Introduction

With the ongoing enrichment of mobile services and the explosive surge in internet traffic, the widely deployed cloud radio access network (C-RAN) has garnered increasing attention1,2. Fronthaul, acting as the critical “last mile” connection between the optical core network and mobile wireless network in C-RAN architecture, enables the transmission of carrier-aggregated wireless signals between the centralized/distributed unit (CU/DU) and the remote radio unit (RU)2,3. Since it directly manages a multitude of mobile terminals, it plays a vital role in ensuring seamless internet access, particularly in bustling business hubs, densely populated residential areas, and smart industrial zones2,4. The fronthaul now faces the daunting task of providing a large number of parallel channels for low latency service and supporting carrier aggregated wireless bandwidths of up to hundreds of GHz, which corresponds to common public radio interface (CPRI)-equivalent rates of tens of Tb/s3,5,6,7. The capacity demand will further expand to Pb/s level according to Cisco Visual Networking Index, which reported that global mobile traffic would reach 796.8 EB in 2022 and continue to trend exponentially8. Meanwhile, due to limited wireless spectrum resources, ultra-high-order modulation formats are required, e.g., 1024-ary quadrature amplitude modulation (1024-QAM) as discussed in 3GPP standards9, setting new requirements for the signal-to-noise ratio (SNR) of the fronthaul link.

To establish an extensive network of parallel channels, wavelength division multiplexing (WDM) is the technique of choice for fronthaul links2,3. The current scheme uses uncooled lasers array with coarse WDM spacing as multi-wavelength light sources2,3, resulting in large guard bands and an underutilized spectrum. On the other hand, the broadband electro-optical (EO) comb, capable of compact integration on lithium niobate platforms10, offers a simple, cost-efficient solution by utilizing a single narrow-linewidth laser as a seed source and only a handful of active components for comb lines duplication11,12,13,14,15,16. It can support tens of channels with stable channel spacing (repetition frequency fr), therefore obviating the need for a guard band in WDM and expanding the pool of available channels13,16. Also, the flexibility in channel spacing adjustment can easily accommodate upgrading changes in interface rates. Furthermore, by using highly nonlinear materials for spectral broadening, hundreds of comb lines spanning a bandwidth of tens of nanometers can be generated, making full use of the C-band telecom windows14,17,18,19,20. It’s worth highlighting that all the generated comb lines inherit the exceptional phase noise characteristics of the seed laser, resulting in the successful construction of superchannels.

However, the optical signal-to-noise ratio (OSNR) will deteriorate when increasing the number of parallel WDM channels21,22. The non-ideal optical transceivers can also cause SNR impairments. To break through the noise bottleneck of the fronthaul link, the digital-analog radio-over-fiber (DA-RoF) has shown competitiveness among existing schemes in terms of bandwidth-SNR trade-off12,23,24,25. The DA-RoF technique separates the wireless waveforms into digital and residual analog parts with adjustable parameters. After DA-RoF demodulation, the recovered wireless signals can obtain an exponential SNR improvement. It also eliminates the need for complicated forward error correction, ensuring low-latency and low-power application-specific integrated-circuit chip design12. In order to support high capacity, large coverage, and good power sensitivity, DA-RoF necessitates a coherent architecture, rather than the traditional intensity-modulation direct-detection systems26. However, in the presence of laser phase noise, this transition introduces a challenge, due to the fact that the noise-like analog part of DA-RoF symbols hinders the traditional digital phase-locked loop in coherent systems12. The additional carrier recovery process also increases the cost- and power consumption of coherent detection. To cancel out the phase noise in optical domain, we previously proposed a cloned comb architecture achieving self-homodyne detection in single-mode fiber12. Yet cloning requires a rather complex set of components. As the number of RUs grows in the future, this method may become less appealing due to cost considerations in fronthaul networks. To address this problem, space-division multiplexing (SDM) is a rivaling approach. By using uncoupled multicore fiber, one core transmits a copy of unloaded comb lines as the remote local oscillators (LOs) for self-homodyne detection at the RUs side, while the others transmit the modulated signals25,27. In this way, not only can the need for stringent time delay matching between the signal and LO paths be eliminated thanks to the narrow linewidth characteristic shared by all comb lines, but also the capacity can be further boosted using SDM. This simple architecture can better accommodate the denser and smaller cellular for future millimeter-wave and THz applications.

Here, we propose self-homodyne DA-RoF fronthaul enabled by comb-based WDM and SDM superchannel. As shown in Fig. 1, by transmitting the DA-RoF superchannel and the unloaded EO comb as local oscillators through different cores of a 7-core fiber, we achieve ultra-high capacity and spectral efficiency with massive parallel channels in a simple and scalable manner. Leveraging the EO comb with 1- to 4-order DA-RoF modulation, we successfully demonstrate fronthaul transmission with CPRI-equivalent rates of a 100 Tb/s class, supporting modulation formats ranging from 256-QAM up to 1,048,576-QAM. To push the boundary of capacity, we accomplish a record-breaking CPRI-equivalent rate of 0.879 Pb/s by generating a 3.3 THz continuum-spectrum EO comb distilled with a Fabry–Pérot cavity. This capacity is on par with the current cumulative global mobile internet traffic. Furthermore, to underscore the feasibility of employing integrated optical source devices, we present a packaged on-chip EO comb-enabled fronthaul transmission, representing a pivotal convergence of integrated photonics within telecommunication.

Fig. 1: The high-fidelity and high-capacity DA-RoF fronthaul enabled by comb-based WDM and SDM superchannel.
figure 1

The proposed fronthaul is a fiber-wireless converged architecture addressing the access of massive wireless users in data-intensive urban areas such as smart factories, stadiums, and shopping malls for the coming 6G era. It can simultaneously support 750 RoF channels (Fig. 5), corresponding to an aggregated wireless bandwidth of several thousand GHz. In the CU/DU, we implement an electro-optic comb using advanced techniques including nonlinear spectral broadening and cavity distillation. This results in the generation of a frequency comb with wide spectrum spanning several THz, high optical carrier-to-noise ratios and exceptional phase coherence. Furthermore, this sole comb source can be packaged and integrated into chip-scale devices, driving fronthaul systems into the integrated photonics era. By simultaneously leveraging the uncoupled multicore fiber as the fronthaul link, we can transmit phase-coherent superchannels carrying both signals and unloaded combs. This enables self-homodyne coherent detection at the RUs, reducing the cost and power consumption. In addition, the spatial dimension significantly expands the fronthaul capacity. To break through the link signal-to-noise ratio limitation for recovered wireless signals, we employ DA-RoF techniques. This approach divides the original wireless signals into digital and residual analog parts. The digital part exhibits excellent anti-noise performance, while the analog part is of high resolution to accurately reflect the details of wireless waveforms. With the successful recovery of DA-RoF symbols through self-homodyne coherent detection, the reconstructed wireless signals maintain high fidelity and support ultra-high-order formats at each RU. CU centralized unit, DU distributed unit, RU remote radio unit, IM intensity modulator, PM phase modulator, HNLF highly-nonlinear fiber, EO electro-optic, Mod. modulation, LO local oscillator, DA-RoF digital-analog radio-over-fiber, WDM wavelength-division multiplexing, SDM space-division multiplexing.

Full size image

Results

100 Tb/s-class CPRI-equivalent rate DA-RoF superchannels with formats up to 1,048,576-QAM

We first implement the comb-seeded WDM and SDM superchannels using off-the-shelf components, demonstrating DA-RoF fronthaul with CPRI-equivalent rates of a 100 Tb/s class. Figure 2a shows the experimental setup. In the CU/DU, by using a seed laser with a 1-kHz linewidth and cascading it with one intensity modulator and one phase modulator, we generate 29 comb lines with a repetition frequency of 28 GHz. We boost the EO comb with a polarization-maintaining erbium-doped fiber amplifier (PM-EDFA) and flatten the comb lines with a wavelength-selective switch (WSS). Then we split the EO comb into 2 copies, which serve as data channels and remote LOs, respectively. For signal modulation, a single-polarization in-phase/quadrature (IQ) modulator is utilized to encode the 20 GBaud DA-RoF symbols on all comb lines, due to the limited device in hand. The polarization division multiplexing is achieved through a split-and-decorrelation-based emulator. With DA-RoF technique, the original wireless waveform is separated into two parts. The digital part achieves a rough approximation of the original wireless waveform through the Round(·) operation. Meanwhile, there is an error signal between the roughly approximated digital part and the wireless waveform. By scaling its amplitude, the tolerance to link noise is improved, and the residual analog part is formed. At the RUs, by recombining these two parts, the wireless signals are reconstructed (More details on the DA-RoF implementation are in Methods). For the 7-core MCF transmission, we amplify and divide the DA-RoF superchannels into 6 replicas and decorrelate them in the CU/DU before sending them into core 2 to 7 using a fan-in multiplexer. Core 1, in the center of the fiber, is used for the LO transmission. The transmission distance is set to 2 km, a typical length for the next generation fronthaul2. At the RUs side, a fan-out is firstly utilized to demultiplex the DA-RoF signal and LO channels. Notably, the remote LOs are amplified and divided into 6 copies for the detection of the DA-RoF superchannel. Each copy can serve as LO for DA-RoF signals detection from core 2 to 7. For self-homodyne coherent detection, we firstly use a pair of WSSs to filter out the target wavelength channel from the data and LO superchannels before they are mixed and detected using a standard coherent receiver (More details on the transmitter/receiver-side digital signal processing (DSP) are in Methods). For all DA-RoF data channels, we maintain the received optical power between  −12 dBm and  −11 dBm throughout all the experiments (see Supplementary Note 1 for received optical power sensitivity evaluation).

Fig. 2: Experimental setup of comb-seeded DA-RoF superchannels and results of 100 Tb/s-class CPRI-equivalent rate fronthaul.
figure 2

a The experimental setup of comb-seeded WDM and SDM superchannels. In the CU/DU, we generate 29 comb lines with a 28 GHz fr using 1 kHz linewidth seed laser and one intensity modulator and one phase modulator and flatten it with WSS. A 50:50 coupler splits the EO comb for DA-RoF signal modulation and remote LOs, respectively. The DA-RoF signals are divided into 6 copies and decorrelated before fed into the 2-km 7-core MCF. At the RUs side, we conduct spatial and then wavelength demultiplexing. The remote LOs also are divided into 6 copies for each core’s detection. The target wavelength channels for both the DA-RoF signal and remote LO are filtered out by WSSs and self-homodyne detected. CU centralized unit, DU distributed unit, IM intensity modulator, PM phase modulator, PM-EDFA polarization-maintaining erbium-doped fiber amplifier, WSS wavelength-selective switch (Waveshaper 4000S), AWG arbitrary waveform generator, SP IQ Mod. single-polarization in-phase/quadrature modulator, PDME polarization-division-multiplexing emulator, MCF multicore fiber, RU remote radio unit, PC polarization controller, TOBPF tunable optical band-pass filter, LO local oscillator, RTO real-time oscilloscope. b The measured optical spectra of the generated DA-RoF superchannels and EO comb at RUs. The wavelength channels are indexed by CH 1 to 29 from long to short wavelengths. c The fluctuation of residual phase noise between the DA-RoF data channel and remote LO. The upper, middle and lower one correspond to CH 1, CH 15, and CH 29, respectively. d The recovered constellation of the 1024-QAM wireless signal of core 2, CH 1. e The recovered SNR of the wireless signal for the 6-core, 29-wavelength multiplexed superchannels. All DA-RoF channels meet the 32.0 dB SNR threshold for 1024-QAM format transmission.

Full size image

Figure 2b illustrates the optical spectra of the DA-RoF modulated superchannel and the generated 29-line comb with a resolution of 0.02 nm. The optical carrier-to-noise ratio (OCNR) is 36.9 dB (defined in 0.1 nm noise bandwidth), with  <2 dB power fluctuation after WSS flattening. The DA-RoF data channels are labeled from CH 1 to 29, arranged from long to short wavelengths. To characterize the fluctuation of the residual phase noise between the signal and the remote LO channels, we transmit standard 16-QAM signals as probes and conduct carrier phase estimation. As shown in Fig. 2c, the center channel, CH 15, demonstrates excellent phase stability with fluctuations within  ±0.5°, while CH 1 and CH 29 exhibit slightly larger phase deviations of  ~±2°. This phenomenon can be attributed to the combined effect of the accumulated phase noise of the radio-frequency source on the outer comb lines and the walk-off from the mismatch between the signal and LO optical paths. Nevertheless, this amount of phase fluctuation is small enough to enable self-homodyne coherent detection without carrier frequency and phase recovery stages in DSP. Thanks to the self-homodyne coherent detection with residual phase fluctuation below  ~±3°, almost no sign of rotation appears in the constellation of DA-RoF symbols. The noise-like analog part of the DA-RoF symbols can be successfully recovered12, and the recovered wireless signals can easily exceed the threshold for 1024-QAM format. Figure 2d shows the recovered constellation of the 1024-QAM wireless signal of core 2, CH 1, with a recovered SNR of 32.1 dB. Figure 2e presents the recovered SNR of the WDM and SDM DA-RoF superchannels. All the 174(=29 wavelength × 6 core) channels can achieve an SNR of 32.0 dB or 2.5% error vector magnitude (EVM) threshold for 1024-QAM transmission. This shows that DA-RoF provides an SNR gain of over 11.0 dB. The CPRI-equivalent rate for each channel is calculated as 1.17(=20 Gbaud × 1000 subcarriers/1024 FFT × 3/2 upsampling × 2 I&Q × 15 quantization bits × 16/15 control word × 10/8 line coding) Tb/s, resulting in an aggregated capacity of 203.6 Tb/s. Drawing parallel with 5G New Radio channels with a 100-MHz bandwidth, this DA-RoF fronthaul can support simultaneous access for 34,800 end users while with superior communication quality.

In this demonstration, we emphasize the scalability of the recovered SNR and the achievable wireless modulation formats using the cascaded DA-RoF technique. In principle, the digital part of the DA-RoF symbols utilizes rounding operation to convey and protect the main information of the wireless waveform. Building upon this concept, the error waveform between the digital part and the original wireless waveform can be further scaled and separated into new digital part and new analog error part. The final analog part will show very fine errors and have ultra-high resolution. Thus, cascaded DA-RoF technique can enhance the SNR with an exponential gain and therefore enable higher-order modulation formats (refer to Methods for detailed implementation of cascaded DA-RoF technique and parameter settings of 1- to 4-order DA-RoF scenarios). For 256-QAM format, 1-order DA-RoF can easily achieve  >30 dB recovered SNR. We thus extend the superchannel to 44-line EO comb using one intensity modulator and two phase modulators to trade the SNR margin for larger capacity. Its measured optical spectra of the modulated DA-RoF superchannel and the generated EO comb are shown in Fig. 3a. Compared with Fig. 2b, the OCNR slightly decreases to 34.6 dB due to higher amplified spontaneous emission (ASE) noise from EDFA. Figure 3b–f illustrates the recovered constellations of 256-QAM, 4096-QAM, 16,384-QAM, 65,536-QAM, and 1,048,576-QAM, respectively. The corresponding DA-RoF orders are 1, 2, 3, 3, and 4, resulting in recovered SNR values of 30.8 dB, 38.5 dB, 52.8 dB, 52.2 dB, and 64.0 dB. The aggregated CPRI-equivalent rates are 309.4 Tb/s, 149.5 Tb/s, 102.0 Tb/s, 102.0 Tb/s, and 81.6 Tb/s, respectively (More details on CPRI-equivalent rate calculation are in Methods). As shown in Fig. 3g, we display the recovered SNR as a function of the modulation formats ranging from 256-QAM to 1,048,576-QAM. With the cascaded DA-RoF technique, high recovered SNR from 30.8 dB to 64.0 dB is feasible. Furthermore, additional SNR gains of 6.3 dB, 13.7 dB, and 11.8 dB are achieved from 2-order to 4-order DA-RoF for each order. It is important to note that in an M-order DA-RoF, there are M digital parts and 1 analog part. Consequently, the cascaded DA-RoF technique can trade the linearly increased bandwidth for ~10.0 dB exponential enhancement in SNR. Ultra-high-order QAM formats beyond 1,048,576-QAM can be achieved since cascaded DA-RoF orders can be further increased. This supremely surpasses the ‘doubled-bandwidth’ laws of angle modulation and delta-sigma modulation found in the literature28,29,30,31.

Fig. 3: Scalability in terms of the SNR and modulation format using cascaded DA-RoF.
figure 3

a The measured optical spectra of the 44-line DA-RoF superchannel and EO comb. The wavelength channels are indexed by CH 1 to 44 from long to short wavelength. b The recovered constellation of the 256-QAM wireless signal of core 2, CH 22, using 1-order DA-RoF. c The recovered constellation of the 4096-QAM wireless signal of core 2, CH 18 using 2-order DA-RoF. d The recovered constellation of the 16,384-QAM wireless signal of core 2, CH 18 using 3-order DA-RoF. e The recovered constellation of the 65536-QAM wireless signal of core 2, CH 18 using 3-order DA-RoF. f The recovered constellation of the 1,048,576-QAM wireless signal of core 2, CH 18 using 4-order DA-RoF. Note that the channel is chosen as representative since it has the lowest optical power. g The recovered SNR as a function of modulation formats from 256-QAM to 1048,576-QAM. We present high recovered SNR from 30.8 dB to 64.0 dB with an efficient SNR-bandwidth trade-off of ~10.0 dB per DA-RoF order (namely linearly increased bandwidth).

Full size image

Sub-Pb/s DA-RoF supperchannels using continuum-spectrum EO combs

On top of the modulator-cascaded EO comb, the nonlinear spectral broadening technique allows us to generate numerous comb lines14,17,32, resulting in a continuum comb spectrum that extends from hundreds of GHz to several THz. This advancement unlocks the capacity of Pb/s-level CPRI-equivalent rate for future fronthaul applications.

We employ a span of highly nonlinear fiber to broaden the spectrum of the EO comb. Other highly nonlinear materials can also realize spectrum broadening19,20. Figure 4a shows the initial broadened comb spectrum. We select 132 comb lines, from 191.5 THz to 194.8 THz, with an fr of 25 GHz, and then flatten it with a WSS. To further enhance the OCNR of the broadband EO comb source, we utilize an in-line air-gap Fabry–Pérot cavity to filter all the comb lines simultaneously14 (see Supplementary Note 3 for more details). As shown in Fig. 4b, more than 90% comb lines in the CU/DU exhibit an OCNR higher than 33 dB. We then split the optical power of the filtered comb into two equal parts using a 50:50 coupler. The first part is used for signal modulation. For each test channel, we select and amplify the channel itself and two adjacent ones as signal carriers as the test band13,16,33,34, due to the fact that we only have one IQ modulator with limited input power. These carriers are then encoded with 20 GBaud dual-polarization DA-RoF symbols using a split-and-decorrelation-based emulator. The remaining channels are firstly filtered out and then filled with a flattened ASE source. The resulting superchannel is amplified and fed into cores 2 to 7 of the MCF. The second part, consisting of unloaded comb lines, directly passes through core 1 and serves as remote LOs. At the RUs side, both the superchannels and the remote LOs are amplified, as shown in Fig. 4c, d. Using a pair of WSSs, we select the central channel of the test band and the corresponding LO. In this way, this test/loading band approach can effectively reflect real-life transmission scenarios and allow comprehensive evaluation of all SDM-WDM channels’ performance. It takes into account the crosstalk from adjacent WDM channels and simulate the SNR conditions of the pratical superchannel. (More detailed setup descriptions are in Methods section.)

Fig. 4: Spectra and phase performance of the broadband continuum-spectrum EO comb.
figure 4

a The optical spectrum of the original nonlinear spectral broadened EO comb. 132 comb lines spanning a range of 3.3 THz are flattened and filtered by a FP cavity. b The filtered broadband EO comb in the CU/DU, exhibiting an OCNR higher than 33 dB for most of the comb lines. c The optical spectrum of DA-RoF superchannels after amplification at the RUs side. d The optical spectrum of the amplified remote LOs. All the spectra are in 0.02 nm resolution. e The residual phase fluctuation between the signal carriers and remote LOs for CH 1, CH 77, and CH 132. Variation of below  ±4° enables the self-homodyne detection and ensures the successful recovery of DA-RoF symbols.

Full size image

Thanks to the phase-coherent characteristic of the EO comb, all the comb lines benefit from the exceptional phase noise level of the 1 kHz narrow linewidth seed laser. This feature greatly simplifies the deployment of self-homodyne fronthaul links. Even in the presence of optical path mismatches between cores equivalent to tens of meters12, there is no need for additional delay matching. We measure the fluctuations of the residual phase noise between channels for signal and corresponding remote LO. As shown in Fig. 4e, CH 77, corresponding to the seed laser frequency, exhibits negligible fluctuations within  ±2°. Even for the leftmost channel (CH 1) and the rightmost channel (CH 132), the residual variation remains below  ±4°. This robustness, enabled by the continuum-spectrum EO comb, ensures the successful recovery of the DA-RoF symbols.

Leveraging the self-homodyne detection, we successfully demonstrate the 132 × 6-superchannel DA-RoF system. We take core 5 and core 7 as representatives, since core 5 exhibits the highest insertion loss of 3.5 dB, and core 7 has the highest crosstalk level of  −33 dB (see Supplementary Note 6 for details). Utilizing 1-order DA-RoF, 125 out of 132 channels across all cores achieve the 3.5% EVM threshold for 256-QAM signals, as shown in Fig. 5a, b. The SNR gain from DA-RoF symbols to recover wireless signals is higher than 10 dB. The remaining 7 channels does not meet the requirement due to the relatively low power of the original comb lines, as shown in Fig. 4b. Figure 5c illustrates the recovered constellation of a representative channel for the 20 GHz/λ 256-QAM signals. The total aggregate bandwidth of the fronthaul link reaches 125 × 6 × 20 = 15,000 GHz, allowing the simultaneous access of 150,000 5G New Radio channels. In turn, we achieve a CPRI-equivalent rate of 0.879 Pb/s for the 256-QAM format. Furthermore, by applying 2-order and 3-order DA-RoF, we achieved bandwidths of 13.3 GHz/λ for 4096-QAM signals and 10 GHz/λ for 16,384-QAM signals. Correspondingly, for the 4096-QAM and 16,384-QAM formats, CPRI-equivalent rates of 585.8 Tb/s and 439.5 Tb/s can be realized. Their recovered constellations are shown in Fig. 5d, e.

Fig. 5: Experimental results of the 6 × 132-channel SDM-WDM superchannel.
figure 5

a, b The recovered SNR for all WDM channels in core 5 and core 7 using 1-order DA-RoF. 125 channels in each core can support 256-QAM format transmission with an aggregated wireless bandwidth of 20 GHz/λ. A total CPRI-equivalent rate of 0.879 Pb/s is achieved. c The recovered 256-QAM constellation for CH 20 in core 5 and core 7. d The recovered 4096-QAM constellation with a 6 × 125λ × 13.3 GHz/λ bandwidth. The EVM is 1.08%, below the threshold of 1.29%28. e The recovered 16,384-QAM constellation with a 6 × 125λ × 10 GHz/λ bandwidth. The EVM is 0.28%, much below the threshold of 0.66%28. The received signals for plotting (d, e) are accumulated by CHs 20, 60, and 100 in core 5 and core 7.

Full size image

100.5 Tb/s DA-RoF supperchannels using packaged on-chip EO comb

With the advent of integrated thin-film lithium-niobate technology, chip-scale cascaded-modulator EO combs can now be manufactured and utilized in the CU/DU. This compact and multi-wavelength optical source significantly reduces the cost and footprint of transmitter-side devices and allows for large-scale production. It becomes well-suited for future C-RAN deployments that demand mass production on a global scale.

In our proof-of-concept demonstration, we integrated one IM and two PMs for EO comb generation, as shown in Fig. 6a. The cascaded modulators are fabricated on a lithium-niobate-on-insulator wafer, which consists of a 600-nm-thick X-cut thin-film lithium-niobate layer and a 4.7-μm-thick buried SiO2 layer. The waveguide width is 1.6 μm, and the sidewalls have an inclination angle of 75°. The traveling-wave electrodes are formed by depositing a 600-nm-thick Ti/Au layer and then lifting it off. They are 10 mm in length and feature a 5 μm gap. Matching resistors are placed at the extremities of the traveling-wave electrodes. The RF signals are amplified to ~1 W to drive the PMs through gold wires. The repetition frequency fr are also set to 25 GHz. Due to the higher half-wave voltage of the modulators, fewer comb lines are generated compared to discrete devices, as shown in Fig. 6b. Yet it has been widely reported that the use of micro-structured electrodes or high permittivity cladding can significantly reduce the required half-wave voltage35,36. The lithium-niobate-waveguide modulators and the coupling loss from packaging introduced an insertion loss of 29 dB, resulting in a total output power of  −14 dBm from the on-chip EO comb. Optical amplification is therefore necessary, resulting in rather high ASE noise. Consequently, the original comb lines exhibit a low OCNR ranging from 20 dB to 25 dB, as shown in Fig. 6b. Similar to the scenario where the continuum-spectrum EO comb is filtered, the same in-line Fabry–Pérot cavity also matches and filters all the comb lines with higher OCNR, as shown in Fig. 6c. Then 13 comb lines are flattened and split into signal carriers and unloaded LOs. After modulation and transmission of the DA-RoF signals, Fig. 6d, e display the DA-RoF supperchannel and remotes LOs at the RUs side. (More detailed setup descriptions are in Methods section.)

Fig. 6: Experimental results for 100.5 Tb/s DA-RoF fronthaul enabled by packaged on-chip EO comb.
figure 6

a A photograph of the packaged integrated EO comb and a micrograph of the chip. b The optical spectrum of the on-chip EO comb after amplification. The central comb lines exhibit OCNR of 20 dB to 25 dB. c The comb lines after filtering through the Fabry–Pérot cavity have a remarkable OCNR improvement. The central 13 comb lines are selected for flattening and served as signal carriers and LOs. d, e The spectra of DA-RoF superchannels and remote LOs at the RUs side. The DA-RoF signals achieved an OSNR of  >32 dB, while the remote LOs had an OCNR of  >38 dB. All spectra are captured with a 0.02 nm resolution. f The recovered SNR exceeds 32.0 dB for all channels from core 2 to core 7. All 13λ × 6 channels can support 1024-QAM signals with an aggregated wireless bandwidth of 22 GHz/λ. A total CPRI-equivalent rate of 100.5 Tb/s is achieved. Figure 6a is adapted with permission from50 © Optica Publishing Group.

Full size image

To compare the performance between using on-chip and discrete devices, we first modulate 1024-QAM signals with a 20 GHz/λ bandwidth. The average recovered SNR and EVM are measured as 33.4 dB and 2.14%, respectively (See Supplementary Note 4 for details). We observe that the performance of the recovered wireless signals is even superior to that achieved using discrete cascaded modulators. This is mainly attributed to fewer channels for OSNR improvement. Moreover, the on-chip EO comb also effectively delivers the excellent phase characteristics of the seed laser to all WDM channels. Additionally, the flexible fr makes the on-chip EO comb compatible with the Fabry–Pérot cavity. Thanks to periodically cavity filtering effect, the OCNR for all comb lines can be further improved by more than 12 dB14.

To achieve a total CPRI-equivalent rate of 100 Tb/s for DA-RoF fronthaul enabled by the on-chip EO comb, we set the bandwidth to 22 GHz/λ for DA-RoF signals. Figure 6f illustrates the SNR for all channels from core 2 to core 7. The DA-RoF symbols exhibit an average SNR of 20.9 dB. The recovered wireless signals have an average SNR of 32.4 dB, demonstrating an SNR improvement of 11.5 dB. All channels can meet the 2.5% EVM threshold for 1024-QAM signals. The recovered constellation is shown in Supplementary Fig. 8 in Supplementary Note 4. The total aggregate bandwidth of DA-RoF fronthaul based on the on-chip EO comb reaches 13 × 6 × 22 = 1716 GHz. When converted to the CPRI-equivalent rate, we achieve a rate of 100.5 Tb/s using the 1024-QAM format.

Discussion

Figure 7 highlights several new records we have achieved in terms of the total CPRI-equivalent rate and recovered SNR, encompassing modulation formats ranging from 256(=28)-QAM to 1,048,576(=220)-QAM12,23,24,25,26,29,30,31,37,38,39,40,41. Firstly, by employing a continuum-spectrum EO comb and multi-dimensional multiplexing (WDM & SDM), we have successfully generated supperchannels with a width of 3.3 THz, supporting self-homodyne detection at the RUs side. This breakthrough has enabled us to achieve a CPRI-equivalent rate of sub-Pb/s with ultra-high modulation formats, ushering in the era of Pb/s capacity for RoF fronthaul. Secondly, we have observed an ~10 dB SNR gain with each increase in the orders of DA-RoF signals, demonstrating an exponential SNR improvement while consuming linear bandwidth. In this way, even when transmitting signals with the highest modulation format of 1,048,576(=220)-QAM, the aggregated bandwidth achieves a multiple orders of magnitude higher than that of other existing solutions (See Supplementary Note 7 for more discussion on optical-wireless conversion efficiency). Thirdly, we have demonstrated the feasibility of employing a packaged on-chip comb in fronthaul transmission for the first time. In comparison to discrete devices, the on-chip modulators requires a small footprint and much lower power consumption. Moreover, the achieved CPRI-equivalent rate of 100.5 Tb/s with 1024-QAM format shows a similar level of performance, validating the possibility of chip-scale comb source employing in future RoF systems.

Fig. 7: State-of-the-art.
figure 7

The recovered SNR of wireless signals versus the aggregated bandwidth and CPRI-equivalent rate. IM/PM parallel intensity/phase modulators, KK Kramers–Kronig receiver, PM phase modulation, DSM delta-sigma modulation, DPCM differential pulse-code modulation, DD direct detection, coh. coherent detection, SDM space-dimension multiplexing, IFoF intermediate-frequency-over-fiber.

Full size image

We also emphasize the advantages of implementing a self-homodyne fronthaul architecture facilitated by the broadband EO comb and multicore fiber. Through the modern EO comb technique10,19,42, we can replace hundreds of laser diodes in the CU/DU with a single comb source, requiring only one seed laser, several modulators, and amplifiers for comb generation. Not only does this significantly reduce cost and power consumption per unit bandwidth by more than an order of magnitude, but also each comb line inherits the narrow linewidth characteristic of the seed laser. Moreover, benefiting from homogeneous cores and low phase noise performance of all channels, the residual phase fluctuation are negligible. Consequently, we can easily implement self-homodyne coherent detection for numerous RUs while ensuring sufficient delay-matching margins and excluding carrier-phase estimation steps at the same time. It further achieves a fourfold capacity compared to intensity-modulation/direct-detection systems while keeping the DSP stages simplified at the same level. Simultaneously, the unloaded LO comb can serve as uplink carriers, so any additional optical sources at RUs are no longer necessary.

We further compare the architectures of self-homodyne coherent detection in terms of average spectral efficiency, which is defined as the information rate divided by the number of channels (see Supplementary Fig. 9 in Supplementary Note 8). In theory, remote delivery of the LOs is required to achieve self-homodyne coherent detection, resulting in a compromise in terms of polarization, wavelength, or spatial channels. In a polarization-division multiplexing based architecture, the information-carrying signal is loaded on one polarization, while the unmodulated LO is placed on the orthogonal polarization. The average spectral efficiency (SE) in this case is 1/2 ⋅ SE43. In a WDM-based architecture, at least two comb lines of the M wavelength channels are reserved for conveying the seed frequency and repetition frequency. This results in an average spectral efficiency of (M − 2)/M ⋅ SE12,15,44. As for our SDM-based architecture, only one of the N parallel spatial channels is utilized for remote LO transmission. The average spectral efficiency in this scenario becomes (N − 1)/N ⋅ SE. It is important to note that as the number of spatial channels increases, the average spectral efficiency of SDM-based self-homodyne coherent detection can approximate that of standard coherent systems45. Another property is that the optical paths mismatch τ in our scheme does not cause difficulty for self-homodyne detection as long as we use a seed laser with a narrow linewidth δ, since the residual phase noise \(\left\langle {\phi }_{o}^{2}\right\rangle=2\pi \delta \tau\)46,47.

In summary, we have successfully demonstrated a high-fidelity fronthaul link with a record-breaking CPRI-equivalent rate of 0.879 Pb/s, thanks to the utilization of the broadband EO comb and multicore fiber. This marks the commencement of the Pb/s era, reaching new heights for the fronthaul capacity. The scalability of modulation formats ranging from 256-QAM to 1,048,576-QAM has also been showcased. Furthermore, we have demonstrated the feasibility of employing on-chip EO comb in the CU/DU and achieved a remarkable CPRI-equivalent rate of 100.5 Tb/s using the 1024-QAM format. This accomplishment represents a significant advancement, propelling fronthaul into the era of integrated photonics. Looking ahead, we envision the integration of high-nonlinear waveguide and Erbium-doped waveguide amplifiers with the EO comb19,48, opening up new possibilities for chip-scale comb systems that cover the entire C+L band21,22,34,42,49. Additionally, the transmission link can be upgraded to fibers with much more cores21,45. At that point, fiber nonlinearity and time skews among cores should be further considered44. Through the implementation of these breakthroughs, we will be able to harness the potential for even a tenfold capacity with a CPRI-equivalent rate beyond 10 Pb/s and enhanced performance in fronthaul and other fiber networks.

Methods

Implementation of DA-RoF modulation and demodulation

For the 1-order DA-RoF modulation, the normalized aggregated wireless signal S0 is separated into a digital part SD and an analog part SA. SD=B ⋅ [Round(AS0+C)−C]/A and SA = (S0−SD/B) ⋅ 2A, where A is the rounding factor, B is the scaling factor, and C = 0 or 0.5+0.5j is the offset factor. During DA-RoF demodulation, the recovered original wireless signal \({S}_{0}^{{\prime} }\) is obtained by \({S}_{0}^{{\prime} }={\rm{Round}}(A/B\cdot {S}_{D}^{{\prime} }+C)-C+1/2\cdot {S}_{A}^{{\prime} }\), where \({S}_{D}^{{\prime} }\) and \({S}_{A}^{{\prime} }\) are the receiver-side digital and analog parts. In the experiment using the discrete-device and on-chip EO comb, we utilize (A, B, C) = (4.3, 1.7, 0) for the 20G Hz/λ 1024-QAM signal and (A, B, C) = (3.9, 1.7, 0) for the 20 GHz/λ 256-QAM signal. In the experiment involving the continuum-spectrum EO comb, we employ (A, B, C) = (4.3, 1.9, 0) for the 20 GHz/λ 256-QAM signal (see Supplementary Note 1 for DA-RoF parameters optimization).

For M-order DA-RoF modulation, each order has its independent parameters Ai, Bi, and Ci. In addition, we introduce a parameter Ei to describe the residual part at each order. For order i, the (i − 1)th-order residual part Ei−1 is separated into the ith-order digital part SDi and the ith-order residual part Ei. Specifically, we have E0 = S0, SDi = Bi⋅[Round(AiEi−1 + Ci) − Ci]/Ai, and Ei = (Ei−1 − SDi/Bi) ⋅ 2Ai. Finally, the analog part SA = EM. (see Supplementary Fig. 3 in Supplementary Note 1 for the flow diagrams of M-order DA-RoF modulation and demodulation). By cascading the M orders of the digital part and the analog part, we obtain the M-order DA-RoF signal SD1 … SDM SA. In the experiment using the cascaded-modulator EO comb, we utilize (A1, A2; B1, B2; C1, C2) = (3, 2; 2, 1.3; 0.5 + 0.5j, 0.5 + 0.5j) for the 14.7 GHz/λ 212-QAM signal. We employ (A1, A2, A3; B1, B2, B3; C1, C2, C3) = (3.5, 2, 2; 2, 1.3, 1.3; 0.5 + 0.5j, 0.5 + 0.5j, 0.5 + 0.5j) for the 10 GHz/λ 214-QAM and 216-QAM signals. Moreover, we used (A1, A2, A3, A4; B1, B2, B3, B4; C1, C2, C3, C4) = (3.5, 2, 2, 2; 2, 1.3, 1.3, 1.3; 0.5 + 0.5j, 0.5 + 0.5j, 0.5 + 0.5j, 0.5 + 0.5j) for the 8 GHz/λ 218-QAM and 220-QAM signals. In the experiment using the continuum-spectrum EO comb, we employ (A1, A2; B1, B2; C1, C2) = (3.9, 2.5; 2, 1.3; 0.5 + 0.5j, 0.5 + 0.5j) for the 13.3 GHz/λ 212-QAM signal. We use (A1, A2, A3; B1, B2, B3; C1, C2, C3) = (3.5, 2, 2; 2, 1.3, 1.3; 0.5 + 0.5j, 0.5 + 0.5j, 0.5 + 0.5j) for the 10 GHz/λ 214-QAM signal.

The CPRI-equivalent rate can be calculated by 1000 subcarriers / 1024 FFT × 3/2 upsampling × 2 I&Q ×15 quantization bits × 16/15 control word × 10/8 line coding × 2 dual-polarization × R symbol-rate / (M + 1) digital & analog parts × 6 cores × N wavelengths.

The detailed experimental setup using continuum-spectrum EO comb

Supplementary Fig. 4 in Supplementary Note 2 illustrates the detailed experimental setup. In the CU/DU, we employ the traditional approach to generate a 25-line EO comb seed with a repetition frequency of 25 GHz. To achieve this, we use a fiber laser (Koheras X15) as seed laser and then cascade two low Vπ phase modulators and one intensity modulator, all driven by low-noise RF amplifiers. The EO comb seed is then amplified to ~18 dBm using a PM-EDFA. We utilize an 80:20 coupler to split the power into two parts. The 80% part, undergoes dispersion compensation via a 173-meter single-mode fiber to achieve optical pulse compression. Subsequently, a spool of highly nonlinear fiber is applied to broaden the spectrum of the EO comb seed17,32, resulting in the generation of numerous comb lines shown in Fig. 4a. Due to relatively higher power fluctuation in the comb lines close to the seed laser frequency, we employ a WSS to filter them out and also flatten the remaining comb lines spanning a spectrum of 3.3 THz. The central comb lines ranging from 193.1 THz to 193.7 THz are complemented using the 20% part of the comb after extraction and flattening. These two sets of comb lines pass through polarization controllers individually and then are combined through a 70:30 coupler. We further amplify the flattened comb to 17 dBm using another PM-EDFA, with most of the comb lines exhibiting an OCNR between 20 dB to 23 dB. An air-gap Fabry–Pérot cavity matching the fr is then added to improve the OCNR of all comb lines by an additional 12 dB14 (see Supplementary Note 3 for details).

We equally divide the power of the filtered comb for remote LOs and the signal carriers at RUs and CUs’ side, respectively. The LO comb has a power of 6 dBm and is directly fed into core 1 for transmission to the RUs side. To generate the DA-RoF superchannel, we establish both test and loading bands. For testing CH n, we employed a WSS to select 3 comb lines from CH n-1 to CH n+1 as signal carriers. After a polarization controller, these carriers are amplified to 13 dBm using a PM-EDFA. Subsequently, a 20 Gbaud DA-RoF signal from a 100 GSa/s 35-GHz arbitrary waveform generator is modulated onto these carriers through a single-polarization IQ modulator. Another PM-EDFA is used to amplify the three adjacent channels. After a polarization-division-multiplexing emulator, we obtain the test band signals. For the loading band which consists of the other 129 channels, we employ an EDFA as an ASE source to fill them up. We apply a WSS to flatten the channels, which filters out the ASE noise in the test-band area. The test and loading bands are then combined to form the superchannel signal and is amplified to 14 dBm using an EDFA. After being divided into six copies, the superchannel are injected into cores 2 to 7, with a power of ~6 dBm/core. At the RUs side, we use an EDFA to amplify the remote LO comb to 14 dBm. A WSS selects the comb line corresponding to CH n, and another PM-EDFA amplified the LO to 9 dBm after passing through a polarization controller. To filter out the ASE noise introduced by the EDFAs, a tunable optical bandpass filter with a 1 nm bandwidth is employed. Simultaneously, we use another EDFA to amplify the superchannel from the test core and utilize a WSS to select the test CH n with a power of  −12 to  −11 dBm. Finally, we mix the filtered test channel and LO using a standard coherent receiver. We sample the 4-channel output signal by a 33-GHz real-time oscilloscope with an 80 GS/s sampling rate.

The detailed experimental setup using on-chip EO comb

Supplementary Fig. 6 in Supplementary Note 3 provides a detailed illustration of the experimental setup. In the CU/DU, the packaged integrated EO comb is seeded by the same fiber laser (Koheras X15) with an optical power of 15.2 dBm. The packaged on-chip modulators are driven by amplified RF signals through encapsulated metal probes and gold wires. A PM-EDFA amplifies the comb to ~13 dBm, and the same Fabry–Pérot cavity matches and filters all the comb lines with a 7 dB insertion loss. We use a WSS to flatten the selected 13 comb lines, and another PM-EDFA further boosts the optical power to 16 dBm. The optical power is then split into two equal parts using a 50:50 coupler. One part becomes the DA-RoF signal carriers, which passes through an IQ modulator, a PM-EDFA, and a polarization-division-multiplexing emulator. These generated dual-polarization DA-RoF signals are then amplified and divided into 6 copies corresponding to core 2~7. Simultaneously, the second part is directly sent into core 1 and serves as the remote LOs. At the RUs side, the same receiving process is utilized, as in the previous experiments.

Digital signal processing procedures

Supplementary Tables 1 and 2 in Supplementary Note 5 depict the DSP streams. At the transmitter, the binary bits are mapped into QAM symbols, which span from 28-QAM to 220-QAM formats. The orthogonal-frequency-division-multiplexing (OFDM) modulation is achieved through a 1536-point fast Fourier transform. In the process, symbols are loaded on 1000 subcarriers12,38. Subsequently, the waveforms undergo 3:2 downsampling and are modulated into DA-RoF symbols. Symbol lengths range from 49,152 to 61,440 varying with different rates across 1-order to 4-order DA-RoF structure. Prior to the DA-RoF symbols, we add a preamble including 256 synchronization symbols and 2048 16-QAM training symbols. All symbols are upsampled, Nyquist pulse shaped with a 0.05 roll-off, and ultimately re-sampled to match the 100 GS/s sampling rate of the 35-GHz bandwidth arbitrary waveforms generator. At the receiver, the initial step involves Löwdin orthogonalization-based front-end correction. The received signals are then re-sampled at a rate of 4 samples per symbol and undergo matched filtering. Following synchronization, we train a 4 × 2 linear recursive-least-square equalizer using the 2048-symbol preamble and apply it to all DA-RoF symbols. The frequency offset and carrier phase estimation steps are omitted owing to the self-homodyne detection. During experiments conducted at different rates, the optimized tap numbers range from 20 to 30. Finally, DA-RoF and OFDM demodulation are performed, and we calculate the recovered SNR and EVM.