Open Access
Add to BibSonomy
Abstract
This work contributes experimental demonstrations and comprehensive comparisons of various modulation and coding techniques for 200 Gb/s intensity modulation and direct detection links including four-level pulse amplitude modulation (PAM-4), PAM-6, trellis-coded modulation (TCM) over PAM and discrete multi-tone (DMT) transmission. Both C-band Mach-Zehnder modulator and O-band electro-absorption modulated laser transmitters were examined for intra-data center applications based on state-of-the-art commercial components.
© 2020 Optical Society of America under the terms of the OSA Open Access Publishing Agreement
1. Introduction
Following the standardization of 400 gigabit ethernet (GE) [1,2] and driven by the upgrade of switch bandwidth for data center network (DCN) applications [3], the next generation (NG) Ethernet pluggable modules targeting 800 Gb/s or 1.6 Tb/s will soon be needed to support the NG switches with a bandwidth of 25.6 Tb/s and 51.2 Tb/s, respectively. Figure 1 illustrates the typical scenarios of DCN based on a mesh architecture and their corresponding optical module evolution roadmap from a historic point of view. 800 Gb/s optical interconnects mainly address the connections from leaf to top of rack (TOR) switches with up to 100 m reach and between spline to leaf switches of 500 m to 2 km. Coherent solutions are promising to convey such high-capacity data over one or two wavelengths, but are not well suited for intra-DC applications due to the relatively high transceiver cost and complexity. Intensity modulation and direct detection (IMDD) solutions are preferred. Recently, an 800G multi-source agreement had been announced [3] and a part of the announcement considered the use of 200 Gb/s/lambda for DCN applications. Indeed, various IMDD systems transmitting 200+ Gb/s per wavelength have been demonstrated [4–22] to verify the technical feasibility of pluggable modules supporting 800 Gb/s or beyond for links of up to 2 km of single-mode fiber (SMF).
Efforts have been made to tackle the technical challenges of realizing >200 Gb/s IMDD links by introducing novel components such as special high-speed DACs [4,5] or self-developed special ultra-broadband optical modulators [6–8,17]. Even with such high-performance components, powerful digital signal processing (DSP) is required to mitigate various linear and nonlinear distortions. Recently, lab demonstrations were performed to show the feasibility of 200 Gb/s/λ short-reach DCN enabled by advanced modulation formats and DSP and based on commercial off-the-shelf components with strong bandwidth limitation [9–15,18–22]. Demonstrated modulation formats mainly include four level pulse amplitude modulation (PAM-4) [10,15,18–21], PAM-6 [12,14,18,20,21], PAM-8 [21,22], and discrete multi-tone (DMT) [4,13,20]. Trellis coded modulation (TCM) can further improve the power sensitivity of PAM and DMT schemes [11,13,16,20]. The commercial transmitters mainly rely upon Mach–Zehnder intensity modulators (MZM) [10–14,18–21] and electro-absorption modulated lasers (EMLs) [9,18]. The former can be viewed as a benchmark transmitter, while the latter are more likely to be incorporated in real 800GE products in order to leverage the 400 GE ecosystem and Ethernet standard compatibility.
However, there has been lack of a comparative study of various modulation formats, DSPs, and hardware, which would be practically important for choosing the right system and component configuration. This work extends our previous work [20] and contributes a comprehensive experimental demonstration and direct comparison of 224 Gb/s data links based on various advanced modulation formats as PAM-4, PAM-6, TCM on two dimensional eight level PAM (2D-PAM8) and DMT, by using both an MZM and an EML. With respect to [20] we offer a significantly more detailed analysis and we include measurement results of an O-band EML-based optical link.
2. Experimental setup
Figure 2 depicts the experimental setup for the investigated 200 G IMDD links. The transmitters rely on offline DSP to generate the waveforms of interest, and use a 120-GS/s 35-GHz arbitrary waveform generator (AWG) to convert the digital waveform into an analog signal. A linear driver with bandwidth of 50 GHz is used to amplify the analog signal before modulating a MZM or an EML. The MZM (EML) has a 33-GHz (40-GHz) 3-dB bandwidth. The laser externally fed to the MZM and the laser integrated with the EML operate at 1550 nm and 1310 nm, respectively. After transmission over SMF, the optical signal reaches a combined optical receiver comprising a variable optical attenuator (VOA), a semiconductor optical amplifier (SOA), and a photodiode (PD) with a bandwidth of approximately 60 GHz. The SOA gain is adjustable so that the input power injected into the PD can be optimized. The detected signal is then converted into a digital signal by a 160 GS/s oscilloscope with a bandwidth of 60 GHz. The digital signal is finally subject to offline DSP. It is interesting to note that the received signal spectra for the MZM and EML links are quite similar in the absence of digital pre-distortion (DPD), as indicated in the inset of Fig. 2. This is because although the MZM has a smaller 3-dB bandwidth than then EML it has a very slow frequency roll-off.
Fig. 2. Experimental setup for 200G IMDD systems using various advanced modulation formats and commercial EML and MZM.
The offline DSP stack for the transmitter and the receiver, which is dependent on the modulation format, is also presented in Fig. 2. At the transmitter, DPD is optionally used for PAM schemes. The DPD aims to compensate the limited bandwidth of the transceiver [12]. Hard clipping is applied after DPD to limit the signal peak-to-average power ratio (PAPR) at 8 dB before loading the samples to the AWG.
It is worth mentioning an alternative scheme for bandwidth pre-compensation, namely the Tomlinson-Harashima precoding (THP) [21,22]. As a replacement of receiver side decision feedback equalizer (DFE), THP can avoid the error propagation issue. THP is beneficial only when extremely strong bandwidth limitation and serious fiber chromatic dispersion issues exist [21], which does not apply to this demonstration. In addition, THP leads to an increased number of amplitude levels, which makes it incompatible with a standard Viterbi equalizer for PAM. Therefore, THP is not considered here.
2.1 PAM
Both PAM-4 and PAM-6 are demonstrated in this work, which correspond to baud rates (bit rates) of 112 Gbaud and 90 Gbaud (224 Gb/s and 225 Gb/s), respectively. The transmitter DSP includes a pseudorandom binary sequence generator, a PAM encoder mapping the bit stream into symbols, and an optional DPD. For PAM-6, the encoder firstly maps the bits to a cross 32-QAM constellation, whose I and Q projections are then sequentially transmitted [11,12].
In the receiver DSP, the signal is first resampled to a rate of 2 samples per symbol. An automatic gain controller followed by timing recovery (TR) and an anti-aliasing low-pass filter is then applied. The signal is then equalized by a full Volterra feed-forward equalizer (FFE) which contains both linear and nonlinear kernels, as indicated in Eq. (1).
2.2 2D PAM-8
2D PAM-8 TCM encodes five bits to two PAM-8 symbols (±1, ±3, ±5, ±7), as indicated in Fig. 3. This is achieved by generating an additional redundant bit by means of an 8-state convolutional encoder similar to that shown in [24], as indicated in Fig. 3(a). The levels are divided into two groups A=(±7, ±3) and B=(±5, ±1) according to the set partitioning rule shown in Fig. 3(b) and four symbols make eight subsets from S0 to S7, as represented by Fig. 3(c). Subsequently, the mapper selects a 2-dimensional PAM-8 symbol pair out of 64 possibilities. The 8-state trellis is shown in Fig. 3(d). The decoder selects the best candidate from each subset S and one of four branches entering each state is selected. The TCM decoder uses 8-state MLSE and decodes 2 PAM-8 symbols (i.e. 5 bits) per step. The combination of DB equalization and 2D PAM-8 is not considered because of its complexity.
Fig. 3. (a) 2D PAM-8 encoder. (b) 2D PAM-8 set partitioning. (c) 2D PAM-8 symbol group subsets. (d) 2D PAM-8 8-state trellis.
2.3 DMT
The DSP at the transmitter is similar to classic DMT systems [25], except that 16-state TCM is employed to increase the effective Euclidean distance and improve the system performance. Before the TCM encoder, the input bit stream is parallelized and bit/power loading based on Chow’s algorithm [25] is applied. The IFFT size is 1024 and a cyclic prefix (CP) of 16 samples is added. The TCM encoder is implemented in a way similar to that shown in Section 2.2. The DMT features a high peak-to-average power ratio (PAPR) compared with PAM systems. Hard clipping is used to limit the PAPR at 10 dB.
In receiver DSP, the signal is firstly re-sampled to 1040 samples per DMT symbol. The synchronization process is realized by a sliding window and correlation technique to identify the beginning of the DMT symbols. A FFE using conventional Volterra base functions or absolute value (ABS) functions by introducing abstract-terms x(k1)|x(k2)| and x(k1)| x(k2)|2 to replace the standard Volterra terms shown in Eq.(1) is used to mitigate system nonlinearities [13], including signal-to-signal beating interference (SSBI) caused by the direct-detection process, and modulator nonlinearity. After serial to parallel (S/P) conversion, CP removal and FFT, the signal is transformed to frequency domain and the 1-tap equalization is used to compensate the linear distortions of the system. A following TCM decoder and a QAM de-mapper yield the estimated bit sequence before the system BER is calculated.
For all modulation schemes described above including PAM-4, PAM-6, 2D PAM-8, and DMT, we consider a Reed Solomon forward error correction (FEC) code with an overhead of 12% and a threshold bit error rate (BER) of 2×10−3 throughout the paper. Note that TCM could reduce error floor, 2D PAM-8 and DMT may consider the KP4 FEC code with 5.9% overhead and a threshold BER of 2×10−4 when low overhead FEC is required.
3. Results
3.1 DPD effects
As described above, a significant challenge of a 200G IMDD link stems from the limited bandwidth of available commercial transceivers. We considered two efficient approaches to address this challenge. The first one is to adopt DB equalization as indicated in Fig. 4(a), which dramatically reduces the bandwidth requirement at the cost of increased SNR requirement. The second one is to apply DPD at the transmitter. Figure 4(b) shows the effect of the DPD on the signal spectrum. Without DPD, there exists a significant gap at high frequencies between the received signal (before the nonlinear Volterra FFE) and both the DB and full response spectra targets. Full response equalization results in a large noise enhancement, whereas DB equalization can partly filter out the high-frequency noise components at the cost of an increased number of amplitude levels and a higher signal-to-noise ratio (SNR) requirement. Such a trade-off determines the choice of the receiver DSP. In the presence of DPD, the received signal before the nonlinear Volterra FFE shows a very flat spectrum and a good match with the full response target up to 40 GHz, beyond which a gap still exists. Whereas a stronger DPD could close the gap, it would dramatically increase the signal PAPR and effectively degrade the overall system performance.
Fig. 4. (a) Simulated spectrum of 224 Gb/s PAM-4 and PAM-6 and their corresponding DB counterparts. (b) Measured 90-Gbaud PAM-6 signal spectra before and after DB and full response equalization. (c) 90-Gbaud PAM-6 system BER versus received optical power with and without DPD.
Figure 4(c) presents the 225-Gb/s PAM-6 system BER versus the received optical power (ROP) for the setup with MZM with and without DPD in the case of optical back-to-back (BtB) and 1-km SMF transmission. In the absence of DPD, DB equalization together with MLSE with a memory length of 1 were found to be the best choice for the receiver DSP. In the presence of DPD, the receiver DSP only needs a full response equalizer. This is because MLSE does not bring about significant gain in this scenario with negligible inter-symbol interference. For reasonably moderate bandwidth limited channel cases such as to PAM-6 here, the benefits of using DPD at the transmitter with full response equalizer at the receiver compared to using FFE + MLSE at the receiver are two-fold: first, it enables better optical power sensitivity. As indicated in Fig. 4(c) 1.6-dB and 1.2-dB improvements in optical power sensitivity are obtained for optical BtB and 1 km SMF transmission, respectively, while, additionally, the error floor is slightly improved. Second, DPD enables a receiver DSP architecture excluding a MLSE without performance degradation.
Although only PAM-6 is presented here as an example to indicate the importance of DPD, the situation is similar for 2D-PAM8 which also has a baud rate of 90 Gbaud and a spectral efficiency of 2.5 bit/s/Hz. For PAM-4, which has higher baud rate (112 Gbaud), a nonlinear Volterra DB FFE and a MLSE are preferred in the receiver DSP even if DPD is applied. This is because the channel shows a very sharp roll-off beyond 45 GHz and, to avoid an excessive increase of the PAPR, it is not desirable to pre-compensate the signal near the Nyquist frequency. If higher bandwidth transceivers are available in future, PAM-4 will require a similar transceiver DSP configuration as described above for PAM-6. The performance shown in the following mainly refers to DSP configurations with DPD. These considerations also apply to the O-band EML-based link since this exhibits a similar end-to-end bandwidth compared to MZM case, as indicated in the inset of Fig. 2.
3.2 Nonlinearity
Figure 5 illustrates the system nonlinearity and the effectiveness of nonlinear Volterra FFE for PAM-4 and PAM-6, respectively. The receiver nonlinear Volterra FFE configuration is denoted as “[M1, M2, M3]”, where M1, M2, and M3 are the memory lengths for linear, 2nd- and 3rd-order Volterra kernels, respectively, as indicated in Eq. (1). A long memory for the linear terms is used because of the reflections in our discrete setup [11]. Note that a compact integrated setup would exhibit less reflection and therefore require a lower number of linear taps. Three cases including a linear FFE, a nonlinear FFE with linear and 2nd-order kernels, and a nonlinear FFE with linear, 2nd and 3rd-order kernels are considered. For both PAM-4 and PAM-6, Fig. 5 clearly shows that the use of nonlinear equalization improves both optical power sensitivity and error floor, regardless of the transmitter choice. For the setup with MZM as indicated in Fig. 5(a), the 3rd order nonlinear FFE brings about much more significant improvement than the 2nd order nonlinear FFE, indicating that the system is dominated by the 3rd order nonlinearity. On the other side, for the EML case, as indicated in Fig. 5(b), the 3rd order nonlinearity is much weaker compared to the MZM case. In general, the EML-based link exhibits a smaller overall nonlinearity. However, it suffers from a higher error floor especially for PAM-6 due to the fact that the EML has a lower extinction ratio (ER) compared to the MZM. A similar conclusion applies to 2D-PAM8 for MZM case.
Fig. 5. BER as a function of ROP for 224 Gb/s DB PAM-4 and 225 Gb/s PAM-6 subject to different Volterra FFE configurations for (a) MZM and 1 km SMF, and (b) EML and 5 km SMF.
The complexity of the Volterra filter can be reduced by introducing restrictions on the tap spacing among samples participating in the same kernel without significant performance degradation [19]. However, we adopt a full Volterra equalizer here to check the best achievable performance.
3.3 System BER performance
The overall BER performance for each modulation scheme is presented in Fig. 6. We have chosen to consider DPD and DB FFE with DB-MLSE for PAM-4 and DPD and full response nonlinear Volterra FFE for PAM-6, respectively. We consider 2D-PAM8 with DPD, full response nonlinear Volterra FFE, and a TCM-MLSE decoder after the FFE.
As indicated in Fig. 6(a)-(b), assuming FEC with a threshold BER of 2×10−3, PAM-6 with DPD achieves the best receiver optical power sensitivity (-9.4 dBm) for the optical BtB case and exhibits about 1 dB penalty for 1 km SMF compared to optical BtB. DB PAM-4 shows 0.7 dB power penalty at the threshold BER compared to PAM-6 for optical BtB. DMT shows similar optical power sensitivity to DB PAM-4 for optical BtB case, while 2D-PAM8 exhibiting slightly better sensitivity than DB PAM-4 and DMT. For the 1 km SMF transmission case, 2D PAM-8 has the best optical power sensitivity of -8.6 dBm, with about 0.2 dB gain compared with all the other three schemes which exhibit similar sensitivity.
Especially, 2D-PAM8 brings about a much improved error floor compared to all other schemes for both optical BtB and 1 km SMF cases, indicating its potential to support simpler FEC with higher BER threshold. Although DMT uses also TCM, its inherit disadvantages of high PAPR and strong intra-sub-carrier beating noise upon square-law direct detection limit its error floor performance.
The benchmark setup already shows the great potential of non-TCM PAM schemes in view of their comparable or better optical power sensitivity but simpler DSP architecture compared with the TCM schemes. Therefore, the EML setup focuses on examining the two PAM schemes. For the EML case shown in Fig. 7, DB PAM-4 and PAM-6 achieve in the optical BtB case a sensitivity of -8.5 dBm and -8 dBm, respectively, at the threshold BER of 2×10−3, which show about 0.6 dB penalty compared to the MZM counterpart as well as increased error floor due to the reduced ER of the EML. PAM-6 shows a high error floor than DB PAM-4 due to its sensitivity to ER. After 5 km SMF, both DB PAM-4 and PAM-6 show negligible penalties and similar error floors compared to the optical BtB case, since the accumulated fiber chromatic dispersion at 1310 nm is very small.
The overall performance of each modulation scheme for MZM and EML cases are summarized in Table 1. It is interesting to note that the EML achieves a better optical power sensitivity compared to the MZM for DB PAM-4 at a BER of 2×10−3 after SMF transmission. This is attributed to the lower nonlinearity of the EML based link and the negligible dispersion compared to C-band transmission with the MZM, as indicated in section 3.2. Concerning the error floor, in the EML case DB PAM-4 is preferable to PAM-6. For the MZM case, 2D-PAM8 and DMT show no power penalty compared to DB PAM-4 and PAM-6.
Table 1. Summary of performance and FEC/DSP requirement
3.4 Complexity
In addition to the performance, Table 1 also lists the DSP requirements for the considered modulation schemes. For both MZM and EML cases, PAM-6 requires the least complex DSP under the current component bandwidth constraints. Although 2D PAM-8 with superior error floor performance in the MZM case could potentially allow less complex FEC, the DSP complexity advises against its practical implementation, unless a simple FEC is required. DMT in the MZM case shows no advantages on optical power sensitivity, error floor and complexity over PAM-4 and PAM-6, indicating it is not a strong candidate.
4. Discussions
It is clear that the choice of the modulation format is a trade-off between optical link power budget and DSP complexity. The optical link power budget is determined by both optical power sensitivity and transmitter power. Table 1 clearly shows that DB PAM-4 using the EML achieves the best optical power sensitivity due to the negligible dispersion in O band. Depending on the implementation platform, an EML and a MZM could have similar output power, but an EML requires a smaller driving voltage and is more power efficient [26]. In the EML case PAM-6 shows a relatively high error floor as a consequence of the ER limitation. Regarding the complexity, PAM-6 needs the simplest DSP. However, when the next generation of components with 45 GHz bandwidth or beyond is available, PAM4 could reduce the DSP requirements significantly (like PAM-6 does) and become the overall best choice with respect to performance and complexity because of its intrinsic noise tolerance.
5. Summary
Experimental demonstrations of 200 Gb/s per lane IMDD short-reach links have been undertaken based on modulation candidates including PAM-4, PAM-6, TCM-encoded 2D-PAM8, and DMT using both a C-band MZM and an O-band EML. PAM-6 shows better trade-off between performance and DSP complexity using current commercial components with strong bandwidth limitation but it exhibits the worst error floor when using EML due to ER limitation, making it very undesirable. When the next generation of components with improved bandwidth is available, PAM4 could reduce the DSP requirements significantly and become the best choice because of its intrinsic noise tolerance. TCM-encoded 2D-PAM8 gives rise to the best error floor but its complexity may prohibit its practical use unless the system requires a FEC with lower overhead but higher threshold BER. DMT does not show any advantage in either performance or DSP complexity.
Disclosures
The authors declare no conflicts of interest.
References
1. IEEE 802.3bs-2017 - IEEE Standard for Ethernet Amendment 10: Media Access Control Parameters, Physical Layers, and Management Parameters for 200 Gb/s and 400 Gb/s Operation.
2. J. Wei, Q. Cheng, R. V. Penty, I. H. White, and D. G. Cunningham, “400 Gigabit Ethernet Using Advanced Modulation Formats: Performance, Complexity, and Power Dissipation,” IEEE Commun. Mag. 53(2), 182–189 (2015). [CrossRef]
3. 800G Pluggable MSA White Paper, “Enabling the next generation of cloud & AI using 800 Gb/s optical module,” Mar. 2020. Available at https://www.800gmsa.com/news/white-paper
4. H. Yamazaki, M. Nagatani, H. Wakita, Y. Ogiso, M. Nakamura, M. Ida, T. Hashimoto, H. Nosaka, and Y. Miyamoto, “Transmission of 400-Gbps Discrete Multi-Tone Signal Using >100-GHz-Bandwidth Analog Multiplexer and InP Mach-Zehnder Modulator,” Proc. ECOC, PDP paper, (2018).
5. X. Chen, S. Chandrasekhar, S. Randel, G. Raybon, A. Adamiecki, P. Pupalaikis, and P. J. Winzer, “All-electronic 100-GHz Bandwidth Digital-to-Analog Converter Generating PAM Signals up to 190-GBaud,” Pro. OFC, Paper.Th5C. 5, (2016).
6. S. Kanazawa, H. Yamazaki, Y. Nakanishi, T. Fujisawa, K. Takahata, Y. Ueda, W. Kobayashi, Y. Muramoto, H. Ishii, and H. Sanjoh, “Transmission of 214-Gbit/s 4-PAM signal using an ultra-broadband lumped-electrode EADFB laser module,” Pro. OFC, Paper. Th5B.3, (2016).
7. S. Lange, S. Wolf, J. Lutz, L. Altenhain, R. Schmid, R. Kaiser, M. Schell, C. Koos, and S. Randel, “100 GBd Intensity Modulation and Direct Detection with an InP-Based Monolithic DFB Laser Mach-Zehnder Modulator,” J. Lightwave Technol. 36(1), 97–102 (2018). [CrossRef]
8. J. M. Estarán, H. Mardoyan, F. Jorge, O. Ozolins, A. Udalcovs, A. Konczykowska, M. Riet, B. Duval, V. Nodjiadjim, J.-Y. Dupuy, X. Pang, U. Westergren, J. Chen, S. Popov, and S. Bigo, “140 /180/204-Gbaud OOK Transceiver for Inter- and Intra-Data Center Connectivity,” J. Lightwave Technol. 37(1), 178–187 (2019). [CrossRef]
9. W. Wang, Z. Huang, B. Pan, H. Li, G. Li, J. Tang, and Y. Lu, “Demonstration of 214Gbps per lane IM/DD PAM-4 transmission using O-band 35GHz-class EML with advanced MLSE and KP4-FEC,” Proc. OFC, M4F.4,(2020).
10. J. M. Estarán, S. Almonacil, R. Rios-Muller, H. Mardoyan, P. Jenneve, K. Benyahya, C. Simonnneau, S. Bigo, J. Renaudier, and G. Charlet, “Sub-Baudrate Sampling at DAC and ADC: towards 200G per lane IM/DD Systems,” J. Lightwave Technol. 37(6), 1536–1542 (2019). [CrossRef]
11. N. Stojanovic, C. Prodaniuc, L. Zhang, J. Wei, S. Calabrò, T. Rahman, and C. Xie, “Four-Dimensional PAM-7 Trellis Coded Modulation for Data Centers,” IEEE Photonics Technol. Lett. 31(5), 369–372 (2019). [CrossRef]
12. J. L. Wei, C. Prodaniuc, L. Zhang, and N. Stojanovic, “225 Gb/s per Lambda IMDD PAM Short Reach Links Using Only Commercial Components,” Proc. OECC, PDP paper, (2018).
13. L. Zhang, J. L. Wei, C. Prodaniuc, N. Stojanovic, and C. Xie, “Beyond 200-Gb/s DMT transmission over 2-km SMF based on a low-cost architecture with single-wavelength, single-DAC/ADC and single-PD,” Proc. ECOC, Paper We1H.1, (2018).
14. J. L. Wei, L. Zhang, C. Prodaniuc, N. Stojanovic, and C. Xie, “Linear Pre-equalization Techniques for Short Reach Single Lambda 225 Gb/s PAM IMDD Systems,” Proc. ECOC, (2018).
15. Q. Zhang, N. Stojanovic, J. Wei, and C. Xie, “Single-lane 180 Gb/s DB-PAM-4-signal transmission over an 80 km DCF-free SSMF link,” Opt. Lett. 42(4), 883–886 (2017). [CrossRef]
16. S. Yamamoto, A. Masuda, H. Taniguchi, and Y. Kisaka, “92-Gbaud PAM4 Transmission Using Spectral-Shaping Trellis-Coded-Modulation with 20-GHz Bandwidth Limitation,” Proc. OFC, Paper W4I.5, (2019).
17. X. Pang, O. Ozolins, L. Zhang, A. Udalcovs, R. Lin, R. Schatz, U. Westergren, S. Xiao, W. Hu, G. Jacobsen, S. Popov, and J. Chen, “Beyond 200 Gbps per Lane Intensity Modulation Direct Detection (IM/DD)Transmissions for Optical Interconnects: Challenges and Recent Developments,” Proc. OFC, Paper W4I.7 (2019).
18. J. Wei, S. Calabrò, T. Rahman, and N. Stojanovic, “System aspects of the next-generation data-center networks based on 200G per lambda IMDD links,” Proc. SPIE 11308, Metro and Data Center Optical Networks and Short-Reach Links III, 1130805 (2020 invited paper).
19. T. Wettlin, S. Pachnicke, T. Rahman, J. Wei, S. Calabrò, and N. Stojanovic, “Complexity Reduction of Volterra Nonlinear Equalization for Optical Short-Reach IM/DD Systems,” ITG Symposium Photonic Networks, to be published (2020).
20. J. Wei, N. Stojanovic, L. Zhang, S. Calabrò, T. Rahman, C. Xie, and G. Charlet, “ Experimental comparison of modulation formats for 200 G/λ IMDD data center networks, “ Proc. ECOC, Paper Tu.3.D.2, (2019).
21. T. Wettlin, S. Calabro, T. Rahman, J. Wei, N. Stojanovic, and S. Pachnicke, “DSP for High-Speed Short-Reach IM/DD Systems using PAM,” J. Lightwave Technol. (to be published). [CrossRef]
22. Q. Hu, M. Chagnon, K. Schuh, F. Buchali, and H. Bülow, “IM/DD Beyond Bandwidth Limitation for Data Center Optical Interconnects,” J. Lightwave Technol. 37(19), 4940–4946 (2019). [CrossRef]
23. J. Wei, N. Stojanovic, and C. Xie, “Nonlinearity mitigation of intensity modulation and coherent detection systems,” Opt. Lett. 43(13), 3148–3151 (2018). [CrossRef]
24. Y. Fu, M. Bi, D. Feng, X. Miao, H. He, and W. Hu, “Spectral efficiency improved 2D-PAM8 Trellis coded modulation for short reach optical system,” IEEE Photonics J. 9(4), 1–8 (2017). [CrossRef]
25. P. S. Chow, J. M. Cioffi, and J. A. C. Bingham, “A practical discrete multitone transceiver loading algorithm for data transmission over spectrally shaped channels,” IEEE Trans. Commun. 43(2/3/4), 773–775 (1995). [CrossRef]
26. K. Zhang, Q. Zhuge, H. Xin, W. Hu, and D. V. Plant, “Performance comparison of DML, EML and MZM in dispersion-unmanaged short reach transmissions with digital signal processing,” Opt. Express 26(26), 34288–24304 (2018). [CrossRef]
