Performance Comparison of 112 Gb/s DMT, Nyquist PAM4 and Partial-Response PAM4 for Future 5G Ethernet-based Fronthaul Architecture

0733-8724 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JLT.2018.2790579, Journal of Lightwave Technology JOURN AL OF L A T E X CLASS FILES, V OL. 14, NO. 8, A UGUST 2015 1 Performance Comparison of 112 Gb/s DMT , Nyquist P AM4 and P artial-Response P AM4 for Future 5G Ethernet-based Fronthaul Architecture Nicklas Eiselt, Student Member , IEEE, Daniel Muench, Member , IEEE, Annika Dochhan, Member , IEEE, Helmut Griesser , Member , IEEE, Michael Eiselt, F ellow , OSA, Senior Member , IEEE, Juan Jos ´ e V e g as Olmos, Senior Member , IEEE, Idelfonso T afur Monro y , Senior Member , IEEE, and J oer g-P eter Elber s, Member , IEEE Abstract —F or a futur e 5G Ether net-based fr onthaul ar chitec- tur e, 100G tru nk lines of a transmission distance up to 10 km standard single mode fiber (SSMF) in c ombination with cheap gr ey optics to daisy chain cell site netw ork interfaces ar e a pr omising cost- and po wer -efficient solution. F or such a scenario, differ ent intensity modulation and dir ect detect (IMDD) f ormats at a data rate of 112 Gb/s, namely Nyquist f our -le v el pulse amplitude modulation (P AM4), discr ete multi-tone transmission (DMT) and partial-r esponse (PR) P AM4 ar e experimentally in v estigated, using a lo w-cost electr o-absor ption modulated laser (EML), a 25G dri v er and curr ent state-of-the-art high speed 84 GS/s CMOS digital-to-analog con v erter (D A C) and analog- to-digital con v erter (ADC) test chips. Each modulation f ormat is optimized independently f or the desir ed scenario and their digital signal pr ocessing (DSP) r equir ements ar e in v estigated. The perf ormance of Nyquist P AM4 and PR P AM4 depend v ery much on the efficiency of pr e- and post-equalization. W e sho w the necessity f or at least 11 FFE-taps f or pr e-emphasis and up to 41 FFE coefficients at the r ecei v er side. In addition, PR P AM4 r equir es an MLSE with f our states to decode the signal back to a P AM4 signal. On the contrary , bit- and po wer -loading (BL, PL) is crucial f or DMT and an FFT length of at least 512 is necessary . W ith optimized parameters, all modulation f ormats r esult in a v ery similar perf ormances, demonstrating a transmission distance of up to 10 km o v er SSMF with bit err or rates (BERs) belo w a FEC thr eshold of 4.4E-3, allo wing err or fr ee transmission. Index T erms —Digital Signal Pr ocessing, Optical Fiber Com- munication, Modulation. I . I N T R O D U C T I O N S EVERAL architect u r es for future 5G fronthaul netw orks are currently discussed. One potential approach is based on a w a v elength di vision multiple xing (WDM) passi v e optical netw ork (PON) [1], where multiple 10G feeds from lar ge cell sites are transported to the central of fice without an y statistical multiple xing. This is in particular interesting in en vironments The results were obtained in the frame w ork of the SEND A TE Secure-DCI, ICirrus and SpeeD projects, partly funded by the German ministry of education and research (BMBF) under contracts 16KIS0477K and 13N1374 and by the European Commission under grant agreement No. 644526 and in the Marie Curie project AB A CUS. N. Eiselt is with the Department of Photonics Engineering, T echnical Uni v ersity of Denmark and with AD V A Optical Netw orking SE in Meiningen, German y (neiselt@adv aoptical.com). D. Muench, A. Dochhan, H. Griesser , M. H. Eiselt and J.-P . Elbers are with AD V A Optical Netw orking SE in Meiningen and Martinsried, German y . J. J. V e g as Olmos and I. T afur Monro y are with the Department of Photonics Engineering, T echnical Uni v ersity of Denmark. Fig. 1. Ethernet-based fronthaul architecture for 5G. where fibers are a scarce resource or acti v e equipment has to be a v oided at intermediate sites. Another potential approach is based on Ethernet, which daisy chains cel l site netw ork interf ace de vices with one or fe w 100G trunk lines and uses statist ical multiple xing. Although this approach can be combined also with WDM, it is most beneficial in terms of po wer consumption, footprint and cost to transmit just a single w a v elength in combination with cheap gre y optics. As Ethernet data are transmitted, the required data rate is 103.125 Gb/s plus optional o v erhead. Fig. 1 sho ws the proposed Ethernet-based fronthaul architecture. P articular benefits from using Ethernet are its widespread- ness and ubiquity making Ethernet a cost-ef fecti v e technology . Furthermore, it allo ws in principle all necessary topologies, lik e star -topologies for aggre g ating, tree topologies for adding and dropping of cells, and ring topologies for pro viding higher a v ailability . In addition, it enables the reuse and adaptation of already a v ailable means for operation, administration and maintenance as well as for service le v el agreements or for self-optimizing netw orks. The aggre g at ion and treatment of multiple traf fic classes are also allo wed. And finally , as a transparent transport layer it is agnostic to an y ne w upcoming functional split between radio remote head (RRH) and base- band processing unit (BB U) [2]. Using CPRI (Common Public Radio Interf ace) o v er Ethernet (CPRIoE) mapper enables coe xistence with le g ac y CPRI-based equipment [3]. All these benefits come at the price of stringent requirements for latenc y , latenc y v ariation and time synchronization. One approach to address the time sensiti v eness is IEEE802.1CM, which discusses the application of e xisting means out of 0733-8724 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JLT.2018.2790579, Journal of Lightwave Technology JOURN AL OF L A T E X CLASS FILES, V OL. 14, NO. 8, A UGUST 2015 2 the toolbox for time-sensit i v e netw orking (IEEE802.1 TSN). Further constraints and assumptions for fronthaul netw orks are a t opology with at least 6 hops or switches and at least 20 nodes [4]. These stringent requirements limit the maximum achie v able distance of each 100G line. The allo wed one-w ay latenc y is 75 µs and bit error rates (BERs) of 1E-12 are required [5], [6]. Deri v ed from the lo w-cost requirements are the consid- erations of using intensity modulation and direct detection (IMDD) together with gre y optics for each 100G trunk line. T o mak e the system as simple as possible, optical amplification and dispersion compensation are not an option. Consequently , the 1300 nm transmission windo w is required to pre v ent se v ere limitations from chromatic dispe rsion. F or the modulation format, non-return to zero (NR Z) w ould be the simplest solu- tion, requiring ho we v er e xpensi v e high-bandwidth optics and electronics. Hence, adv anced modulation formats combined with DSP and FEC-encoding are considered, with potential candidates being P AM4 [7], [8], [9], duobinary/partial re- sponse P AM4 (PR P AM4) [10], [11], [12], DMT [13], [14] and carrier -less amplitude and phase modulation (CAP) [15]. Out of these, P AM4 and DMT ha v e also been hea vily dis- cussed during the standardization acti vities of the IEEE802.3bs 400 GbE T ask F orce for ne xt generation of intra-data center interconnects [16], resulting already in demonstrations of 100G real-time PHYs and of fer good opportunities for reuse [17], [18]. A thorough comparison between the dif ferent modulation formats and their required DSP is essential in order to find the optimum solution for the scenario of interest. In this paper , we e xperimentally e v aluate and compare Nyquist P AM4, PR P AM4 and DMT at a data rate of 112 Gb/s as potential modulation formats for the discussed transmission. T o sho w its commercial feasibility , current state-of-the-art high speed CMOS D A C and ADC test chips with 84 GS/s are used together with a lo w-cost EML and a 25G dri v er . W e demonstrate, that each format requires careful optimization, which at the end results in a v ery similar performance. A transmission distance of up to 10 km is successfully bridged. The paper is structured as follo ws. Section II presents the e xperimental setup and the used components. The applied modulation formats and their required DSP are introduced in section III and IV. In section V the e xperimental results are demonstrated, discussed and compared. A conclusion is dra wn in section VI. I I . E X P E R I M E N T A L S E T U P Fig. 2 sho ws the emplo yed e xperimental setup. Of fline DSP is applied at the transmitter as well as at the recei v er side and requires the use of a high-resolution D A C and ADC. Both operate at a sampling speed of 84 GS/s, ha v e a nominal bit resolution of 8 bit and sho w a 3-dB bandwidth of around 15 GHz and 18 GHz, respecti v ely . At the transmitter , the dif ferential outputs of the D A C are amplified by a linear , dif ferential input and single-ended output modulator dri v er amplifier (D A, MA OM-003115) dri ving directly the succeed- ing electro-absorption modulated laser (EML) [19]. This D A Fig. 2. Experimental transmission setup. Fig. 3. Characteristics of the EML: a) measured output spectrum (unmodu- lated) and b) meas ured transfer function at a DFB-current of 100 mA and a temperature of 45 ◦ C. has inte grated high frequenc y coils for biasing the EML, it pro vides an adjustable g ain of up to 9 dB with a maximum output swing of 2 V and a 3-dB bandwidth of around 25 GHz. The EML operates at a fix ed temperature of 45 ◦ C and the current of the distrib uted feedback laser (DFB) section is set to a maximum of 100 mA, which gi v es the highest linear range and the highest optical output po wer . At these operating conditions, the transmission w a v elength of the EML is around 1308 nm and the 3-dB bandwidth is measured to be around 27 GHz b ut with a smooth roll-of f. Fig. 3a) illustrates the unmodulated output spectrum of the EML, while Fig. 3b) sho ws the measured optical po wer vs. EML bias v oltage. F or transmission, a bias of -1.25 V w as found to be optim um for all modulation formats and is therefore used throughout this paper . The optical link setup consists of a con v entional SSMF with an attenuation of around 0.32 dB/km at 1300 nm, a v ariable optical attenuator (V O A) with an inte grated po wer monitor and a PIN-photodetector (Picometri x PT -40E) inte grated with a linear trans-impedance amplifier (PIN/TIA) with a combined bandwidth of 35 GHz. Finally , the signal is sent back to the ADC and stored for of fline processing. As the memory of this ADC is limited to 500000 sam ples, se v eral blocks of data are used to mak e a v alid statement of the bit error rate (BER) performance. I I I . M O D U L A T I O N S C H E M E S A N D T H E I R D S P A. Discr ete Multi-tone T r ansmission DMT as a special v ariant of orthogonal frequenc y di vision multiple xing (OFDM) emplo ys the properties of Hermitian symmetry and the IFFT to create a real-v alued signal with the frequenc y spectrum di vided into orthogonal subcarriers. Each subcarrier can be modulated and the po wer of each 0733-8724 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JLT.2018.2790579, Journal of Lightwave Technology JOURN AL OF L A T E X CLASS FILES, V OL. 14, NO. 8, A UGUST 2015 3 Fig. 4. DSP blocks of the implemented DMT system. subcarrier can be allocated based on the w ater filling method. This process is kno wn as bit and po wer loading (BL, PL) and enables the ef fecti v e compensation of channel impairments and component bandwidth limitations without applying com- ple x signal processing, e.g. a simple 1-tap equalizer at the recei v er side is ef ficient. T o apply BL and PL, the transfer function of the transmission system is first estimated in terms of the signal-to-noise ratio (SNR) at the recei v er with 16- QAM constellations with equal po wer on each subcarrier . Afterw ards, Cho w’ s mar gin-adapti v e bit loading algorithm and Ciof fi’ s po wer loading are applied to ef ficiently distrib ute the bits and allocate t he po wer [20]. Fig. 5 sho ws the estimated SNR of the transmission setup for the optical back-to-back case, together with the corresponding bit and po wer allocation for a 112 Gb/s DMT si g na l. From the estimated SNR, we can also estimate the a v ailable bandwidth of the transmission system: a n SNR of 15 dB or more is a v ailable up to 25 GHz, while it drops belo w 0 dBm for frequencies abo v e 30 GHz. What is more, the EML e xhibits a frequenc y drop at around 7 GHz in its transfer function, de grading the transmission performance. The c lock-line of the D A C and ADC e xplains the frequenc y null at 21 GHz. Fig. 4 illustrates the DSP blocks of the analyzed DMT system and T able I summarizes the most important DMT system parameters. T o meet the mentioned memory requirements of the D A C and the ADC, a DMT frame consists of 124 data symbols and four training symbols (in total 128 DMT symbols), which are used for channel estimation and synchronization. B. Nyquist P AM4 F our -le v el pulse amplitude m odu l ation (P AM4) encodes 2- bits into one symbol, resulting in a four -le v el signal and reduc- ing the transmission bandwidth by a f actor of tw o compared to on-of f-k e ying. Utilizing Nyquist pulse shaping with a small roll-of f f actor ( β ≈ 0 . 1 ), the signal bandwidth can be further reduced, resulting in an electrical bandwidth of around 30 GHz for a 112 Gb/s P AM4 signal. Fig. 6 sho ws the implemented of fline DSP blocks for the Nyquist P AM4 system. A 4-ary deBruijn sequence of order eight ( 4 8 = 65536 symbols) is used and gray-mapped onto a P AM4 signal . Compared to DMT , P AM4 of fers the possibility to easily compensate the nonlinear transfer function of the modulator by adjusting the le v els to w ar ds equally spaced po wer le v els after the modulator . Afterw ards, the signal is Fig. 5. a) Estimated SNR per subcarrier at the recei v er for optical back-to- back and the corresponding b) bit and c) po wer loading. T ABLE I D M T S Y S T E M P A R A M E T E R S Modulation formats BPSK to 64-QAM Frame length (data symbols) 124 T raining symbols (TS) 4 FFT length 512 Usable carriers 255 (max. used 242) Cyclic prefix (CP) 1/64 Clipping ratio to be optimized Equalizer decision directed, 1-tap Fig. 6. DSP-flo w of the implemented Nyquist P AM4 system. upsampled to 3 samples/symbol, raised cosine shaped in frequenc y domain with β = 0 . 1 , and do wnsampled by a f actor of tw o in order to achie v e an o v ersampling of 1.5 samples per symbols to generate a 112-Gb/s Nyquist-P AM4 signal with the 84 GS/s D A C. The subsequent digital pre-emphasis compensates the bandwidth limitations of the D A C and dri v er , and the signal is quantized into inte ger v alues between 0 to 255 to use the full 8-bit resolution of the D A C. Fig. 9a) illustrates the obtained e ye diagrams after the dri v er amplifier as well as after the EML, e xhibiting the typical o v er -and undershoots of a Nyquist P AM4 signal. Furthermore, the po wer spectrum density (PSD) of the transmit-signal before the D A C demonstrates the ef fect of pre-emphasis (gray area = uncompensated signal). At the recei v er , the signal is resampled to 2 samples/symbol and for the sak e of con v enience an ideal clock-reco v ery by means of a Gardner phase detector is assumed [21]. The 0733-8724 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JLT.2018.2790579, Journal of Lightwave Technology JOURN AL OF L A T E X CLASS FILES, V OL. 14, NO. 8, A UGUST 2015 4 Fig. 7. DSP blocks of the implemented PR P AM4 system. detector output o v er the phase error (S-curv e) is calculated and a v eraged o v er the whole block of recei v ed data to determine the optimum phase shift. Subsequently , an adapti v e symbol- spaced equalizer is applied to reco v er the P AM4 signal. T w o dif ferent equalizer structures are e v aluated: a simple adapti v e feedforw ard equalizer (FFE) based on the least-mean square (LMS) algorithm and a combination of FFE and maximum lik elihood sequence estimation (MLSE). F or the latter , the MLSE basically replaces the hard decision after the FFE. C. P artial-Response P AM4 This paper performs partial-response encoding in the digital domain at the transmitt er -side by passing the mapped P AM4 symbols s ( k ) through a simple delay-and-add filter s ( k ) = s ( k ) + s ( k − 1) . (1) This operation correlates adjacent symbols and redistrib utes the PSD such, that it is more concentrated at lo wer frequencies. Furthermore, an input P AM4 signal results in a se v en-le v el signal at the output, which is also referred to duobinary P AM4 [11], [22]. Usually , the correlation requires pre-coding prior to the delay-and-add operation in order to a v oid error - propag ation at the recei v er side. Ho we v er , applying an MLSE at the recei v er side tak es the correlation into account for a decision, a v oiding the need for pre-coding. Fig. 7 sho ws the DSP bl ocks of the implemented PR P AM4 system, which resembles with some e xceptions the P AM4 system. The described PR encoding is done after the P AM4 mapping and the le v el adjustment is no w based on the obtained se v en le v els. Subsequently , the signal is resampled to 1.5 samples/symbols and also shaped with a Raised-Cosine filter of β = 0 . 1 . Note, also PR P AM4 requires frequenc y domain pulse-shaping in order to a v oid Aliasi ng . Ag ain, pre-emphasis is applied and the signal is quantized before loading it into the D A C-memory . Fig. 9b) depicts the obtained PR-P AM4 e ye diagrams after the dri v er and the EML as well as the PSD with and without pre-emphasis. The e ye diagrams sho w clearly the se v en le v els due to partial response encoding. At the recei v er , the signal is first resampled to 2 sam- ples/symbol, clock-reco v ery is applied and a combination of FFE and MLSE is emplo yed to detect the symbols, both op- erating at one sample per symbol. Ag ain, t he FFE-coef ficient calculation is based on the LMS algorithm, b ut with a se v en- le v el hard decision for error estimation to reco v er the PR P AM4 signal. Fig. 8. Experimental setup and DSP-flo w of the pre-emphasize algorithm. In the inset the obtained time domain filter coef ficients are sho wn. Fig. 9. Nyquist P AM4 and PR-P AM4 e ye diagrams obtained after the dri v er amplifier and after the EML at 112 Gb/s emplo ying digital pre-equalization. The PSDs of the transmit-signal sho w the ef fect of pre-emphasis. I V . D I G I T A L P R E - E M P H A S I S F O R S I N G L E C A R R I E R M O D U L A T I O N T o operate at 100 Gb/s and be yond in combination with higher order modulation formats, it becomes ine vitable to consider the limitations of the transmitter components such as D A C, dri v er , electric al cables and modulator . While DMT of fers the ability to do BL and PL to ef ficiently compensate channel distortions, single carrier modulation formats require the use of ef ficient pre- and post-equalization to mitig ate the performance limitations at the transmitter and recei v er side. This w ork uses the indirect learning architecture to calculate 0733-8724 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JLT.2018.2790579, Journal of Lightwave Technology JOURN AL OF L A T E X CLASS FILES, V OL. 14, NO. 8, A UGUST 2015 5 the in v erse transfer function of the transmitter components consisting of D A C, dri v er amplifier and the electrical ca- bles [23], [24]. Fig. 8 depicts the e xperimental setup. F or the estimation, a 4-ary deBruijn sequence of order eight is directly loaded onto the 84 GS/s D A C and digitized after transmission with a real-time scope of 70 GHz bandwidth and a sampling rate of 200 GS/s. The use of such high bandwidth oscillo- scope ensures that only the transmitter component bandwidth distortions are tak en into account. In digital domain, after resampling and clock-reco v ery , the pre-equalize r coef ficients are calculated b y comparing the recei v ed sequence with the transmitted one using the LMS algori thm. The inset of Fig. 8 displays the obtained equalizer coef ficients (in this case 61 coef ficients) as well as the corresponding frequenc y domain beha viour . Comparing the pre-emphasis for both formats, PR P AM4 requires a weak er pre-emphasis as it concentrates more ener gy at lo wer frequencies. The stronger pre-emphasis for Nyquist P AM4 with up to 10 dB amplification of the higher frequencies comes at the cost of an increased peak-to-a v erage- po wer ratio (P APR). This decreases the dynamic range of the D A C, resulting in a smaller amplitude after the D A C and dri v er and thus, a smaller optical modulation amplitude (OMA) and e xtinction ratio as demonstrated by the e ye diagrams of Fig. 9 and by the optical spectrum of Fig. 17. F or instance, we measured an e xtinction ratio (highest to lo west le v el) of around 4 dB for Nyquist P AM4 and an e xtinction ratio of around 7 dB for PR P AM4. V . E X P E R I M E N T A L R E S U L T S A N D D I S C U S S I O N A. DMT Each format needs to be optimi zed indi vidually in order to ensur e a f air comparison between the dif ferent modulation formats. F or DMT , se v eral parameters such as clipping ratio, the FFT length, number of trai ning symbols, etc. ha v e to be optimized at first. T o not go be yond the scope of this paper , some parameters with minor influence on the performance such as number of training symbols or the c yclic prefix are set to fix ed v alues based on pre vious e xperiences/publications [25]. Fig. 10a) demonstrates the influence of dif ferent clipping ratios using an FFT length of 512, re v ealing an optimum of around 15 dB for this particular transmission scenario. The influence of the FFT length is sho wn in Fig. 10b) using a fix ed clipping ratio of 15 dB and a fix ed CP of 1/64 of the symbol duration. A longer FFT increases the granularity and enables a more precise adaption of the signal to the channel characteristics, resulting in an impro v ed performance. Ho we v er , the performance g ain by a longer FFT decreases as only little impro v ement is achie v ed with an FFT length of 2048 compared to an FFT length of 512, whereas the implementation comple xity of DMT raises proportionally to N FFT log 2 ( N FFT ) . Hence, the use of an FFT length of 512 is a good trade-of f between implementation comple xity and BER- performance [25]. Fig. 11 sho ws the achie v ed BER vs. recei v ed optical po wer (R OP) into the PIN/TIA for dif ferent data rat es as well as for dif ferent transmission distances. Since DMT allo ws to easily switch between dif ferent data rates by loading a dif ferent Fig. 10. Optimization of DMT parameters: a) influence of clipping ratio at an FFT -length of 512 and b) influence of FFT length at a fix ed clipping ratio of 14 dB. Fig. 11. T ransmission results of DMT at dif ferent data rates and for dif ferent transmission distances. number of bits onto each subcarrier , the performance of 112 Gb/s, 89.4 Gb/s, 74.7 Gb/s and 56 Gb/s is in v estig ated. Note that 89.4 Gb/s or 74.7 Gb/s are no Ethernet data rates and a rate of 56 Gb/s w ould require tw o w a v elengths to achie v e 100G transmission. Ne v ertheless these rates are illustrati v e in sho wing the increasing performance penalties introduced by the limited bandwidth of the system. T w o dif ferent FEC thresholds are added as a solid and a dashed line, representing the standardized KP4-FEC (RS(544,514,10)) with a BER- limit of 2E-4 and the continuously-interlea v ed BCH FEC (CI- BCH(1020,988)) with a BER-limit of 4.4E-3 [26], [27]. T rans- mitting at a data rate of 112 Gb/ s, BERs belo w the CI-BCH FEC are achie v ed only for the optical back-to-back (b2b) case, while BERs around the FEC threshold are achie v ed in case of 10 km transmission distance. At a bias of 1.25 V , the output po wer of the deplo yed EML is around 1 dBm, which results in a maximum achie v able input po wer of − 5 . 5 dBm after 20 km transmission at this w a v elength ( 0 . 32 dB/km ∗ 20 km). This input po wer is not suf ficient to achie v e BERs belo w the FEC- limits in case of 112 Gb/s. Indeed, a v ery similar performance is demonstrated for the dif ferent transmission distances up to the achie v able input po wers. The performance impro v es with decreasing bitrate. 0733-8724 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JLT.2018.2790579, Journal of Lightwave Technology JOURN AL OF L A T E X CLASS FILES, V OL. 14, NO. 8, A UGUST 2015 6 Fig. 12. Optical back-to-back transmission results of 112 Gb/s Nyquist P AM4 emplo ying dif ferent numbers of pre- and post-FFE coef ficients: a) sho ws the optical e ye diagrams obtained directly after the EML using a pre-equalizer of 5 coef ficients and 61 coef ficients, b) and c) illustrate the BER vs. R OP results using dif ferent numbers of post-FFE coef ficients and d) depicts the BER performance for dif ferent Tx-FFE/Rx-FFE combinations at R OP=0 dBm. Fig. 13. T ransmission results of 112 Gb/s Nyquist P AM4: a) o v er dif ferent transmission distances using only 11 Tx-FFE and 41 Rx-FFE coef ficients and b) optical back-to-back transmission applying dif ferent MLSE memory length after the FFE. B. Nyquist P AM4 F or Nyquist P AM4, the performance of the pre-equalizer (Tx-FFE) in combination with the appli ed FFE at the re- cei v er (Rx-FFE) is e v aluated in terms of BER-performance in a first step. In principle, we w ould lik e to answer the question, ho w man y Tx-FFE and Rx-FFE coef ficients are necessary for such a transmission scenario to achie v e BERs belo w the desired FEC threshold. Fig. 12b) and c) depict the BER vs. R OP results for optical b2b, using 5 and 61 Tx- FFE coef ficients, respecti v ely , in combination with a dif ferent number of applied Rx-FFE coef ficients. The number after the term ”FFE” notates the number of used coef ficients. Ag ain, the pre viously discussed FEC thre sholds are sho wn as black lines. The interaction between the amount of Tx-FFE and Rx-FFE coef ficients is further illustrated as a contour plot in Fig. 12d), demonstrating the achie v ed BERs for dif ferent Tx- FFE/Rx-FFE combinations at a fix ed R OP=0 dBm. Basically , no significant BER impro v ement is seen with more than 11 Tx-FFE coef ficients, while at the recei v er at least 21 Rx- FFE coef ficients are required. T o achie v e BERs belo w 1E-3 ho we v er , more than 40 coef ficients for both Tx-FFE and Rx- FFE are necessary . Based on the results of Fi g. 12, 11 Tx-FFE and 41 Rx-FFE coef ficients of fer a good tradeof f between performance and comple xity and are used for further e v aluation. W ith these settings, the performance for optical b2b, 10 km and 20 km is compared in Fig. 13a). F or optical b2b and for 10 km SSMF , the results stay well belo w the CI-BCH FEC threshold, ho we v er , the KP4-FEC threshold is not reached. In addition, the limited output po wer of the EML pre v ents a possible transmission o v er 20 km. Up to the achie v able input po wer a similar performance of the dif ferent transmission distances is sho wn. Fig. 13b) illustrates the performance of dif ferent FFE-MLSE combinations for optical b2b . Ag ain, 11 Tx-FFE coef ficients are used for pre-emphasis. The number after ”MLSE” indicates the used memory , e.g. an MLSE1 indicates a memory m of one, which corresponds to 4 m = 4 states for the V iterbi algorithm. Although applying a m emory length of up to three, which corresponds to 64 states, only little performance impro v ement is achie v ed at higher input po wers. C. P artial-Response P AM4 Fig. 14 sho ws the transmission results of PR P AM4 at 112 Gb/s. Ag ain, the BER performance in dependence of the number of Tx-FFE and Rx-FFE coef ficients is e v aluated in a first step. Note, that an MLSE of memory one is emplo yed after the Rx-FFE, in order to detect the PR P AM4 signal (see section III-C). A significant performance impro v ement is demonstrated also for PR P AM4 when suf ficient pre-emphasis is applied. BERs belo w the CI-BCH FEC-limit are not achie v- able without an y pre-emphasis. Based on the contour plot of Fig. 14c), around 11 Tx-FFE and 21 Rx-FFE coef ficients are suf ficient, as no further significant performance impro v ement is achie v ed. W ith these settings, BERs belo w 4E-4 can be achie v ed at an optimum input po wer of -1 dBm. At higher input po wers the PIN/TIA starts to saturate, compressing the outer le v els of the signal and resulting in a de gradation of performance. The obtained histograms after the FFE in the inset of Fig. 14b) demonstrate this ef fect, as the outer le v els ha v e a smaller distance at higher R OPs. 0733-8724 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JLT.2018.2790579, Journal of Lightwave Technology JOURN AL OF L A T E X CLASS FILES, V OL. 14, NO. 8, A UGUST 2015 7 Fig. 14. Optical back-to-back transmission results of 112 Gb/s PR P AM4 emplo ying di f ferent numbers of pre- and post-FFE coef ficients: a) and b) illustrate the BER vs. R OP results using dif ferent numbers of post-FFE coef ficients and d) depicts the BER performance for dif ferent Tx-FFE/Rx-FFE combinations at R OP= − 1 dBm. Fig. 15. T ransmission results of 112 Gb/s PR P AM4: a) using dif ferent MLSE memory length after the FFE in case of optical back-to-back transmission, b) o v er dif ferent transmission distances using 11 Tx-FFE and 21 Rx-FFE coef ficients. Fig. 16. T ransmission results of 112 Gb/s DMT , Nyquist P AM4 and PR P AM4 for a) optical b2b and b) 10 km. Fig. 15a) e v aluates the performance using dif ferent memory length for the MLSE. Similar to Nyquist P AM4, a significant performance impro v ement is not obtained with a longer mem- ory length. BERs around the KP4-FEC threshold are achie v ed with an MLSE3, which impro v es the performance from 4E- 4 do wn to 2E-4. This ”small” impro v ement does not justify the use of such comple x equalizat ion and thus, an MLSE1 is used to e v aluate the performance up to a transmission distance of 20 km. Finally , Fig. 15b) sho ws the performance for b2b, 10 km and 20 km. A transmission distance of up to 10 km is possible assuming the CI-BCH FEC threshold. D. A comparison of DMT , Nyquist P AM4 and PR P AM4 Finally , Fig. 16 compares the performance of 112 Gb/s DMT , Nyquist P AM4 and PR-P AM4 for optical b2b and for a transmission distance of 10 km. The optimum parameters found in the pre vious sections are applied for each modulation format: DMT emplo ys an FFT length of 512, Nyquist P AM4 utilizes 11 Tx-FFE and 41 Rx-FFE coef ficients and PR P AM4 uses 11 Tx-FFE and 21 Rx-FFE coef ficients together with an MLSE1. F or the optical b2b case, PR P AM4 sho ws the best transmission performance as BERs belo w 4E-4 are achie v able. Ho we v er , one has to k eep in mind, that PR P AM4 uses an additional delay-and-add filter at the transmitter side and an MLSE at the recei v er side, to generate and detect the partial- response signal. This is not necessary for Nyquist P AM4. DMT and Nyquist P AM4 sho w v ery similar results, only at R OPs of more than 0 dBm Nyquist P AM4 achie v es lo wer BER v alues. The results re v eal further , a higher sensiti vity of PR P AM4 to w ards the nonlinear beha viour of the PIN/TIA, resulting in a significant performance de gradation at higher input po wer v alues. This is also due to the f act, that the used PIN/TIA saturates f aster for a signal e xhibiting a higher OMA and e xtinction ratio. Fig. 17 sho ws the optical po wer spectrum obtained after the EML for the dif ferent modulation schemes, which on the hand re v eals a significant hi gher e xtinction for PR P AM4 and on the other hand v ery similar spectral shapes for Nyquist P AM4 and DMT . F or the transmission distance of 10 km nearly no performance dif ferences are seen an y more between the dif ferent modulation formats. F or all three modulati on formats, the CI-BCH FEC is required to transmit o v er 10 km SSMF . The latenc y of this FEC is 1 Mbits, which leads to a processing delay of around 10 µ s at 100 Gb/s [27]. The latenc y of the DSP is mainly determined by the number of equalize r taps for P AM4 or the length of the FFT for DMT . Supposing an ASIC clock of 0733-8724 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JLT.2018.2790579, Journal of Lightwave Technology JOURN AL OF L A T E X CLASS FILES, V OL. 14, NO. 8, A UGUST 2015 8 Fig. 17. Optical spectra of the dif ferent modulation formats. 1 GHz and that 56 symbols are a v ailable per ASIC clock (56 GHz/1 GHz), an FFE could process 56 symbols at the same time. The latenc y calculation for an N -tap FFE includes time for N multiplications and K =ceil(log2( N -1)) additions, which w ould lead to a minimum latenc y of around 2 ns for Nyquist P AM4 with 11 Tx-FFE taps and 41 Rx-FFE taps (1 ns for Tx-FFE and 1 ns for Rx-FFE). If one multipl ication and six additions cannot be performed in one clock period, e.g. when only a single operation is done in 1 ns, the w orst case latenc y of the FFE w ould be 12 ns. F or PR P AM4 an additional MLSE1 is required, which is here assumed to process 56 decoded symbols per single block in parallel plus 10 o v erhead symbols (the o v erhead of the MLSE is usually 5xMLSE- memory , which requires in this case 5 symbols before and after the main symbol for PR P AM4). The MLSE processing time includes one multiplication to calculate the branch metric, one addition to sum the branches and state metrics, tw o add- compare-select operations to select one candidate per state and one decoding update. Supposing fur ther that a 1 GHz ASIC can process 4 symbols per clock, the MLSE1 w ould add around 16 ns of latenc y . Ag ain, for w orst case, when one clock c ycle per symbol is required t he MLSE1 latenc y is 66 ns. The o v erall processing delay of the DSP e v en in w orst case is therefore ne gligible compared to the latenc y of the FEC. Estimating in addition the propag ation delay with 5 µ s/km, the transmission distance of 10 km leads to an o v erall latenc y of around 60 µ s, which is well belo w the allo wed one-w ay latenc y of 75 µ s mentioned at the be ginning. The achie v ed results demonstrate at best a mar gin of around 2 dB with this FEC. P art of the performance penalties are due to the limited bandwidth of the electrical components (D A C, ADC and dri v er), which we e xpect to be increased in future. The EML of fe red suf ficient bandwidth, ho we v er it e xhibited a frequenc y dip at around 7 GHz, which unfortunately de graded the transmission performance. V I . C O N C L U S I O N W e ha v e e xperimentally in v estig ated DMT , Nyquist P AM4 and partial-response P AM4, all at a data rate of 112 Gb/s as candidate modulat ion formats for an Ethernet-based fronthaul architecture for future 5G with 100G trunk lines between dif ferent cell sites. Emplo ying current state-of-the-art 28 nm CMOS D A C and ADC in combination with a lo w-cost EML and a 25G dri v er , a transmission distance of 10 km SSMF is successfully bridged with all modulation formats if an HD- FEC of 4.4E-3 is assumed, allo wing error free transmission. Each modulat ion format requires careful optimization, result- ing in a v ery similar performance. F or Nyquist P AM4 and PR P AM4 the performance depends v ery much on the applied pre- and post-equalization, re v ealing the necessity of at least 11 pre-equalizer and 21 post-equalizer FFE coef ficients. On the other hand, BL and PL is essential for DMT and an FFT length of 512 is required. R E F E R E N C E S [1] J. Zou, C. W agner , and M. Eiselt, “Optical Fronthauling for 5G Mobile: A Perspecti v e of P assi v e Metro WDM T echnology, ” in Optical F iber Communication Confer ence , OFC , Los Angeles, CA, USA, 2017, paper W4C.2. [2] “IEEE 1914.1 P ack et-based Fronthaul T ransport Netw ork, ” 2016. [Online]. A v ailable: http://sites.ieee.or g/sagroups-1914/p1914-1/ieee- p1914-1-draft-specifications/ [3] “IEEE 1914.3 Draft 1.2 Radio o v er Ethernet Encapsulations and Mappings, ” 2016. [Online]. A v ailable: http://sites.ieee.or g/sagroups- 1914/p1914-3/ieee-p1914-3-draft-specifications/ [4] H. Jinri and Y . Y annan, “White P aper of Ne xt Generation Fronthaul Interf ace, ” White P aper , 2015. [5] Nokia, “Ev olution to centralized RAN with mobile fronthaul, ” T ec hnical White P aper , 2016. [6] “iCirrus D3.2 Preliminary Fronthaul Architecture Proposal, ” 2016. [Online]. A v ailable: http://www .icirrus-5gnet.eu/cate gory/deli v erables/ [7] M. Chagnon, M. Osman, M. Poulin, C. Latrasse, J.-F . Gagn ´ e, Y . P ainchaud, C. P aquet, S. Lessard, and D. Plant, “Experimental study of 112 Gb/s short reach transmission emplo ying P AM formats and SiP intensity modulator at 1.3 µ m.” Optics Expr ess , v ol. 22, no. 17, pp. 21 018–36, 2014. [8] N. Kikuchi and R. Hirai, “Intensity-modulated/direct-detection (IM/DD) Nyquist pulse-amplitude modulation (P AM) signaling for 100-Gbit/s/ λ optical short-reach transmission, ” in Eur opean Confer ence on Optical Communication, ECOC , Cannes, France, 2014, paper P .4.12. [9] Y . Gao, J. C. Cartledge, S. S. Y am, A. Rezania, and Y . Matsui, “112 Gb/s P AM-4 Using a Directly Modulated Laser with Linear Pre- Compensation and Nonlinear Post-Compensation, ” in Eur opean Confer - ence on Optical Communication, ECOC , Duesseldorf, German y , 2016, paper M2.C2. [10] N. Stojano vic, Z. Qiang, C. Prodaniuc, and F . Karinou, “Performance and DSP comple xity e v aluation of a 112-Gbit/s P AM-4 transcei v er emplo ying a 25-GHz T OSA and R OSA, ” in Eur opean Confer ence on Optical Communication (ECOC) , V alencia, Spain, 2015, paper T u.3.4.5. [11] L. F . Suhr , J. J. V . Olmos, B. Mao, X. Xu, G. N. Liu, and I. T . Monro y , “112-Gbit/s x 4-Lane Duobinary-4-P AM for 400GBase, ” in Eur opean Confer ence on Optical Communication, ECOC , Cannes, 2014, p. T u.4.3.2. [12] X. Xu, E. Zhou, G. N. Liu, T . Zuo, Q. Zhong, L. Zhang, Y . Bao, X. Zhang, J. Li, and Z. Li, “ Adv anced modulation formats for 400-Gbps short-reach optical inter -connection, ” Optics Expr ess , v ol. 23, no. 1, p. 492, 2015. [13] C. Xie, P . Dong, S. Randel, D. Pilori, P . W inzer , S. Spig a, B. K ¨ ogel, C. Neume yr , and M.-C. Amann, “Single-VCSEL 100-Gb/s Short-Reach System Using Discrete Multi-T one Modulation and Direct Detection, ” in Optical F iber Communication Confer ence , Los Angeles, CA, USA, 2015, paper T u2H.2. [14] Y . Kai, M. Nishihara, T . T anaka, T . T akahara, Lei Li, Zhenning T ao, Bo Liu, J. Rasmussen, and T . Drenski, “Experimental comparison of pulse amplitude modulati on (P AM) and discrete multi-tone (DMT) for short-reach 400-Gbps data communication, ” in Eur opean Confer ence on Optical Communication (ECOC) , 2013, paper Th.1F .3. [15] M. I. Olmedo, T . Zuo, J. B. Jensen, Q. Zhong, X. Xu, S. Popo v , and I. T . Monro y , “Multiband carrierless amplitude phase modulation for high capacity optical data links, ” J ournal of Lightwave T ec hnolo gy , v ol. 32, no. 4, pp. 798–804, 2014. [16] [Online]. A v ailable: http://www .ieee802.or g/3/bs/ 0733-8724 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JLT.2018.2790579, Journal of Lightwave Technology JOURNAL OF L A T E X CLASS FILES, VOL. 14, NO. 8, A UGUST 2015 9 [17] D. Lewis, S. Corbeil, and B. Mason, “Practical Demonstration of live- traffic Optical DMT Link using DMT T est Chip. ” [Online]. A vailable: http://www .ieee802.org/3/bs/public/14 09/le wis 3bs 01a 0914.pdf [18] F . Caggioni, “100G single Lambda Optical link, experimental data, ” in IEEE Interim Meeting - Dallas , Sep. 2016. [Online]. A vailable: http://www .ieee802.org/3/cd/public/Sept16/caggioni 3cd 01 0916.pdf [19] U. Troppenz, M. Narodovitch, C. K ottke, G. Przyrembel, W . D. Molzow , A. Sigmund, H. G. Bach, and M. Moehrle, “1.3 µ m electroabsorption modulated lasers for P AM4/P AM8 single channel 100 Gb/s, ” in Interna- tional Conf. on Indium Phosphide and Related Materials , Montpellier, France, 2014, paper Th.B2.5. [20] J. M. Ciof fi, “Data T ransmisssion Theory: Course T ext for EE379A-B and EE479, ” in Multi-channel modulation , Stanford Univ ersity , Stanfort, CA, USA, 2015. [Online]. A v ailable: http://www .stanford.edu/group/ciof fi/book [21] F . M.Gardner, Phaselock T echniques . W iley-Interscience, 1980. [22] S. W alklin and J. Conradi, “Multilevel signaling for increasing the reach of 10 Gb/s lightwave systems, ” Journal of Lightwave T echnology , vol. 17, no. 11, pp. 2235–2248, 1999. [23] G. Khanna, B. Spinnler, S. Calabr ` o, E. De Man, and N. Hanik, “A Robust Adaptive Pre-Distortion Method for Optical Communication T ransmitters, ” IEEE Photonics T echnolo gy Letters , vol. 28, no. 7, pp. 752–755, 2016. [24] C. Eun and E. J. Po wers, “A new V olterra predistorter based on the indirect learning architecture, ” IEEE Tr ansactions on Signal Processing , vol. 45, no. 1, pp. 223–227, 1997. [25] A. Dochhan, H. Griesser , N. Eiselt, M. Eiselt, and J.-P . Elbers, “Opti- mizing Discrete Multi-tone T ransmission for 400G Data Center Inter - connects Results an Discussion Launch power, ” ITG Symp. Photonic Networks , 2016. [26] “IEEE P802.3bs/D1.2 Draft Standard for Ethernet Amendment: Media Access Control Parameters, Physical Layers and Management Parameters for 400 Gb/s Operation, ” pp. 1–269, 2016. [Online]. A v ailable: http://www .ieee802.org/3/bs/ [27] M. Scholten, T . Coe, and J. Dillard, “Continuously-interleav ed BCH (CI-BCH) FEC delivers best in class NECG for 40G and 100G metro applications, ” in 2010 Conference on Optical Fiber Communication (OFC/NFOEC), collocated National Fiber Optic Engineers Confer ence , 2010, paper NTuB3.