through port while \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:t\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\tau\\:\\)\u003c/span\u003e\u003c/span\u003e represent the fast time and slow time. In the simulation, the pump power (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\left|{E}_{p}\\right|}^{2}\\)\u003c/span\u003e\u003c/span\u003e) is set to be 200 mW, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\alpha\\:}_{p}\\)\u003c/span\u003e\u003c/span\u003e=0.00615 is the cavity loss, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\delta\\:}_{p}\\)\u003c/span\u003e\u003c/span\u003e is the normalized absolute detuning and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\theta\\:\\)\u003c/span\u003e\u003c/span\u003e=0.011 is the power coupling ratio between the bus and microring waveguides. \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:L\\)\u003c/span\u003e\u003c/span\u003e = 246\u0026times;2π \u0026micro;m is the cavity length which determines the comb spacing of the microring and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\gamma\\:\\)\u003c/span\u003e\u003c/span\u003e = 1.22 m2W-1 is the nonlinear refractive index. We added an AMX-induced resonance shift of 200 MHz at the pump mode. Meanwhile, a weak noise is induced as the initial field to seed the intracavity field. The detuning is increased gradually as the intracavity field goes through the roundtrips while the total number of roundtrips is 200000.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e\n\u003ch3\u003eChip fabrication\u003c/h3\u003e\n\u003cp\u003eThe Si\u003csub\u003e3\u003c/sub\u003eN\u003csub\u003e4\u003c/sub\u003e chip was fabricated by Shanghai Industrial \u0026micro;Technology Research Institute (SITRI) using an 8-inch Si wafer. The Si\u003csub\u003e3\u003c/sub\u003eN\u003csub\u003e4\u003c/sub\u003e thin film with a thickness of 400 nm was deposited based on the LPCVD process. The waveguides are patterned using 193-nm photolithography. The thicknesses of the bottom and top SiO2 claddings are 3.18 \u0026micro;m and 3.7 \u0026micro;m.\u003c/p\u003e \u003cp\u003eThe MRM was fabricated on an 8-inch SOI wafer with a top silicon layer thickness of 220 nm by Advanced Micro Foundry (AMF) using the standard fabrication processes. After the waveguide etching process, the modulation region was doped, with boron as the p-type dopant and phosphorus as the n-type dopant. Subsequently, silicon dioxide was grown, followed by the formation of contact openings and the deposition of metal electrodes and the heater.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eMethod for dark soliton generation\u003c/h2\u003e \u003cp\u003eWe put the Si\u003csub\u003e3\u003c/sub\u003eN\u003csub\u003e4\u003c/sub\u003e chip on a thermo-electric cooler (TEC) and used a commercial wavelength-tunable laser (Santec TSL-710) as the pump laser source. The pump laser wavelength was initially set at a fixed wavelength on the blue-detuned side of the resonance and tuned with a step of 0.05 pm. At the output, we used a notch filter to filter out the pump for measuring the comb power. The comb power was recorded after each pump tuning, and the spectrum was recorded by an optical spectrum analyzer (Yokogawa, AQ6370D) every ten detunings. For the measurement of the low-frequency noise, the comb line that locates 2-comb spacing away from the pump on the left side was filtered out. We used a variable optical attenuator (VOA) to adjust the line power to around 1.8 dBm and sent it to the PD with a bandwidth of 50 GHz. The RF spectrum between 0 and 2 GHz was recorded by an electrical signal analyzer (Keysight, N9000B), indicating the results of low-frequency noise.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eMethod for error-free parallel data transmission\u003c/h2\u003e \u003cp\u003eFor the error-free parallel data transmission, we used a bandpass filter (Santec, OTF-350) to filter out individual comb lines and encoded each carrier with 64 Gbit/s OOK by the MRM, respectively. The high-speed OOK was generated by an arbitrary waveform generator (AWG, Keysight M8199A). At the receiver, due to the loss of the MRM, another EDFA (Amonics AEDFA-23-B-FA) was used before being sent to a digital communications analyzer (DCA, Keysight DCA-X N1000A) to measure the eye diagram and BER. Another bandpass filter with a 3-dB bandwidth of 85 GHz (II-VI, WS-04000B) was used before the PD for simulating demultiplexing. A VOA was used before the PD to adjust the received optical power. For the MRM, a thermal phase shifter was employed to shift the micro-ring resonance which is closest to the selected comb line wavelength. This adjustment ensures that the comb line wavelength is positioned at the 6-dB IL operation point of the resonance.\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe thank SJTU-Pinghu Institute of Intelligent Optoelectronics for their assistance with the high-speed transmission experiments, and chip packaging. Besides, we thank SITRI for the fabrication of the Si\u003csub\u003e3\u003c/sub\u003eN\u003csub\u003e4\u003c/sub\u003e chip and AMF for the fabrication of the SOI chip.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding:\u003c/strong\u003e We acknowledge the funding from the National Natural Science Foundation of China (62120106010, 62090052, 62135010, 62075128), and the Shanghai Science and Technology Committee Rising-Star Program (23QA1404500).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions:\u003c/strong\u003e H.Z. designed, simulated, and characterized the efficient dark soliton microcomb source. S.R. designed, simulated, and characterized the microring modulator. H.Z. and S.R. conceived the link architecture and performed the high-speed comb-based data transmission experiments. S.W. and Y.C. supported the fabrication of the Si\u003csub\u003e3\u003c/sub\u003eN\u003csub\u003e4\u003c/sub\u003e chip. All authors helped analyze the data. L.L., Y.Li, Y.Liu, H.Z., and R.S. prepared the manuscript. H.Z., R.S., L.L., Y.Li, Y.G., and L.Z. revised the manuscript. J.C., Y.Li., L.L., and L.Z. provided suggestions and feedback during the revisions. L.L. and Y.L. co-supervised the research.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests:\u003c/strong\u003e The authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData and materials availability:\u003c/strong\u003e All data needed to evaluate the conclusions in the paper are present in the paper and/or the Supplementary Materials. Additional data related to this paper may be requested from the authors.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eNaffziger S (2023) Innovations For Energy Efficient Generative AI, in \u003cem\u003e2023 International Electron Devices Meeting (IEDM)\u003c/em\u003e, 1\u0026ndash;4\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOpenai \u003cem\u003eUpdate: ChatGPT runs 10K Nvidia training GPUs with potential for thousands more\u003c/em\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.fierceelectronics.com/sensors/chatgpt-runs-10k-nvidia-training-gpus-potential-thousands-more\u003c/span\u003e\u003cspan address=\"https://www.fierceelectronics.com/sensors/chatgpt-runs-10k-nvidia-training-gpus-potential-thousands-more\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMeta \u003cem\u003eBuilding Meta\u0026rsquo;s GenAI Infrastructure https\u003c/em\u003e\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e://engineering.fb.com/2024/03/12/data-center-engineering/building-metas-genai-infrastructure/\u003c/span\u003e\u003cspan address=\"http://://engineering.fb.com/2024/03/12/data-center-engineering/building-metas-genai-infrastructure/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eI. AI. \u003cem\u003eAnnouncing our collaboration with NVIDIA and CoreWeave on MLPerf\u003c/em\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://inflection.ai/blog/nvidia-coreweave-mlperf\u003c/span\u003e\u003cspan address=\"https://inflection.ai/blog/nvidia-coreweave-mlperf\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eUtils G \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://gpus.llm-utils.org/nvidia-h100-gpus-supply-and-demand/#fn:23\u003c/span\u003e\u003cspan address=\"https://gpus.llm-utils.org/nvidia-h100-gpus-supply-and-demand/#fn:23\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChung J-W, Gu Y, Jang I et al (2024) Reducing Energy Bloat Large Model Train\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGholami A, Yao Z, Kim S et al (2024) AI and Memory Wall. IEEE Micro 44:33\u0026ndash;39\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSahni S, Abhyankar A, Aggarwal A et al (2024) \u003cem\u003eBeyond the Beachfront: Integration of Silicon Photonic I/Os under a High-Power ASIC\u003c/em\u003e. in. \u003cem\u003eOptical Fiber Communications Conference and Exhibition (OFC)\u003c/em\u003e. 2024. IEEE\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRizzo A, Daudlin S, Novick A et al (2023) Petabit-Scale Silicon Photonic Interconnects With Integrated Kerr Frequency Combs. IEEE J Sel Top Quantum Electron 29:1\u0026ndash;20\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYoung IA, Mohammed E, Liao JTS et al (2010) Optical I/O Technology for Tera-Scale Computing. IEEE J Solid-State Circuits 45:235\u0026ndash;248\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYuan S, Chen L, Wang Z et al (2021) On-chip terahertz isolator with ultrahigh isolation ratios. Nat Commun 12\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHosseini K, Kok E, Shumarayev SY et al (2022) \u003cem\u003e5.12 Tbps Co-Packaged FPGA and Silicon Photonics Interconnect I/O\u003c/em\u003e, in \u003cem\u003e2022 IEEE Symposium on VLSI Technology and Circuits (VLSI Technology and Circuits)\u003c/em\u003e. pp. 260\u0026ndash;261\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFeng L, Ayache M, Huang J et al (2011) Nonreciprocal light propagation in a silicon photonic circuit. \u003cem\u003eScience\u003c/em\u003e 333, 729\u0026thinsp;\u0026ndash;\u0026thinsp;33\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYu Z, Fan S (2009) Complete optical isolation created by indirect interband photonic transitions. Nat Photonics 3:91\u0026ndash;94\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRizzo A, Novick A, Gopal V et al (2023) Massively scalable Kerr comb-driven silicon photonic link. Nat Photonics 14\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKhilwani D, Lee S, Ou C et al (2024) \u003cem\u003e3D-Integrated, Low Power, High Bandwidth Density Opto-Electronic Transceiver\u003c/em\u003e, in \u003cem\u003e2024 IEEE International Symposium on Circuits and Systems (ISCAS)\u003c/em\u003e. pp. 1\u0026ndash;5\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLu J, Liu S, Tang Q et al (2015) Multi-wavelength distributed feedback laser array with very high wavelength-spacing precision. Opt Lett 40\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang J, Zheng W, Liu A (2023) Polarized Vertical-Cavity Surface-Emitting Laser Arrays With Asymmetrical Structures for Three-Dimensional Sensing With Polarization Imaging. IEE J Quantum Electron 59:1\u0026ndash;8\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhuang R, Ni K, Wu G et al (2023) Electro-Optic Frequency Combs: Theory, Characteristics, and Applications. Laser Photonics Rev 17\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHu Y, Yu M, Buscaino B et al (2022) High-efficiency and broadband on-chip electro-optic frequency comb generators. Nat Photonics 16:679\u0026ndash;685\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePfeifle J, Vujicic V, Watts RT et al (2015) Flexible terabit/s Nyquist-WDM super-channels using a gain-switched comb source. Opt Express 23\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWeimann C, Schindler PC, Palmer R et al (2014) Silicon-organic hybrid (SOH) frequency comb sources for terabit/s data transmission. Opt Express 22\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLu Z, Liu J, Mao L et al (2019) \u003cem\u003e12.032 Tbit/s coherent transmission using an ultra-narrow linewidth qantum dot 34.46-GHz C-Band coherent comb laser\u003c/em\u003e, in \u003cem\u003eNext-Generation Optical Communication: Components, Sub-Systems, and Systems VIII\u003c/em\u003e. p. 109470J\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBernal S, Dumont M, Berikaa E et al (2024) 12.1 terabit/second data center interconnects using O-band coherent transmission with QD-MLL frequency combs. Nat Commun 15\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKippenberg TJ, Gaeta AL, Lipson M et al (2018) Dissipative Kerr solitons in optical microresonators. Science 361\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShu H, Chang L, Tao Y et al (2022) Microcomb-driven silicon photonic systems. Nature 605:457\u0026ndash;463\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShen B, Chang L, Liu J et al (2020) Integrated turnkey soliton microcombs. Nature 582:365\u0026ndash;369\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGeng Y, Zhou H, Han X et al (2022) Coherent optical communications using coherence-cloned Kerr soliton microcombs. Nat Commun 13:1070\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePicqu\u0026eacute; N, H\u0026auml;nsch TW (2019) Frequency comb spectroscopy. Nat Photonics 13:146\u0026ndash;157\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLukashchuk A, Riemensberger J, Tusnin A et al (2023) Chaotic microcomb-based parallel ranging. Nat Photonics\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhou H, Geng Y, Cui W et al (2019) Soliton bursts and deterministic dissipative Kerr soliton generation in auxiliary-assisted microcavities. Light Sci Appl 8:50\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMazur M, Suh M-G, Fulop A et al (2021) High Spectral Efficiency Coherent Superchannel Transmission With Soliton Microcombs. J Lightwave Technol 39:4367\u0026ndash;4373\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCorcoran B, Tan M, Xu X et al (2020) Ultra-dense optical data transmission over standard fibre with a single chip source. Nat Commun 11:2568\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhou X, Urata R, Liu H (2020) Beyond 1 Tb/s Intra-Data Center Interconnect Technology: IM-DD OR Coherent? J Lightwave Technol 38:475\u0026ndash;484\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAlam MS, Li X, Jacques M et al (2021) Net 300 Gbps/λ Transmission Over 2 km of SMF With a Silicon Photonic Mach-Zehnder Modulator. IEEE Photonics Technol Lett 33:1391\u0026ndash;1394\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBerikaa E, Alam MS, Li W et al (2023) TFLN MZMs and Next-Gen DACs: Enabling Beyond 400 Gbps IMDD O-Band and C-Band Transmission. IEEE Photonics Technol Lett 35:850\u0026ndash;853\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu Y, Zhang H, Liu J et al (2024) Parallel wavelength-division-multiplexed signal transmission and dispersion compensation enabled by soliton microcombs and microrings. Nat Commun 15\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu X, Gong Z, Bruch AW et al (2021) Aluminum nitride nanophotonics for beyond-octave soliton microcomb generation and self-referencing. Nat Commun 12\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGu J, Li X, Qi K et al (2023) Octave-spanning soliton microcomb in silica microdisk resonators. Opt Lett 48\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDong B, Dumont M, Terra O et al (2023) Broadband quantum-dot frequency-modulated comb laser. Light: Sci Appl 12\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHelgason \u0026Oacute;B, Girardi M, Ye Z et al (2023) Surpassing the nonlinear conversion efficiency of soliton microcombs. Nat Photonics\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJang JK, Okawachi Y, Zhao Y et al (2021) Conversion efficiency of soliton Kerr combs. Opt Lett 46:3657\u0026ndash;3660\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRahim A, Hermans A, Wohlfeil B et al (2021) Taking silicon photonics modulators to a higher performance level: state-of-the-art and a review of new technologies. Adv Photonics 3\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eReed GT, Mashanovich G, Gardes FY et al (2010) Silicon optical modulators. Nat Photonics 4:518\u0026ndash;526\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePCI-SIG \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://pcisig.com/pci-express-6.0-specification\u003c/span\u003e\u003cspan address=\"https://pcisig.com/pci-express-6.0-specification\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eParsons R (2024) Silicon Photonic Integration of DWDM and Mode-Division Multiplexing for Advancing Multi-Dimensional Data Transmission. \u003cem\u003eCLEO Europe and EQEC 2023\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIntel I-S-OCI -Optical-I-O-Chiplet-Co-packaged-with-CPU. \u003cem\u003eOFC 2024\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSun C (2024) Photonics for Die-to-Die Interconnects: Links and Optical I/O Chiplets. \u003cem\u003eIEEE International Solid-State Circuits Conference (ISSCC)\u003c/em\u003e (2024)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRaj M, Silicon Photonics-Based High-Throughput Optical Transceivers (2024). \u003cem\u003eIEEE International Solid-State Circuits Conference (ISSCC)- Forum 6.5\u003c/em\u003e (2024)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWade M, Anderson E, Ardalan S et al (2021) An Error-free 1 Tbps WDM Optical I/O Chiplet and Multi-wavelength Multi-port Laser. \u003cem\u003eOptical Fiber Communication (OFC) Conference 2021\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHosseini K, Kok E, Shumarayev SY et al 8 Tbps Co-Packaged FPGA and Silicon Photonics Optical IO. \u003cem\u003eOptical Fiber Communication (OFC) Conference 2021\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOng KYK, Chia XX, Rahim AA et al (2024) IMDD Data Transmission with Microresonator Soliton Crystals. \u003cem\u003eOptical Fiber Communication (OFC) Conference 2024\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTanikawa K, Tasaka S, Fujii S et al (2022) FEC-free dense WDM transmission with Kerr soliton microcombs in crystalline MgF2 microresonators. \u003cem\u003ePacific Rim Conference on Lasers and Electro-Optics (CLEO/Pacific Rim)\u003c/em\u003e (2022)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRan S, Zhou G, Li Y et al (2024) Micro-Ring Modulators With Integrated Inductor to Mitigate Bandwidth and Extinction Ratio Trade-Off. IEEE Photonics Technol Lett 36:231\u0026ndash;234\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDub\u0026eacute;-Demers R, LaRochelle S, Shi W (2016) Ultrafast pulse-amplitude modulation with a femtojoule silicon photonic modulator. \u003cem\u003eOptica\u003c/em\u003e 3\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShu H, Chang L, Lao C et al (2023) Submilliwatt, widely tunable coherent microcomb generation with feedback-free operation. Adv Photonics 5\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLin Q, Painter OJ, Agrawal GP (2007) Nonlinear optical phenomena in silicon waveguides: modeling and applications. Opt Express 15:16604\u0026ndash;16644\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang Y, Zhang S, Bi T et al (2023) Geometry optimization for dark soliton combs in thin multimode silicon nitride microresonators. Opt Express 31\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu J, Huang G, Wang RN et al (2021) High-yield, wafer-scale fabrication of ultralow-loss, dispersion-engineered silicon nitride photonic circuits. Nat Commun 12\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGuo Y, Li X, Jin M et al (2022) Hybrid integrated external cavity laser with a 172-nm tuning range. APL Photonics 7:066101\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXiang C, Liu J, Guo J et al (2021) Laser soliton microcombs heterogeneously integrated on silicon. Science 373:99\u0026ndash;103\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-5365298/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5365298/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe escalating demand for high-speed, low-power data transmission between processing units (XPUs) has underscored the limitations of traditional electrical input/output (I/O) technologies. Silicon photonics emerges as a promising solution for chip-level optical I/O by integrating Kerr microcombs, microring-based modulators, and photodetectors. In this study, we demonstrate a record-breaking error-free optical I/O transmission achieving 2.3 Tbit/s per fiber port. This feat is enabled by dark soliton microcombs generated in a 400-nm-thick Si\u003csub\u003e3\u003c/sub\u003eN\u003csub\u003e4\u003c/sub\u003e microring, exhibiting a high conversion efficiency of 49% and an on-chip spectral bandwidth of 28 nm at -5 dBm, achieved through precise coupling and dispersion engineering. Utilizing a silicon microring modulator with an electro-optic bandwidth of 61.7 GHz, 36 comb lines are encoded with PCIe6.0-compatible 64 Gbit/s on-off keying (OOK) signals. Additionally, these comb lines support 100 Gbit/s OOK per channel with a bit error rate (BER) of 10\u003csup\u003e\u0026minus;\u0026thinsp;10\u003c/sup\u003e. The successful integration of these foundry-compatible platforms confirms the viability of microcomb-based optical I/O, paving the way for the next generation of high-speed, energy-efficient data communication systems.\u003c/p\u003e","manuscriptTitle":"High-capacity data transmission for optical I/O driven by efficient microcomb and microring modulator","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-11-07 06:28:58","doi":"10.21203/rs.3.rs-5365298/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"78f87945-7aeb-438f-a6a8-a6b8cbe5014b","owner":[],"postedDate":"November 7th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":39904720,"name":"Physical sciences/Optics and photonics/Other photonics/Solitons"},{"id":39904721,"name":"Physical sciences/Optics and photonics/Applied optics/Integrated optics"}],"tags":[],"updatedAt":"2024-12-11T19:55:13+00:00","versionOfRecord":[],"versionCreatedAt":"2024-11-07 06:28:58","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-5365298","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5365298","identity":"rs-5365298","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}