The relentless growth of data traffic demand in short-reach intra-datacenter interconnects (DCI) necessitates an acceleration of the capacity increase of cost-effective intensity modulation direct detection (IMDD) systems. Currently, the 800G multi-source agreement (MSA) has declared the specifications for 200 G/lane systems [1]. 400 G/lane is the next milestone for IMDD systems enabling 4λ 1.6-Tbps interfaces. However, increasing the system capacity of IMDD systems mandates either transmitting higher symbol rate signals or adopting higher pulse amplitude modulation (PAM) formats. The former choice requires a higher sampling rate (beyond 200 GSa/s) digital-to-analog converter (DAC) that does not currently exist, while the latter requires DAC and analog-to-digital converter (ADC) with higher effective number of bits (ENoB). A solution adopted in academic research is to interleave multiple DAC channels operating between 90 and 128 GSa/s, yielding a higher sampling rate interleaved DAC assembly that can support data transmission at higher symbol rates.

In [2], the authors interleaved three DACs, referred to as digital-band-interleaved (DBI) DAC, to transmit 200 Gbaud probabilistically shaped (PS)-PAM16 (net 494.5 Gbps) over 120 m of standard single-mode fiber (SSMF) with a C-band thin-film lithium niobate (TFLN) Mach–Zehnder modulator (MZM) above the 26% overhead (OH) soft-decision (SD)-FEC normalized generalized mutual information (NGMI) threshold of 0.8456. Additionally, the authors in [3] used an analog multiplexer (AMUX) to multiplex two DAC channels, enabling the transmission of 162-Gbaud PS-PAM16 (net 420 Gbps) above the 0.857-NGMI threshold.

The cost-effectiveness of IMDD systems favors single-DAC operation; however, achieving net 400-Gbps transmission using a single DAC channel has not yet been demonstrated. Figure 1 reviews the high-speed IMDD reports employing conventional PAM formats [2–14]. In this Letter, we demonstrate the transmission of 128-Gbaud PAM16 under the 25% OH SD-FEC bit error rate (BER) threshold and 128-Gbaud PS-PAM16 below the 20% OH SD-FEC threshold using a C-band TFLN MZM driven directly with a single DAC (128 GSa/s, 800 mV_pp at 1 sps), which respectively represent net rates of 410 and 400 Gbps.

 

Fig. 1. Summary of the high-speed IMDD demonstrations.

Download Full Size | PDF

A schematic of our experimental setup and the receiver digital signal processing (DSP) routine is given in Fig. 2. We employ a DSP-free transmitter; thus, we create a random sequence of 2¹⁹ length and generate the PAM symbols, which are loaded directly without any further processing to the memory of an arbitrary waveform generator (AWG: Keysight M8199A). The PAM4/8/16 symbols are generated directly from a random binary sequence, whereas the PAM6 and PAM12 symbols are respectively constructed from the standard 32QAM and 128QAM constellations by demapping the complex-valued (QAM) symbol into two consecutive real-valued (PAM) symbols. We operate the AWG at 1 sample per symbol (sps) with neither pulse shaping nor frequency pre-emphasis filtering; this minimizes the peak-to-average power ratio (PAPR) of the signal and maximizes the driving signal swing (maximum of 800 mV_pp). As we operate without an RF driver, maximizing the driving swing is crucial for optimal transmission performance. The TFLN MZM is driven directly with the AWG output using a 67-GHz ground-signal-ground (GSG) probe. Optically, a 15-dBm external cavity laser (ECL) feeds the MZM via vertical grating couplers with a back-to-back coupling loss of 11 dB (at 1565 nm). The output of the MZM is transmitted over 120 m of SSMF, which corresponds to ∼2-ps/nm dispersion. This is equivalent to the dispersion induced by 2-km O-band transmission for the edge channels in the coarse wavelength division multiplexing (CWDM) grid. Given that our receiver employs a 70-GHz PIN photodiode (PD) without a trans-impedance amplifier (TIA), we needed to employ an EDFA to compensate for the coupling loss and boost the optical signal to 7 dBm. In practice, edge couplers with a much lower coupling loss (1.5 dB/facet) will be used alongside a PIN PD with TIA; this will improve the receiver sensitivity significantly and dispense the need for optical amplification. Before the receiver, a variable optical attenuator is added to sweep the received optical power (ROP). The 100-GHz 256-GSa/s real-time oscilloscope (RTO) digitalizes the PD output, and the DSP is carried out offline. At the receiver, the signal is resampled to 2 sps (for symbol rates below 128 Gbaud), and then processed with a polynomial nonlinear equalizer (PNLE), unless mentioned otherwise. PNLE is a simplified form of the Volterra nonlinear equalizer (VNLE) that uses only the self-beating terms and yields a considerable reduction in computational complexity [15]. The BER is calculated from approximately half a million symbols after demapping to a bit sequence.

 

Fig. 2. Experimental setup and receiver DSP blocks. The inset shows the MZM EO response and RF V_π.

Download Full Size | PDF

The TFLN MZM has 23-mm-long coplanar waveguide electrodes with a nearly 50-Ω on-chip termination and is biased through thermal phase shifters. The MZM has an excess loss of less than 1.5 dB and 25-dB DC extinction ratio. The electro-optic (EO) frequency response (S₂₁) and the extrapolated RF V_π are given in the inset of Fig. 2. The MZM has a 24(66)-GHz 3(6)-dB EO bandwidth (normalized to 5 GHz) with a low-MHz V_π of 1.25 V. The low V_π is a key enabler of the achieved driver-less transmission performance.

Figure 3 presents the transmission performance of the standard PAM formats employing a DSP-free transmitter. The BER versus the symbol rate is given in Fig. 3(a), with the summary in Table 1. We sweep the symbol rate by changing the sampling rate of the AWG (100 to 128 GSa/s) while operating at 1 sps. With an 800-mV_pp driving swing, we transmit 128-Gbaud PAM16 at a BER of 3.4 × 10⁻² under the 25% OH SD-FEC threshold, which corresponds to a net rate of 410 Gbps. The assumed 25% OH SD-FEC has a pre-FEC BER threshold of 4 × 10⁻² and is based on a spatially coupled low-density parity-check (LDPC) [16]. We also transmit 108-Gbaud PAM8 below the 2.4 × 10⁻⁴-KP4-FEC BER threshold, corresponding to net 306 Gbps.

 

Fig. 3. Summary of the standard PAM transmission performance. (a) BER versus the symbol rate. (b) BER sensitivity to the ROP with different receiver equalization schemes.

Download Full Size | PDF

Table 1. Summary of Net Bitrates after 120-m Transmission (ROP, 7 dBm; Driving Swing, 800 mV_pp)

View Table

Figure 3(b) shows the BER sensitivity of the 128-Gbaud PAM16 signal to the ROP, defined as the input power to the PD, with different receiver equalizers. We considered linear feed-forward equalizer (FFE), second-order PNLE, and third-order PNLE with/without maximum likelihood sequence detection (MLSD). The lengths of the equalizer kernels are indicated in the figure. PAM16 signaling is more sensitive to the nonlinearities stemming from the AWG, MZM transfer function, and bias point drifting. Thus, we observe a considerable improvement in the BER with third-order PNLE. MLSD improves the BER marginally, which does not justify the added complexity. Therefore, we limited ourselves to third-order PNLE with 101 first-order, 21 second-order, and 11 third-order beating terms; the further increase in the number of terms improves the BER negligibly.

A major aspect of this work is operating without an RF driver, owing to the high modulation efficiency of TFLN. The BER sensitivity to the driving swing of the different PAM formats at 128 Gbaud is shown in Fig. 4. The system is swing-limited, and the BER keeps improving with increasing the swing. We transmit 128-Gbaud PAM4 under the KP4-FEC threshold with only a 200 mV_pp. CMOS technology can readily support these swing requirements, which highlights the potential to discard the transmitter RF driver. Dispensing the RF driver and the transmitter DSP (pulse shaping and pre-emphasis filtering) reduces the power consumption significantly, which is attractive for the short-reach DCI because of their stringent power constraints.

 

Fig. 4. BER sensitivity to the driving signal swing at 128 Gbaud for different PAM orders.

Download Full Size | PDF

Given the high SNR requirement for PAM16 transmission and its sensitivity to nonlinearities, we consider PS-PAM16. PS offers more granular control over the transmitted signal source entropy (bits/symbol) at the expense of adding a DSP block (power consumption and latency) to the transmitter. We generate PS-PAM16 signals at different source entropies using the constant composition distribution matcher (CCDM) with a Maxwell–Boltzmann distribution. Figure 5(a) shows the BER versus the source entropy and the corresponding line rate for PS-PAM16 compared with standard PAM formats. Here, PAM8 (PAM16) corresponds to an entropy of 3 (4) bits/symbols. We transmit PS-PAM16 at a source entropy of 3.75 bits/symbol under the 2.4 × 10⁻² BER threshold of the 20% OH SD-FEC, corresponding to a net rate of 400 Gbps. The 20% OH SD-FEC is widely adopted and employs a LDPC convolutional code [17]. Due to the higher number of inner levels, PS-PAM16 at an entropy of 3 bits/symbol is significantly worse than PAM8, whereas both PAM-12 and PS-PAM16 show similar performance. However, our range of interest is primarily between 3.5 and 4.0 bits/symbol. Figure 5(b) shows the PAPR of the transmitted PS-PAM and standard PAM signals before loading to the DAC and the corresponding received root mean square (rms) swing as a function of the source entropy or PAM format. For standard PAM (solid curves), the PAPR and the received rms swing are approximately constant as we operate at 1 sps without any signal processing. However, PS leads to an increase in the transmitted signal PAPR, which reduces the AWG driving swing and subsequently the received signal rms leading to a degradation in performance.

 

Fig. 5. (a) BER versus the source entropy (line-rate) of PS-PAM16 and standard PAM at 128 Gbaud. The inset shows the histogram of PS-PAM16 (3.75 bits/symbol) after equalization. (b) PAPR of the transmitted signal and rms of the received signal versus the entropy.

Download Full Size | PDF

The BER sensitivity to the ROP of 128-Gbaud PS-PAM16 (3.75 bits/symbol) is shown in Fig. 6. Compared with PAM16, PS-PAM16 has a higher nonlinearity tolerance, as the edge symbols are sent with lower probabilities [inset of Fig. 5(a)]. Thus, the improvement incurred from employing third-order PNLE compared with the second-order PNLE is less pronounced for the PS-PAM16 case, compared with the PAM16 [Fig. 3(b)]. Interestingly, the BER keeps improving with the ROP; hence, a higher ROP is expected to improve the performance further.

 

Fig. 6. BER sensitivity to the ROP of 128-Gbaud PS-PAM16 (3.75 bits/symbol).

Download Full Size | PDF

This work demonstrates the C-band IMDD transmission of net 400 Gbps on a single carrier using an 800-mV_pp driving signal from a single DAC channel (128 GSa/s) without external RF amplification, with two different schemes: (1) DSP-free transmitter supporting 128-Gbaud PAM16 with the 25% OH SD-FEC; and (2) employing PS-PAM16 at 128 Gbaud with the 20% OH SD-FEC. Therefore, the decisive parameters are the power consumption and induced latency of the PS module versus employing a lower code rate FEC. Our results indicate the promise of operating high-speed transmission systems beyond net 400 Gbps with a single DAC channel directly driving the MZM with simplified DSP requirements and without an RF driver, which will reduce the system power consumption considerably. To the best of the authors’ knowledge, the reported transmission rates (Table 1) are the highest achieved with PAM formats using a single-DAC transmitter at all the considered FEC thresholds.

In conclusion, using a single-DAC channel (128 GSa/s) without an RF amplifier, we demonstrate the transmission of 128-Gbaud PAM16 under the 25% OH SD-FEC BER threshold with a DSP-free transmitter and 128-Gbaud PS-PAM16 below the 20% OH SD-FEC threshold using a TFLN MZM; representing record net rates for the single-DAC operation of 410 and 400 Gbps, respectively.