real-time digital signal processing for optical ofdm-based future optical access networks

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 32, NO. 4, FEBRUARY 15, 2014 553

Real-time Digital Signal Processing for OpticalOFDM-Based Future Optical Access Networks

Roger Giddings

(Invited Tutorial)

Abstract—In this paper, key aspects associated with realizingreal-time digital signal processing (DSP)-based optical transceiversfor use in future optical access networks are considered. The fun-damental principles of optical orthogonal frequency division mul-tiplexing (OOFDM)-based transceivers and their associated DSPimplementation are presented. This is followed by an extensive re-view of the real-time DSP implemented in the end-to-end OOFDMtransceivers experimentally demonstrated over the past severalyears.

Index Terms—Digital signal processing (DSP), optical access,optical communication, orthogonal frequency division multipleaccess (OFDMA), orthogonal frequency division multiplexing(OFDM), passive optical network (PON).

I. INTRODUCTION

D IGITAL signal processing (DSP) is widely exploited in themodern world to enable a vast array of high performance

services and devices that were unimaginable several years ago.As a highly pervasive technology, DSP considerably enhanceseveryday life by enabling applications ranging from antilockbreaking systems to satellite navigation and sophisticated med-ical imaging. DSP has also been an enabler for many of thehighly successful communications technologies over the last20 years [1]–[3].

It is only in recent years that advanced DSP has been utilizedin optical communications to realize commercial long-haul op-tical systems in the form of DSP-enable coherent optical re-ceivers [4], which not only offer high transmission capacitiesof the order of 100 Gb/s per wavelength, but achieve ultrasen-sitive receivers for radically increasing unrepeated transmissiondistances.

To enable the wide use of DSP in other areas of optical com-munications, there is growing interest in the exploitation ofDSP to solve the challenges facing the future optical accessnetworks. The passive optical network (PON) [5], [6] has beenwidely adopted as one of the main fiber-to-the-home (FTTH)solutions capable of meeting the low-cost demands of the ac-cess networks. PON technologies are expected to deliver an

Manuscript received June 23, 2013; revised August 4, 2013 and August 31,2013; accepted September 3, 2013. Date of publication September 16, 2013;date of current version January 10, 2014.

The author is with the School of Electrical Engineering, Bangor University,Bangor LL57 1UT, U.K. (e-mail: [email protected]).

Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/JLT.2013.2281628

aggregate capacity of 40 Gb/s in the near future and the NG-PON2 standardization work has addressed this by the decisionto adopt a time-division multiplexing/wavelength-division mul-tiplexing (TDM/WDM) approach [7], this maintains the use ofconventional on-off keying (OOK) modulation with transmis-sion speeds preserved at 10 Gb/s per wavelength. It is widelyaccepted that in the long term the future generation PON tech-nologies must exceed the 10 Gb/s per wavelength threshold tofurther increase network capacity throughput. It is technicallyhighly challenging to achieve this with the conventional binaryOOK modulation. DSP-enabled PON technologies on the otherhand offer far greater flexibility in signal generation and decod-ing allowing compensation of signal distortions and/or utiliza-tion of advanced modulation formats which are inherently moretolerant to the fiber distortion effects. DSP can also allow theuse of spectrally efficient modulation techniques, which meansincreased network capacity can be achieved through efficientexploitation of component bandwidths, thus great commercialbenefits may be attained if the established, mature optical com-ponent sources for NG-PON1 and NG-PON2 can be exploited.DSP can also enable adaptive modulation techniques which canadapt to the varying spectral characteristics of the network dueto the natural variations in optical fiber, optical component, andradio frequency (RF) component characteristics. DSP can alsoenable important features such as dynamic bandwidth allocation(DBA) to improve capacity utilization efficiency. DSP-enabledoptical access networks can thus potentially provide network ad-ministrators with on-demand adaptability down to the physicallayer making the networks highly adaptable to the fluctuatingend-user service demands.

This paper presents a tutorial on real-time DSP for opticalaccess networks, first exploring general implementation as-pects of any DSP-based optical transceiver and then provid-ing an in-depth examination of real-time DSP for OOFDM-based transceivers based on the world first real-time end-to-endOOFDM transmission systems developed at Bangor Univer-sity. The rest of this paper is organized as follows; Section IIdiscusses the basic DSP-based transceiver structure and key ele-ments with a detailed discussion of digital-to-analog converters(DAC) and analog-to-digital converters (ADC). In Section III,the implementation of real-time DSP with field programmablegate arrays (FPGAs) is presented covering issues such asDAC/ADC interfacing and parallel processing and pipeliningtechniques. Example state of the art FPGAs are also described.In Section IV, the use of DSP in OOFDM-based access net-works is discussed. The basic principles of the OFDM modu-lation format are also covered, with the fundamental concepts

0733-8724 © 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See http://www.ieee.org/publications standards/publications/rights/index.html for more information.

554 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 32, NO. 4, FEBRUARY 15, 2014

Fig. 1. System elements of DSP-based optical transceivers.

of adaptively modulated OOFDM explained. The section alsodescribes in detail the architecture and fundamental functionsof a DSP-based OOFDM transceiver. In Section V, the real-time OOFDM transceiver implemented at Bangor Universityis described with sections describing in detail the IFFT/FFTfunction, channel estimation and equalization functions, andsymbol synchronization. Section VI presents the advantages ofthe multiband OOFDM technique over the single-band OOFDMtechnique when considering PON applications. Section VII con-cludes the paper.

II. REAL-TIME DSP-BASED OPTICAL TRANSCEIVERS

A. Transceiver Structure and Key Elements

The basic structure of a DSP-based optical transceiver isshown in Fig. 1. The key elements in the transmitter are: high-speed digital logic for hardware-based DSP, high speed DAC forconversion of the digital signal samples to an analog electricalsignal, a wideband RF section to amplify, filter and possibly up-convert the signal onto an RF carrier, and finally an electrical-to-optical (EO) converter that converts the analog electrical signalinto an optical signal for fiber launching. The key elements in thereceiver are: an optical-to-electrical (OE) converter to detect theoptical signal and convert to an electrical signal, a wideband RFsection to filter, amplify and possibly down-convert the signal,and a high speed ADC to convert the analog electrical signal todigital samples for processing by the high-speed digital logic.

The DSP functions must be implemented in digital hardwaredue to the ultrahigh processing speeds necessary to supportthe multi-Gb/s optical signals. It may be feasible to implementsome transceiver DSP functions in software which can oper-ate by subsampling the received signal, (e.g., synchronizationfunctions). However, for the majority of DSP functions, it isessential to employ digital hardware operating at clock speedsof several 100 MHz to achieve sufficient processing through-put. For prototyping real-time DSP hardware, FPGAs offer theideal solution due to their reprogrammability. This enables rapid

evaluation, exploration, and optimization of the hardware-basedalgorithms. The high cost and power consumption of FPGAs,however, makes them inappropriate for the cost and power sensi-tive PON applications. It is therefore necessary to employ cus-tom designed application specific integrated circuits (ASICs)for real-time DSP in commercial products. ASICs obviouslyrequire significant capital investment for development but reapthe benefits of low costs associated with high volume mass pro-duction of integrated circuits. ASICs also offer the advantage ofsignificant power reduction compared to FPGAs. The DAC andADC are highly critical components in the transceiver and arediscussed in more detail in Section II-B.

The transceiver structure in Fig. 1 is for operation withbaseband electrical signals. It is also possible to constructa transceiver which employs modulation of a single RF up-converted signal or multiple RF signals in a single or multi-band architecture. Moreover, transceiver architectures can com-bine baseband signals with RF up-converted signals. Generallyspeaking, in a multiband transceiver, multiple DACs and ADCsare required according to the number of subbands and alsodepending on whether in-phase/quadrature (IQ) modulation isused. As DACs/ADCs are critical components, it is greatly ben-eficial if advanced DSP algorithms can be used to relax boththe requirements and the number of DACs/ADCs required bya specific transmission system. A transceiver employing RFsignals obviously requires more complex RF sections, and is-sues such as RF carrier phase and frequency offsets must beaddressed. Unless specifically stated, the baseband transceiverwill be considered throughout this paper. However, the multi-band transceiver architecture is discussed in more detail in Sec-tion VI where its advantages over the single-band basebandtransceiver are analyzed. It should be noted, however, that dataconveyed by all signals are generated and recovered at basebandregardless of the transmission frequency band, such that the im-plemented DSP functionality is similar for all sub-bands. It isof course possible to use ultrawideband DACs and ADCs fordirect digital-to-RF conversion [8] thus eliminating the analogRF front-ends, but this approach is, at least for the present time,most likely too costly for application in cost-sensitive PONs.

Due to the cost sensitivity of the optical access network it isnecessary to employ low-cost optical front ends. For low costoptics, the intensity-modulation direct-detection (IMDD) tech-nique [9] is unrivalled. IMDD operates by either direct modula-tion or external modulation of a laser source. Directly modulatedlasers (DML) offer the lowest cost solution. However, DMLssuffer from the phenomenon of frequency chirp [10] whichcan degrade transceiver performance compared with the almostchirp-free external modulation scheme. For direct detection, aphotodiode or avalanche photodiode is employed which is a so-called square-law detector as the electrical current generated isproportional to the square of the optical field and therefore theoptical signal intensity. The photodiode is followed by a tran-simpedance amplifier to convert the detected current to a voltagefor the following RF section. For ultralow cost IMDD optics,a highly promising laser source is the vertical cavity surfaceemitting laser (VCSEL) [11] as these lasers can be producedat extremely low cost mainly due to the reduced manufacturingprocesses involved.

GIDDINGS: REAL-TIME DIGITAL SIGNAL PROCESSING FOR OPTICAL OFDM-BASED FUTURE OPTICAL ACCESS NETWORKS 555

A key point to emphasize is that the utilization of DSP enablesthe use of low-cost optical components in high performanceoptical transceivers, as their associated characteristics of limitedbandwidth, higher signal distortions, and wider tolerances canbe compensated for by the DSP algorithms either directly orindirectly through advanced modulation formats that are moretolerant to the component deficiencies. For example, the highspectral efficiency and variable signal spectrum of adaptivelymodulated OFDM, is able to fully utilize the lower bandwidth,and adapt to the varying frequency response, of low cost opticalcomponents.

As with conventional optical transceivers, the downstreamand upstream optical interfaces can either operate at the samewavelength for a dual feeder fiber-based PON, or as is moretypical, at different wavelengths for operation with a singlefeeder fiber-based PON. Other advanced optical transmissionschemes, such as wavelength remodulation [12] and a lightwave-centralized architecture [13] can also be employed with DSP-based optical transceivers.

B. DACs and ADCs

The DAC and ADC are highly critical components in DSP-based optical transceivers. The required DAC/ADC basic char-acteristics are: high sample rates of the order of several GS/s,bit resolutions in the region of 8 bits (modulation format depen-dent), high linearity and low noise. DAC/ADC aspects that canhave impact on transceiver performance include: quantizationnoise due to the discrete signal levels, non-ideal linear behav-ior which causes the effective number of bits (ENOB) to belower than the physical resolution, and the ENOB decreasingwith signal frequency. The full-scale of the DAC/ADC shouldbe utilized to minimize the effect of quantization noise, whichcan necessitate automatic gain control (AGC) before the ADC.DACs also typically have a characteristic roll-off in frequencyresponse due to the inherent sin(x)/x shaping due to the zero-order-hold output format, as well as low pass filtering effects ofthe on-chip analog front end. The sampling clock quality canmoreover affect performance due to clock jitter and frequencyoffset. It should be emphasized here that DSP algorithms can beexploited to mitigate some of the nonideal DAC/ADC propertiesand/or relax the required DAC/ADC performance requirements.

The required DAC/ADC sampling rate for a given line rateof R (bits/s) is dependent on the electrical spectral efficiencyE (b/s/Hz) of the adopted modulation format. The required sig-nal bandwidth is B = R/E (Hz). Therefore assuming operationover the entire Nyquist band and single-band transmission, therequired sampling rate is S = 2·B = 2·(R/E) (samples/s). Fig. 2shows graphically the variation of sample rate against line ratefor different spectral efficiencies. It can be seen, for example,that if the sampling rate is limited to 20 GS/s, a 40 Gb/s line ratewould require a modulation format with at least 4 b/s/Hz spec-tral efficiency. Modulation formats with high spectral efficiencyare thus important to minimize DAC/ADC sample rates.

Fig. 3 shows the bit resolution and sample rates of some com-mercial high speed DACs and ADCs currently available. Thetrend in DAC/ADC sampling rates has shown a steady growthover the last 5 years [14] and developments are generally led

Fig. 2. DAC/ADC sample rate versus line rate for different spectralefficiencies.

Fig. 3. Bit resolutions and sample rates of commercially availableDAC/ADCs.

by the progress in high-end test equipment such as digital sam-pling oscilloscopes (ADCs) and arbitrary waveform generators(DACs). The DAC/ADC can contribute a significant portion ofthe total power consumption in an optical transceiver [15], sothis is obviously a key area to be addressed in the developmentof future DAC/ADC targeted at access network applications.

III. REAL-TIME DSP IMPLEMENTATION WITH FPGAS

A. FPGA Technology

State-of-the-art FPGAs are unrivalled as a development plat-form for high-speed real-time signal processing. Modern FPGAssupport features such as:

� Vast array of logic elements;� Ultra-high speed transceivers (10 s Gb/s);� Huge resource of high speed IO (Gb/s);� Embedded memory;� Dedicated multiplier units;� High-performance embedded DSP blocks;� Embedded hard functions such as phase-locked loops

(PLLs);� Soft microprocessor support.To illustrate the performance available from high-end FPGAs,

Table 1 summarizes the features of Altera’s Stratix V familyof FPGAs [16] implemented in 28 nm complementary metal–oxide–semiconductor (CMOS) technology. Not only are somedevices offering almost 1 million logic elements and hundreds


TABLE IFEATURES OF ALTERA’S STRATIX V FPGA FAMILY [16]

Fig. 4. Parallel and pipelined processing.

or thousands of dedicated DSP blocks they also support animmense digital interface bandwidth which is essential for in-terfacing to the multi-GS/s DACs and ADCs. The huge digitalinterface bandwidth is provided by the multi-Gb/s embeddedtransceivers, which can offer bidirectional peak bandwidths ofover 1 Tb/s.

B. Parallelism and Pipelining for High Speed DSP

For the DSP-based optical transceiver, analog signal samplerates are of the order of several GS/s, whereas the digital logiccan be clocked at speeds on the order of several 100 MHz.To overcome this speed disparity parallelism and pipeliningtechniques must be fully exploited to achieve the required pro-cessing throughput. Fig. 4 shows the principle of the technique.Incoming digital samples at multi-GS/s from the ADC are firstpassed through a serial-to-parallel (S/P) converter which gener-ates parallel samples at a reduced sample rate compatible withthe FPGA logic speed. In order to maintain the necessary samplethroughput, the parallel samples are processed simultaneously.Furthermore, to maximize the clock speed of a digital logic func-tion, it is partitioned into a series of sequential functions withthe intermediate samples stored by registers, this is known as apipelining. Each sequential function is then clocked simultane-ously with its input samples taken from the registered outputs ofthe previous function. The maximum achievable clock speed is

Fig. 5. FPGA SERDES.

thus determined by the function with the longest propagation de-lay, which is significantly shorter than that of the correspondingnonpipelined function. Skillful partitioning of the higher levelfunction can thus enable maximization of a DSP function’s clockfrequency.

The sample throughput of a function (samples/s) is the prod-uct of the number of parallel samples and the clock frequency.Thus, if more (less) parallel samples are employed for a giventhroughput, the necessary clock speed is reduced (increased).The required logic resources are proportional to the number ofparallel samples. There can thus be a tradeoff between logicresources and clock speed. As power consumption is a functionof clock speed, it is possible to tradeoff die area and powerconsumption [17].

The pipelining approach may lead to an increase in total func-tion propagation delay. However, the high clock speeds involvedmean that this is unlikely to cause unacceptable processing la-tency. Pipelining a function increases clocking frequency, andas new output samples are generated on every clock cycle, theoverall function throughput is significantly increased.

C. FPGA Interfacing to DACs and ADCs

The multi-GS/s DACs and ADCs employed mean that ultra-high bandwidth digital interfaces between the FPGA and theDAC/ADC are necessary. A 56 GS/s, 8-bit converter requires adigital bandwidth of 448 Gb/s for example. As previously indi-cated, FPGAs offer large resources of high speed input/output(I/O) supporting speeds in the order of 1 Gb/s, as well as highspeed digital transceivers supporting speeds of several 10 s Gb/s,as illustrated in Table I. As the FPGA logic cannot operate atthese speeds, high speed serializer and deserializer (SERDES)circuits implemented in dedicated circuitry are employed in theFPGA, as illustrated in Fig. 5. The SERDES circuits typicallyhave a programmable range of parallelization ratios thus per-mitting the logic array to operate with parallel data at a suitableclock frequency, which is a sub-multiple of the interface clockfrequency. Due to the high frequencies involved, the logic sig-nals at the digital interface are typically differential logic carriedby controlled impedance signal pairs, thus requiring impedancematched interconnections. An example common interface logicstandard is low voltage differential signaling (LVDS), which hasan impedance of 100 Ω and logic levels of ±350 mV. As the


Fig. 6. FPGA to ADC digital interface example.

digital transceivers operate at higher frequencies than the I/O,they also incorporate embedded circuits to ensure high signalintegrity. These include clock data recovery (CDR) and pro-grammable equalization in the individual deserializer inputs,and programmable preemphasis at the individual serializer out-puts. To programme the equalization and preemphasis features,characterization of the interconnection is of course necessary.The SERDES can also have arbitrary phase offsets at power up,so it can be necessary to synchronize all SERDES when initial-izing the system. Test pattern generation by the ADC may thusbe necessary in order to correctly synchronize the deserializers.

Fig. 6 shows an example interface between an FPGA and a10 GS/s, 8-bit, 4 port ADC. The interface consists of 32 signalsoperating at 2.5 GHz. 32 × 10:1deserializers are used in theFPGA to give 320 parallel signals at 250 MHz.

IV. DSP IN OOFDM-BASED OPTICAL ACCESS NETWORKS

A. OFDM Modulation

OFDM is a multicarrier modulation (MCM) technique firstproposed in the 1960s [18]–[20] but at that time its implementa-tion was impractical. Salz and Weinstein [21] first proposed theuse of the discreet Fourier transform (DFT) [22] for the gener-ation of OFDM signals in 1969. It was not until semiconductorelectronics achieved sufficient processing power, however, thatimplementation of OFDM with the DFT was feasible. Today theOFDM modulation technique is widely adopted in numerouscommunication standards such as digital subscriber line (DSL)and its many variants, wireless local area networks (WLAN)and digital audio and video broadcast (DAB, DVB). OFDM isnow widely recognized as a potential modulation technique forapplication in future optical access networks [14], [23]–[28].In [14], [23] authors provide an extensive coverage of OFDM inoptical communications. The following sections, of this paper,provide an overview of the key principles of OFDM and adap-tive OFDM, pertinent to the real-time OOFDM implementationpresented in this paper.

To illustrate the principle of the MCM method, Fig. 7 shows ablock diagram of a generic multicarrier transmission system. Inthe transmitter, the incoming binary data stream is converted toa number of lower bit rate binary data streams, with the separate

Fig. 7. Block diagram of a generic multicarrier transmission system.

data streams encoded onto multiple carrier frequencies, alsoknown as subcarriers, and subsequently summed together fortransmission over the channel.

At the receiver, the same carrier frequencies are used to ex-tract the separate signals from the received signal containing allsubcarriers. The low pass filters remove the unwanted productsgenerated from other subcarriers before the signal is decoded torecover the binary data stream. The multiple binary data streamsare then combined to form a single high speed bit stream. It isextremely difficult to implement the subcarrier generation anddetection functions with discrete RF components even with alimited number of subcarriers. However, DSP is highly suitedto performing these functions, even when the number of sub-carriers is of the order of several hundreds or even approaches1000s [29].

OFDM is a special case of multicarrier modulation wherethere is a harmonic relationship between subcarrier frequencies,the frequency spacing Δf is defined as

Δf =1Tb

(1)

where Tb is the symbol period. Symbols are defined as the fixedlength time intervals into which the OFDM signal is divided fortransmission. Data are encoded onto the subcarriers in each sym-bol using both amplitude and phase modulation such as M-aryQAM modulation [30], where M is the number of distinct com-binations of subcarrier phase and amplitude. An M -ary QAMsignal thus encodes b bits were b = log2(M). For example, ifthere are 64 subcarriers each encoded with 16-QAM the numberof bits carried by each symbol is 64 × (log216) = 256. As anOFDM symbol encodes multiple bits its period is significantlylonger than a single bit period in a conventional OOK modulatedsignal for the same bit rate. This makes OFDM more tolerant todispersive channels such as wireless channels and optical fibers,as the extended symbol period greatly reduces the impact of thedispersive channel’s pulse spreading effect.

The frequency, fk , of the kth subcarrier is thus

fk = f0 +k

Tbk = 1, 2, 3, . . . , Ns (2)


Fig. 8. Principle of OFDM signal generation in the time domain.

where f0 is a fixed frequency offset and NS is the total numberof subcarriers. f0 is 0 for a baseband system and represents afrequency offset due to the frequency up-shifting in a carrier-based system. Fig. 8 illustrates the OFDM signal generationprinciple in the time domain for the case of 4 subcarriers overone symbol period.

The spreading effect associated with a dispersive channelcauses adjacent received symbols to overlap, a phenomenonknown as intersymbol-interference (ISI). To further improve thedispersion tolerance of OFDM, an intersymbol gap can be in-serted between two adjacent symbols to avoid the ISI occurringin the wanted signal region. To ensure the temporal spreadingof the signal in the intersymbol gap does not distort the wantedsignal region, each subcarrier is simply extended into the inter-symbol gap [23]. As the subcarriers are all cyclic the simplestway to achieve this is to take an appropriate portion from theend of the symbol and prefix it to the front of the symbol, thisis thus known as a cyclic prefix (CP). The CP causes a trans-mission overhead and so reduces the net bit rate, the length ofthe CP should thus be only as long as necessary to eliminate ISIfrom the wanted signal region of the symbol and to also providesufficient system operation robustness.

In addition to dispersion tolerance, OFDM has the extremelybeneficial characteristic of high spectral efficiency due to thesubcarrier orthogonality property allowing the subcarriers tooverlap in the frequency domain without interference. To un-derstand the spectrum of an OFDM signal the general form ofthe signal s(t) should be considered

s (t) =+∞∑

m=−∞

NS C∑

k=1

[Rk,m ejθk , m

]Υ (t) ej2πfk (t−mTb ) (3)

where k and m denote the mth symbol and the kth subcarrierand Rk,m and θk,m are the amplitude and phase respectively ofthe encoded data. Υ (t)is a rectangular pulse shaping waveformwhich ensures the mth encoded data is zero outside the mthsymbol, Υ (t)is given by

Υ (t) ={

1, (0 < t ≤ Tb)

0, (t ≤ 0, t > Tb) .(4)

A subcarrier within each symbol can thus be considered as theproduct of a modulated yet continuous sine wave and the rectan-gular waveform Υ (t) . The spectrum of Υ (t) is the sinc function

Fig. 9. OFDM signal spectrum showing four orthogonal subcarriers.

which has the form of sin(x)/x, multiplication by the continu-ous sine wave causes an up-shifting of the sinc function to besymmetrical about the sine wave frequency. This is illustratedby the four subcarrier spectra shown in Fig. 9, each is a sincfunction centered at the corresponding subcarrier frequency.An important property of the sinc function is the zero-crossingpoints which are spaced at the subcarrier spacing frequency Δf ,this means that although the subcarrier spectra overlap at anysubcarrier frequency all other subcarriers are zero valued. Asthe OFDM signal is processed digitally using the DFT only thediscrete subcarrier frequencies are considered, at these discretefrequencies only one subcarrier is nonzero thus the individualsubcarriers can spectrally overlap without intersubcarrier inter-ference. This spectral overlapping thus makes the OFDM signalvery spectrally efficient.

B. Adaptively Modulated OFDM

An important characteristic of OFDM is the ability to modu-late each subcarrier independently [31], [32] which allows thesignal to adapt to the spectral characteristics of the completetransmission channel which includes the fiber and transceivercomponents. As for any modulated signal its bit error rate (BER)performance is dependent on the received signal-to-noise ratio(SNR) thus for a desired BER there is a corresponding mini-mum SNR requirement. The minimum SNR will be modulationformat dependent, as the number of encoded bits increases thesignal’s tolerance to noise and distortion decreases thus the min-imum required SNR will increase. Each OFDM subcarrier canexperience different noise and distortion due to their frequencydependent nature thus SNR is subcarrier frequency dependent.There are two basic methods to ensure the minimum SNR isachieved for a specific subcarrier. First, for a fixed modulationformat an individual subcarrier’s transmitted power level can beadjusted to achieve the minimum SNR at the receiver. Second, ifthe transmitted subcarrier power is fixed the SNR at the receivercannot be adjusted, however the modulation format adopted ona particular subcarrier can be varied to change the minimumrequired SNR to be below but as near as possible to the actualSNR.

To illustrate the adaptive loading principle we consider achannel where noise and distortion are similar at all subcar-rier frequencies and the received subcarrier SNRs are only


Fig. 10. (a) Power loading and (b) bit loading of OFDM subcarriers.

dependent on the channel loss which is frequency dependent.Fig. 10(a) shows how subcarrier power loading is employedat the transmitter to compensate for the channel frequency re-sponse and ensure a constant SNR across all subcarriers at thereceiver. Fig. 10(b) illustrates how subcarrier bit loading can beemployed for the same case. The transmitted modulation formatis selected to ensure the minimum required SNR is aligned to theactual subcarrier SNR resulting from the frequency dependentchannel loss. It is important to note that power loading aloneallows the optimization of BER across subcarriers at a fixed bitrate whereas bit loading provides the added flexibility of con-trolling aggregated bit rate and can thus maximize the bit ratefor a given channel while meeting the required BER target. Bitloading also allows dynamic bandwidth redistribution betweensubcarriers enabling increased control of bandwidth allocationto end users. It is of course possible to employ a combination ofboth power and bit loading [33] for optimum flexibility.

It is important to consider the impact on DSP complexity ofthe three possible adaptive loading techniques of power loading(PL), bit loading (BL), and combined bit and power loading(BPL). Implementing BL or BPL requires significantly morecomplex DSP compared to PL due to the additional logic re-source required to implement the encoders and decoders sup-porting reconfigurable modulation formats [34].

Fig. 11. Functional DSP architecture of an OOFDM transceiver employingbaseband transmission of real-valued signals.

C. DSP-Based Optical OFDM

Fig. 11 illustrates the basic functional architecture of the DSPemployed within an optical OFDM transceiver which operateswith real-valued baseband signals. The DSP architecture con-sists of various functional blocks. However, the core functionsat the heart of the transceiver are the inverse DFT (IDFT) andDFT. The IDFT in the transmitter converts a series of N complexcoefficients, Xk , representing the amplitude and phase of thediscrete subcarrier frequencies, fk , (both positive and negative)to a series of N coefficients xn representing the correspondingdiscrete complex time domain samples. The IDFT is defined asfollows:

xn =1N

N2 −1∑

k= −N2

Xkej 2 πN kn n = −N

2, . . . , 0, . . . ,

N

2− 1.

(5)

The IDFT thus represents a bank of IQ modulators each mod-ulating one of the Δf spaced subcarriers. Where Δf is as de-fined in (1).

The DFT in the receiver converts a series of complex time do-main coefficients, xn , representing the OFDM symbol samplesinto the corresponding complex frequency domain coefficients,Xk , and is defined as follows

Xk =

N2 −1∑

n= −N2

xnej −2 πN kn k = −N

2, . . . , 0, . . . ,

N

2− 1.

(6)The DFT thus represents a bank of IQ demodulators each

demodulating one of the Δf spaced subcarriers.The signal flow through the transmitter DSP blocks is as fol-

lows: Incoming serial binary data are serial-to-parallel convertedto a sequence of parallel data words, each word is combinedwith pilot data for use in channel estimation and subdivided


into (N /2)-1 data words representing the data to be encodedon each of the (N/2)-1 subcarriers. A bank of encoders thenconvert each data word to a complex number according to themodulation format used by the encoder, where the complexnumber represents the subcarrier’s frequency domain amplitudeand phase. It should be highlighted that the encoders can beof a fixed type or adaptive encoders can be employed to al-low the selection of different modulation formats (bit loading)and/or power levels (power loading). To ensure the generatedtime domain signal is real valued only for use with the IMDD-based optics, the frequency coefficients at the IDFT inputs arearranged with Hermitian symmetry [22] such that the encodedcomplex data are applied to the positive (N /2)-1 frequency binsand the complex-conjugates are applied to the corresponding(N /2)-1 negative frequency bins, such that

X−k = X∗k k = 1, 2, ...,

(N

2

)− 1 (7)

and X0 = X−N/2 = 0 as the subcarrier at zero frequency can-not carry data. Also, due to the need for symmetry, there is nocorresponding positive frequency for X−N/2 . The IDFT thusconverts the N frequency-domain coefficients to N time do-main samples which represent one OFDM symbol. Due to thelarge peak-to-average-power-ratio (PAPR) of an OFDM signal,the time domain signal is clipped to reduce the PAPR and is alsoquantized to the number of bits of resolution required by theDAC. Although clipping distorts the signal, it increases the av-erage signal power, this reduces the effect of both quantizationnoise and independent analog noise sources, such as thermalnoise, due to the resultant increase in SNR at the receiver. Thereis however an optimum clipping level [35] which maximizessystem performance. A cyclic prefix is then added by prefixingη · N samples at the start of the symbol, where η is the cyclicprefix parameter to create a symbol of (η+1)·N samples, whereη is typically <1. The parallel samples of the whole symbolare then passed to the DAC typically via a number of serializercircuits as discussed in Section III-C.

The signal flow through the receiver DSP blocks is essen-tially the reverse of that in the transmitter, with the additionof symbol alignment and channel estimation and equalizationfunctions. The data from the ADC passes through a number ofdeserializer circuits to provide parallel signal samples typicallycorresponding in length to one symbol period, but arbitrarilylocated and so not aligned to the symbol boundaries. A symbolalignment block is employed to detect the symbol offset so thatit can be compensated for in the symbol offset adjustment blockand achieve symbol realignment. The cyclic prefix can then beremoved as it has no useful data. The remaining N time domainsamples are then processed by the DFT block to recover the Nfrequency domain coefficients from which (N /2)-1coefficientsare selected from the positive frequency bins. At the output of theDFT, a channel estimation function must detect the location ofthe subcarriers carrying pilot data. From these pilot-subcarriers,the channel can be estimated based on the known transmittedpilot-subcarriers. Once the channel transfer function (CTF) isknown for each subcarrier frequency, it can be used to equalizethe corresponding received subcarriers by a simple multiplica-

Fig. 12. Basic DSP architecture for an OOFDM transceiver employing an IQmodulated RF carrier.

tion by the inverse CTF at the given frequency. This is commonlyreferred to as single-tap equalization. The equalized subcarrierfrequency domain coefficients are then fed to a bank of decoderscorresponding to the encoders used in the transmitter to recoverthe transmitted data words. The data words are recombined andthe pilot data removed to form a single parallel data word con-taining all data bits transmitted in one symbol. The parallel wordis then passed to a parallel-to-serial converter for output at thetransceiver’s digital data output. It should be noted that the dig-ital data interfaces of the transceiver would in reality typicallyconsist of multiple serial interfaces, as the transceiver wouldeffectively multiplex several lower bit rate signals into one highspeed optical signal. For example, the transceiver may have 4× 10 Gb/s Ethernet ports which are multiplexed into a 40 Gb/soptical signal.

As a comparison, the DSP functional architecture for anOOFDM transceiver employing an IQ modulated signal isshown in Fig. 12. It can be seen that the DSP is essentiallythe same as that employed in the transceiver employing base-band transmission. The main difference is that N -1 subcarrierscan be used to carry data (zero frequency cannot be used) andthe time domain signal is now a complex value. The phase offsetbetween the received carrier and the receiver’s local oscillator(LO) results in a phase rotation of the I and Q componentswhich must be detected and corrected. Any frequency offsetbetween the transmitter’s carrier frequency and the receiver’sLO must also be accommodated, although this is not explicitlyshown in Fig. 12.

V. REAL-TIME DSP FOR OPTICAL OFDM

Since 2009, there have been a number of real-time OOFDMdemonstrations world-wide. Nearly all demonstrations are tar-geted at coherent long-haul systems [29], [36]–[44]. A real-timeOOFDM receiver targeted at PON applications was also demon-strated by NEC [45].


Fig. 13. DSP architecture of the adaptively modulated OOFDM transceiver.

TABLE IIKEY SYSTEM REAL-TIME OOFDM TRANSCEIVER PARAMETERS

Bangor University’s Optical Communications ResearchGroup were the first to demonstrate real-time end-to-end IMDD-based OOFDM transmission over SMF and MMF fibers in a se-ries of world first real-time OOFDM transmission experiments[35], [46]–[53]. More recently a group at ETRI also demon-strated a real-time end-to-end OOFDM system [54] targeted atPON applications. This section gives a detailed overview of thereal-time DSP employed in Bangor’s OOFDM transceivers.

A. Implementation

The DSP architecture of Bangor’s most recent real-timeOOFDM transceiver design, based on Altera’s Stratix II GXFPGAs, is shown in Fig. 13 and key transceiver parameters arepresented in Table II.

The architecture is similar to the functional DSP architectureshown in Fig. 11. The key differences are discussed here. In thetransmitter, a data generator block is implemented to provideparallel pseudorandom binary data for transmission, 15 adap-tive modulators are employed each of which can perform either16, 32, 64, or 128-QAM encoding by selecting one of fourdistinct encoders [34], and a power loading block allows liveadjustment of individual subcarrier power. The 32 point IDFTis implemented with an inverse fast Fourier transform (IFFT),

as this is a resource efficient form of the IDFT as discussedin Section V-B. The clipping and quantization block have anonline adjustable clipping level for live optimization. After thecyclic prefix is added the signed digital samples are converted tounsigned samples required by the DAC. This block also insertsa low power synchronization signal for symbol alignment as de-scribed in Section V-D. The samples and bits are then correctlyorganized for interfacing to 32×10:1 serializers feeding 32 I/Ooperating at 1 GHz. The interface to the DAC thus consists of4 × 8 bit ports such that 4 digital samples are transferred inparallel to the 4 GS/s DAC. All online controlled parametersare controlled via embedded memory accessed via the FPGA’sJoint Test Action Group (JTAG) [55] interface.

In the receiver, the ADC interface is the reverse of the DACinterface such that 32 1 GHz I/O feed 32×10:1 deserializers.A reorganization block restructures the parallel samples cor-rectly. These samples are used by the symbol offset detectionblock to determine the arbitrary sample offset from the symbolboundaries (described in Section V-D). The incoming samplesare thus realigned to the symbol boundary and the cyclic prefixremoved to provide 32 real-valued time domain samples cor-responding to one OFDM symbol. The 32 real-valued samplesare converted to 32 signed complex samples by removing theDC offset added by the ADC and setting all imaginary compo-nents to zero. The 32 complex time domain samples are thenfed to a 32 point FFT. The FFT is used as it provides a highlyefficient implementation of the DFT. From the 32 complex fre-quency domain coefficients output from the FFT, only the 15positive frequencies are selected. These correspond to the 15data-encoded frequency domain subcarriers. The pilot detec-tion block operates on the 15 frequency domain subcarriers todetect the pilot subcarriers which are used by the channel es-timation block to determine the CTF as described in SectionV-C. The 15 subcarriers will subsequently be equalized usingthe estimated CTF. The encoded binary data are then decodedusing adaptive decoders with modulation formats selected tocorrespond to those used in the transmitter. Finally, the totalchannel BER and individual subcarrier BERs are calculatedin a BER analyzer block in combination with a parallel pseu-dorandom binary data generator identical to that used in thetransmitter.

The real-time adaptively modulated OOFDM transceiver hasbeen used to achieve the following world first demonstrationsinvolving real-time end-to-end OOFDM transmission in bothSMF and MMF links by fully exploiting its ability to dynami-cally adapt to the system spectral characteristics.

� Aggregated upstream baseband OOFDM transmission at11.25 Gb/s in an OFDMA PON [56].

� Colourless OLT-seeded REAM-based baseband OOFDMupstream transmission at 10 Gb/s over bidirectional 25 kmSSMF [57].

� 19.125 Gb/s dual-band OOFDM transmission over 25 kmSSMF employing an EML as intensity modulator [51].

� 19.25 Gb/s dual-band OOFDM colourless transmissionover 25 km SSMF employing a REAM as intensity modu-lator [52].

� 20 Gb/s dual-band OOFDM transmission over 500 m OM2MMF [58].


� 25.25 Gb/s triple-sub-band OOFDM transmission over300 m OM2 MMF.

� Triple subband OOFDM transceivers operation at 30 Gb/s[53].

B. IFFT and FFT

To explicitly compute xn and Xk from the definitions of an Npoint IDFT and DFT, as given in (5) and (6) respectively, wouldrequire N 2 complex multiplications and N 2–N complex addi-tions. For a hardware-based implementation of the transforms,it is highly advantageous to minimize computational complex-ity in order to minimize design complexity. Furthermore, theextremely high IDFT/DFT real-time computational throughputinherent to OOFDM implies that a highly parallel and pipelinedarchitecture is necessary. This makes it difficult to reuse complexfunctions for more than one calculation during one transformcycle. Therefore, minimizing the number of discrete instancesof complex functions in the algorithm is vitally important ifchip cost and power consumption targets are to be met. Thefast Fourier transform (FFT) and inverse FFT (IFFT) are highlycomputationally efficient algorithms for computing the DFTand IDFT, respectively. The FFT was first introduced by Tukeyand Cooley in 1965 in their seminal paper “An algorithm forthe Machine Calculation of Complex Fourier Series” [59]. Thedrastic reduction in computational complexity offered by theFFT and IFFT makes them highly appropriate for implemen-tation in physical hardware and thus ideal for use in real-timeOOFDM transceivers. The IFFT can be created from an FFT bysimple modification. Therefore, the following discussions willconcentrate on the FFT although they are equally applicable tothe IFFT.

The fundamental principle of the FFT is to take the inputsequence xn of N coefficients and split it into shorter subse-quences. For this example, we split the original N point se-quence into two sequences of length N /2. Two DFTs of lengthN/2 can now be performed and the resulting two N /2 outputsequences can be recombined to form the N point output se-quence of the original DFT. The recombination process takescare of the associated time shift between the sequences. Thissplitting approach can be continued on the subsequences by fur-ther division of the subsequences for M = log2N total stepsuntil the N -point DFT is replaced by N DFTs each of length1. The DFT operation is now trivial as the DFT of a 1- pointsequence is itself, thus requiring no computational stage. Thetask of the FFT is then to correctly recombine the 1-point se-quences according to the splitting method, thus recombining theN 1-point sequences into N /2 2-point sequences, and then theseinto N /4 4-point sequences, and so on, until the final single N -point transform sequence Xk is formed. The N -point FFT thusconsists of a total of log2N recombination stages.

When the original sequence and all subsequences are equallydivided at each step, this is a radix-2 FFT with N = 2M , whereM is the number of recombination steps. Other radices are cre-ated when the sequence is split into more than two subsequencesat each stage. For example, if the original sequence is first splitinto four subsequences, this is repeated for M = log4N steps.In this case N = 4M and so it is a radix-4 FFT. The possible

Fig. 14. Radix-2 decimation in time butterfly element.

radix values that can be employed are therefore dependent ofthe required value of N . To allow more flexibility in the valueof N , it is also possible to create mixed radix FFTs [60].

A detailed examination of the conversion of one N -point DFTinto two N /2-point DFTs will be presented as this also explainsthe origin of the fundamental building block of the FFT: thebutterfly operator. If the original time domain sequence xn issplit into its even and odd sequences of yn = x2n and zn =x2n+1 , respectively, for n = 0,1,..,(N /2)–1 and substituted intothe DFT as defined in (6), this becomes

Xk =

N2 −1∑

n=0

{ynω2nk

N + znω(2n+1)kN

}(8)

for k = 0,1,..(N–1) and ωN = e−j2π/N . (8) can be rewrittenusing the relation ωN

2nk = ωnkN/2 as

Xk =

N2 −1∑

n=0

ynωnkN/2 + ωk

N

N2 −1∑

n=0

znωnkN/2 . (9)

The original DFT has now been expressed as a simple com-bination of two DFTs each of length N /2. The DFT on the rightof (9) is multiplied by the factor ωk

N which accounts for therelative time shift between the sub-sequences and is known asthe twiddle factor. If these N /2-point DFTs are denoted as Yk

and Zk respectively we can write

Xk = Yk + ωkN Zk (10)

and

Xk+ N2

= Yk+ N2

+ ω(k+ N

2 )N Zk+ N

2(11)

for k = 0,1,..,(N /2)–1. (11) can be further simplified as ωNN/2

= −1 and Yk and Zk have a period of N /2, which gives thefollowing pair of equations known as butterfly operators:

Xk = Yk + ωkN Zk (12)

Xk+ N2

= Yk − ωkN Zk . (13)

These butterfly operators are the fundamental FFT buildingblocks used at the recombination stages of the radix-2 FFT, andare depicted by the example symbol shown in Fig. 14. To converttwo subsequences Yk and Zk to the single N -point sequence Xk

will thus require N /2 discreet butterfly operators.Different radix values and sequence splitting methods will

have their own corresponding butterfly elements. The radix-4butterfly, for example, has 4 coefficient inputs, three twiddlefactor inputs, and 4 coefficient outputs.

The selected splitting method can affect the order of the se-quence reordering and the sequence recombination. The afore-mentioned example splits the original sequence and the subse-quences into even and odd sequences. This requires the inputcoefficients xn to first be reordered and then the sequences to be


recombined. As xn is reordered, this type of FFT architecture isknown as decimation-in-time. If, however, the splitting methoddivides according to the first-half and last-half subsequences, xn

and xn+N/2 , where n = 0,1,..,(N /2)–1, the FFT first performsthe recombination of the naturally ordered input sequence andthen reorders the Xk coefficients. This type of FFT architectureis therefore known as decimation-in-frequency.

The selected architecture for the FFT and IFFT implementedin the real-time OOFDM transceiver is based on the fact that a32-point DFT is required. A radix-2 FFT can be used as N =32 = 25 . There is no difference in the complexity of decimation-in-time and decimation-in-frequency, therefore decimation-in-time is selected. The implemented FFT architecture is thereforethe Cooley–Tukey radix-2 decimation-in-time.

To implement the IFFT it is only necessary to modify thetwiddle factors, which is apparent from the opposite sign of theexponential power in the IDFT in (5) compared to the DFT in(6). Thus, for the IFFT, the twiddle factors are now ωk

N . Toconvert the FFT function to an IFFT function, the twiddle factorvalues are thus simply replaced with their complex conjugates.

It is important to consider the savings in computational com-plexity achieved by the implemented FFT architecture comparedto the explicit computation according to the DFT definition in(6). Explicit computation requires N 2 complex multiplicationsand N 2 − N complex additions, whereas a radix-2 N -point FFTwill have (N /2)log2N butterfly operators, each consisting of onecomplex multiplier and two complex additions. Thus, in totalthere are (N /2)log2N complex multipliers and N log2N com-plex additions. For the case of the implemented 32-point FFT,the computational saving is ∼92% for the complex multiplica-tions and ∼84% for the complex additions. The actual saving ishigher when taking into account the instances where the twid-dle factors are unity. The immense computational efficiency ofthe IFFT is clear from the savings achieved. Furthermore, thecomputational saving increases further with higher values of N .For a hardware-based implementation of the DFT, the FFT istherefore indispensable due to the advantages associated withthe vast reduction in the required logic resources. The FFT andIFFT will still constitute one of the largest logic functions, ifnot the largest logic function in the OOFDM transceiver, and sothe optimization of the FFT logic is an important issue.

As an example, Fig. 15 shows the structure of a 16 point radix-2 decimation-in-time FFT. The repeated divide by 2 structureis apparent in the right-to-left direction and the decimation-in-time architecture results in the reordering of the xn values. Animportant issue with hardware implementation is bit resolutioncontrol of the intermediate stage sample values. As the butterflyelements within each stage contain multiplication and additionfunctions, the bit resolutions of the output samples will increaseat each stage. If this is not restricted, the final stage will have anexcessive bit resolution and so large logic resources will be con-sumed. It is thus critical that the intermediate sample resolutionsare truncated to limit the excessive escalation of sample bit res-olutions while maintaining sufficient calculation precision. Asshown in Fig. 15, sample resolution reduction must be builtinto the FFT structure between stages. Another important factorwhich can affect logic resource usage and calculation precisionis the bit resolution of the twiddle factor values. This must be

Fig. 15. Structure of a 16-point radix-2 decimation-in-time FFT.

carefully selected as overly high resolution can cause excessiveuse of logic, whereas overly low resolution can cause insuffi-cient calculation precision. The implemented twiddle factor is a6 bit signed complex value.

For the implemented real-time transceiver, it is worth notingthe processing power of the FFT and IFFT. There are log2(32) =5 recombination stages each consisting of 32/2 = 16 radix-2 but-terfly elements, giving a total of 80 butterflies. There are thus80 complex multipliers and 160 complex adders giving a to-tal of 240 complex operations. As the FFT/IFFT is clocked at100 MHz, this results in a processing throughput of 24×109

complex operations per second. In comparison, direct computa-tion with the DFT would require 1024 complex multipliers and961 complex adders, resulting in close to 200×109 complexoperations per second.

C. Pilot Detection, Channel Estimation, and Equalization

The frequency response of the transmission channel intro-duces subcarrier amplitude and phase changes during transmis-sion. The received signal is therefore no longer a direct repre-sentation of the transmitted signal. To compensate the effect ofthe channel response, the inverse channel response is applied inthe receiver, which is termed channel equalization. In order toperform equalization, the CTF must be estimated. An advantageof OFDM is that channel equalization can be extremely simple.As the amplitude and phase of each subcarrier are determinedat a discrete frequency, the CTF only needs to be known at thecorresponding frequency to allow the subcarrier to be equal-ized. Equalization can then be achieved by a single complexmultiplication in the frequency domain.

To determine the CTF at the subcarrier frequencies, pilotsymbols Pk are periodically transmitted on each subcarrier withknown amplitude, Ak and phase θk , defined as

Pk = Akej (2π kN n+θk ) . (14)


Fig. 16. Subcarrier equalization to compensate for channel induced amplitudeand phase changes.

Fig. 17. Pilot and data bearing subcarrier mapping in the OOFDM time-frequency symbol space.

The corresponding received pilot symbol Rk is

Rk = Bkej (2π kN n+φk ) + Wk (15)

where the received subcarrier amplitude and phase are Bk , andφk respectively and Wk is the noise component of the kth sub-carrier after the receiver FFT. The CTF in the frequency domain,Hk , is then determined as

Hk =Rk − Wk

Pk=

Bk

Akej (φk −θk ) (16)

the CTF is thus estimated as

Hk =Rk

Pk= Hk +

Wk

Pk. (17)

Large pilot symbol amplitude therefore reduces error dueto noise. To further reduce the effects of channel noise, theestimated CTF can be averaged over many pilot symbols aslong as the channel can be considered to be static over theaveraging period. To equalize the received frequency domaincomplex data, d′k,m , encoded onto the kth subcarrier, a single

multiplication by the inverse CTF estimate, H−1k , is applied.

The equalized encoded complex data value d′′k,m is thereforedefined as

d′′k,m = H−1k d′k,m . (18)

The subcarrier equalization principle is illustrated in Fig. 16.The real-time OOFDM transceiver implements pilot

subcarrier-based channel estimation in the following way. Inthe transmitter, the pilot insertion function follows the paralleldata generator, such that one extra parallel bit sequence of afixed pattern, representing known pilot subcarrier data, is diag-onally mapped into the OOFDM time-frequency symbol spaceas shown in Fig. 17. Mathematically, the pilot and data-bearingsubcarrier mapping onto the frequency domain subcarriers Xk,m

can be expressed as

Xk,m ={

pk,m (m − k) = qNs

dk,m (m − k) �= qNsq = 0, 1, 2, 3, . . .

(19)

where pk,m and dk,m are the encoded complex pilot and datavalues, respectively, and Ns is the total number of data bearingsubcarriers (Ns is 15 in this case). The diagonal pilot mappingapproach was adopted as it has the advantage that no bufferingof the incoming data is required when all subcarriers carry thesame number of bits. However, it is still necessary to direct theNs–1 incoming data streams and the single pilot data value tothe appropriate subcarriers on a per OFDM symbol basis.

In the receiver, the 15 data-bearing OFDM subcarriers in thepositive frequency bins are selected for channel estimation andsubsequent data recovery at the FFT output. When transmissionis first established, the pilot detection block must locate thesymbols such that the first subcarrier is a pilot subcarrier. Thesesymbols are regarded as pilot subcarrier reference points relativeto which all other pilot subcarriers can be located. At the outputof the FFT, the identification of the received pilot subcarriers isfirst made by performing operations (20) and (21) to subcarrier1 of consecutive symbols, where

Dm,1 = Xm,1 · X∗(m+Ns ),1 (20)

Qm,1 =1C

∣∣∣∣∣

C−1∑

i=0

D(m+iNs ),1

∣∣∣∣∣

2

(21)

such that Xm,1 (X∗(m+N s),1) is the received complex (complex

conjugate) value of subcarrier 1 of the mth [(m+NS ) th] symbol.C is a preset integer number determining the total number ofNs symbol-spaced D values used for averaging. The magnitudesquared function is used in (21) as this gives a real-valued Qvalue to simplify peak detection, and is also easier to computethan the absolute magnitude which would require a square-rootoperation. As the pilot mapping sequence repeats every NS

symbols, Qm,1 must be determined for NS adjacent symbol lo-cations. For the implemented design with NS = 15, 15 valuesof Qm,1 for consecutive symbols positions must be determined.The data-bearing subcarriers are modulated with complex val-ues encoded using a random data sequence. This results in mini-mized Q values due to the averaging process. On the other hand,each of the pilot subcarriers is modulated with a fixed complexnumber of maximal amplitude, causing the occurrence of a Qpeak corresponding to the symbol locations where subcarrier1 is the pilot subcarrier, as illustrated in Fig. 18. A large Cwill make the Q peak more distinguishable, but this requires alonger time and more logic resources to conduct the averagingoperation, such that C should be optimized. Experimental mea-surements show that C = 16 is adequate for reliable detectionof pilot subcarriers. By locating the peak in the 15 detected andstored Q values the symbols are identified where subcarrier 1 isthe pilot. Based on this reference pilot, all other pilot subcarriersin subsequent symbols can be easily identified due to their fixedrelative positions. In the implemented design, the pilot detectionfunction operates continuously. However, after identifying the


Fig. 18. Pilot subcarrier identification using Q peaks after FFT in the receiver.

reference pilot subcarriers, the pilot detection process could beterminated and only needs to be reactivated following a breakin the transmission.

Making use of the known transmitted pilot subcarriers and thereceived pilot subcarriers, the channel estimation block deter-mines the complex CTF, Hk (k = 1, 2, . . .,Ns), by performingthe operation defined below in (22) as:

Hk =1M

M −1∑

i=0

R(k+iNs ),k

P(k+iNs ),k=

1M · PC

M −1∑

i=0

R(k+iNs ),k (22)

where R(k+iN s),k (P(k+iN s),k ) is the received (assigned) com-plex value of the kth pilot subcarrier in the (k + iNs ) th symbol.As a constant power, PC , is assigned to the pilot subcarriers thesimplified expression on the right of (22) is used. To reduce thenoise effect associated with the transmission system, frequencyresponse averaging is performed over M pilot subcarriers ateach frequency. Here, M is taken to be 32, which is an op-timum value identified experimentally [61]. Thus, to computeHk , parallel summation functions with suitable scaling are im-plemented over 32 pilots for each subcarrier. The 15 computedcomplex values forming the CTF are stored and fed to the chan-nel equalization block with new values continuously computedevery 32 symbols.

The CTF obtained in the channel estimation function is thenused by the channel equalization block to equalize each individ-ual subcarrier using the following operation:

Xm,k =X ′

m,k

Hk(23)

where X ′m,k is the received complex value of the kth unequal-

ized subcarrier in the mth symbol. The channel equalizationfunction thus consists of 15 parallel complex dividers. Theequalized subcarriers, Xm,k , provide the inputs to the 15 paralleladaptive demodulators.

The real and imaginary parts of the 15 complex CFT parame-ters Hk determined by the channel estimation block are probedby the Signal Tap II embedded logic analyzer. H1 to H15 areextracted by the Signal Tap IIapplication in order to view thelive system frequency response from IFFT input to FFT output.This feature is utilized in combination with individual subcarrierBER measurements to manually determine suitable levels for thevariable power loading profile employed in the transmitter.

It is important to note that, due to the quasi-static nature of theoptical channel, a low CTF estimate update rate can be employedwithout degrading system performance. The channel estimationtechnique can therefore only insert pilot data in periodic burstsof pilot subcarriers. This allows all 15 subcarriers to be used fordata transmission between pilot bursts. The insertion rate of thepilot bursts can be as low as 10 Hz [61], corresponding to anextremely low overhead of 0.001% for the channel estimationfunction.

D. Symbol Synchronization

Symbol timing offset (STO) is the difference between thecorrect symbol start position and the estimated symbol start po-sition. Symbol synchronization is necessary to minimize STO,which is ideally zero, as nonzero STO leads to degraded BERperformance if the processed samples do not all originate fromthe same symbol. It should be noted, however, that if the CPlength exceeds the ISI length by y samples, an STO of up toy samples can be tolerated without performance degradation.STO tolerance can thus be improved by increasing CP length.

A DSP-based symbol synchronization method has been ex-perimentally demonstrated that is highly suitable for applica-tion in OOFDM multiple access-based passive optical networks(OOFDMA-PONs). This is because the technique can achievesymbol, timeslot, and frame alignment of an optical networkunit’s (ONU’s) upstream and downstream signals without theneed to interrupt existing ONU traffic.

The symbol synchronization technique [62], [63] is based onthe principle that the dc level of each OOFDM symbol has noinfluence on the nonzero frequency subcarriers at the output ofthe FFT. A different dc offset can therefore be applied to eachindividual OOFDM symbol. This operation thus produces anencoded synchronization signal, SALIGN , which is added to theOOFDM signal, SOOFDM , at the transmitter. For simplicity butwithout losing generality, two different dc offsets with identicalamplitude but opposite polarities, +P and –P, are alternatelyadded to successive symbols. SALIGN is thus a square wavewith a period of 2 symbol periods (2·TS ), and a peak-to-peakamplitude of 2P . The transmitted signal ST X as illustrated inFig. 19 is as follows:

ST X = SOOFDM + SALIGN . (24)

The corresponding received signal SRX can be written as

SRX = S ′OOFDM + S ′

ALIGN + SN (25)

where SN represents system noise.In the receiver, a cross-correlation method is used to detect the

position of S ′ALIGN . A signal SCORR is generated which has an

identically shaped waveform to SALIGN and amplitude of ±1to simplify computation. By computing the cross-correlationbetween SRX and SCORR , symbol alignment offset can be de-termined based on the location of the correlation peaks. Thisis because there is no correlation between SCORR and eitherS ′

OOFDM or SN due to their Gaussian random characteris-tics. The cross-correlation is therefore entirely dependent on


Fig. 19. OOFDM signal combination with dc offset symbol alignment signal(alignment signal dc levels are exaggerated for clarity).

S ′ALIGN . An arbitrarily positioned sequence of 2·M · Z sam-

ples is processed, where Z is the total number of samples in anOOFDM symbol, and M is a sufficiently large integer selectedto give clear correlation peaks. As SALIGN is cyclic, a sym-bol summation, or accumulation, can be performed before thecross-correlation. A signal SSUM is calculated using (26) whereSSUM (n) is the nth sample within SSUM and n = 1 to Z.

SSUM(n) =M −1∑

k=0

SRX (n + 2kZ) =M −1∑

k=0

S ′OOFDM(n + 2kZ)

+ S ′ALIGN(n + 2kZ) + S ′

N (n + 2kZ) (26)

where SSUM is thus the sum of M sequences of Z consecu-tive samples spaced at intervals of 2·Z samples. If M is largeenough, the waveform of SSUM will take on the shape of SALIGNas the Gaussian random characteristics of S′OOFDM and SN re-sult in their summations both tending to zero. The exact shapeof SSUM will depend on the symbol alignment offset relative tothe arbitrarily selected samples. Signal transitions from positiveto negative, and vice-versa, will thus coincide with the OOFDMsymbol boundaries. The cross-correlation is then performed be-tween SSUM and SCORR with the relative offset, v, of SCORRvaried from 0 to (2·Z)–1 and the correlation value COR(ν)for each offset calculated using (27). The sequence of valuesCOR(0) to COR(2Z–1) provides a correlation profile, CPROF ,where the position of the peaks indicates the offset where thehighest correlation between S′ALIGN and SCORR occurs, thusidentifying the position of the OOFDM symbol.

COR (v) =(2Z )−1∑

k=0

SSUM (k) · SCORR (k + v) . (27)

A positive (negative) peak will occur in CPROF when SCORRand S′ALIGN are in phase (in opposite phase) both of whichindicate symbol alignment as SCORR and SALIGN have a periodof 2·TS . By taking |COR(ν)|, only positive peaks then occur inCPROF and it is only necessary to select Z samples in every 2·Zsamples to ensure a peak is detected. Fig. 20 shows the ideal

Fig. 20. Ideal correlation profile for symbol offset w0 .

Fig. 21. Symbol offset detection block diagram.

variation of |COR(ν)| against offset ν for an arbitrary symbolalignment offset of w0 .

The addition of the dc offset level is performed in the signedto unsigned block in Fig. 13. The added dc offset is onlineadjustable to allow optimization. In the experimental demon-stration [62], it was shown that a dc offset as small as ±1quantization level was sufficient and resulted in no reduction insystem BER performance.

A block diagram of the implemented symbol offset detec-tion function is shown in Fig. 21. The sum and accumulateblock consists of 40 parallel accumulators, corresponding to the40 samples per symbol period, to generate a new SSUM every10 000 symbols as M = 5000. As each accumulator sums a totalof 5000 8 bit samples, between resets, scaling is used to limit theaccumulator outputs to 12 bits. 40 parallel cross-correlators areemployed to generate the correlation profile. A peak detectordetects the position of the correlation profile peak to determinethe symbol offset value.

It should be noted that as the correlation signal SCORRhas values of ±1, the cross-correlation function consists of 40add/subtract operators each with 40 inputs. The offset of the cor-responding SCORR value determining if a sample is added orsubtracted. The use of multipliers is thus avoided to reduce de-sign complexity. Also, it should be noted that although multipleparallel cross-correlators were employed, it would be possible tosignificantly reduce logic resources by implementing the func-tion with a single cross-correlator, and sequentially incrementthe offset of the correlation signal to build up the correlationprofile one value at a time.

As previously discussed, the symbol synchronization tech-nique is designed for application in OOFDMA-PONs to achieve


downstream and upstream alignment of symbols, timeslots andframes. Here, the mechanism for achieving synchronization ofan ONU in a live PON is described. For downstream symbolalignment, the optical line terminal (OLT) continuously trans-mits a synchronization signal which all ONUs use for symbolalignment. In the upstream direction, the OLT controls the syn-chronization process, allowing only one ONU to transmit asynchronization signal at any one time. For each ONU, the OLTdetects its upstream symbol offset and then notifies the ONUvia a control channel, so that it can correctly realign its sym-bol positions. The ONU thus contains a symbol offset detectionfunction in its receiver and a symbol offset adjustment functionin both its transmitter and receiver. The OLT only requires thesymbol offset detection function in its receiver.

By constructing the synchronization signal from a codedsequence of dc offsets, enhancements can be made that of-fer a number of key features. For upstream and downstreamframe/timeslot alignment, a suitably coded synchronization sig-nal with the same length as one or more OOFDMA framesallows the OLT to detect an ONU’s frame alignment offset byperforming a cross-correlation over a period equivalent to oneor more coded sequence lengths. ONU frame alignment offset isdetected and then corrected in the ONU, such that each ONU istimeslot aligned to the network before initiating OOFDMA sig-nal transmission and hence avoiding any upstream ONU signalcollisions in the operational network. As symbol offset can driftslowly over time, the OLT must periodically track and correctany symbol offset drift for each ONU in turn.

Furthermore, coding the downstream synchronization signalwith a sufficiently long encrypted key code will make it virtuallyimpossible for an unauthorized user to achieve synchronization,thus achieving network security at the physical layer.

An alternative symbol synchronization technique employinga subtraction-based correlation method for cyclic prefix locationhas also been implemented and fully verified in the real-timetransceiver [64]–[66].

E. Clock Synchronization

Accurate synchronization of the OOFDM receiver and trans-mitter sampling clocks is essential to minimize their sam-pling frequency offset (SFO) [67]. SFO induces interchannel-interference (ICI) which produces increasing received signaldistortion with increasing SFO. ICI results from the loss ofsubcarrier orthogonality due to the mismatch between the dis-crete subcarrier frequencies in the receiver compared to thosein the transmitter. SFO also induces a drift in symbol alignmentnecessitating periodic symbol realignment.

Due to the noise-like nature of the OOFDM signal clock re-covery is not straightforward. However, asynchronous (nonzeroSFO) and synchronous (zero SFO) clocking techniques in real-time OOFDM transmissions have been demonstrated. The CPdetection-based symbol alignment method [66] supports asyn-chronous clocking. The technique is able to compensate for SFOas it continuously readjusts the symbol alignment and so pre-vents the accumulation of excessive STO. Suitably fast symboloffset tracking and symbol realignment was demonstrated tocompensate for a SFO as high as 4000 ppm. This method [66]

Fig. 22. Multiband OFDM signal generation principle.

can also estimate SFO to an accuracy of <1 ppm which couldallow highly accurate synchronization of the receiver clock viaa feedback loop from a voltage controlled oscillator (VCO).

For synchronous clocking, a DSP-free technique employingan out-of-band auxiliary clock signal was employed [68]–[70].At the transmitter, a 4 GHz sine wave is combined with the 0–2 GHz baseband OFDM electrical signal, before driving theoptical modulator. At the receiver the clock signal and theOFDM signal are separated using electrical filters. The receivedclock signal is prescaled and feeds a PLL generating a low jit-tered clock at the required frequency for clocking the receiver.The synchronous clocking technique has particular significancefor implementing accurate synchronization of OOFDMA-basedPONs.

VI. MULTIBAND OOFDM-BASED PONS FOR IMPROVED

COST EFFECTIVENESS

As the cost of the ONU is critical in an OOFDMA PON, itis important to avoid unnecessary overengineering of the ONU.Employing single-band OOFDM transceivers in a PON leads tothe undesirable scenario, where all ONUs support the full peakPON capacity. In practice, however, an ONU will only ever needto operate at a reduced peak capacity. If a multiband OOFDMapproach [51], [52] is adopted this can overcome the drawbacksof the single-band approach. Fig. 22 illustrates the multibandOFDM signal generation principle. Each OFDM transceivergenerates a baseband signal, which is then up-converted usinga unique RF carrier frequency such that the generated subbandsdo not overlap in the frequency domain. A frequency divisionmultiplexing (FDM) method is thus adopted to combine theOOFDM subbands together. In the downstream direction, theOOFDM subbands are electrically summed in the OLT beforeEO conversion, whereas for the upstream direction, the summa-tion occurs in the optical domain in the optical coupler in thePON’s remote node.

There are many advantages associated with the multibandOOFDM technique particularly when considering the ONU im-plementation. It offers the key advantage of flexibility in adoptedDAC/ADC bandwidth as this is no longer dictated by the totalPON capacity. Moreover, ONU signal processing complexity is


reduced, as only one PON subband is processed. If the subbandtransceiver is designed to support subband tunability, it pro-vides increased network operation efficiency in terms of bothdynamic bandwidth provisioning and equipment logistics. Fur-thermore, the reduced complexity leads to a reduction in ONUpower consumption. Although the OLT must support all sub-bands, the same tunable subband transceiver electronics, as usedin the ONUs, can be employed. This provides the benefits ofeconomies of scale in transceiver manufacturing, and also al-lows a scalable OLT architecture, where capacity can expandin sub-band capacity increments in line with service take up.It would also be possible to implement dynamic traffic redis-tribution across subbands, so that when PON traffic levels per-mit, transceivers can be powered down to reduce OLT energyconsumption. As higher cost and complexity can be toleratedat the OLT side of a multiband OOFDMA-PON, an alterna-tive OLT architecture employing wideband DACs and ADCs,for direct digital-to-RF conversion of all subbands [71], is alsoconceivable.

An OOFDM transceiver, designed to support a single subbandin a multi band system, will additionally require, two RF mixers,a single tunable LO and RF filters to support the up-conversionand down-conversion of the OFDM signal. The IMDD opticalcomponents will be similar for the single-band and multibandtransceivers; although the multiband approach will require wideroptical component bandwidths. This is because the optical bandmust encompass all subbands to support dynamic subband tun-ability. Furthermore, double side-band subbands and the inter-subband spacing will increase the required optical bandwidth.

It is important to compare the difference between subbandgeneration using a single carrier, and using two orthogonalcarriers for IQ modulation. IQ modulated subbands have theadvantage of increasing spectral efficiency. This theoreticallyhalves the subband bandwidth for a given data capacity. As aconsequence, optical component bandwidth requirements aresignificantly reduced. For IQ modulation, DAC and ADC pairsare needed, as illustrated in Fig. 12, though the bandwidths arenow halved compared to a single carrier generated subband ofthe same data capacity.

The savings in IFFT/FFT processing complexity andDAC/ADC bandwidths will now be considered in detail forthe case of an ONU that supports one subband using a singleRF carrier. For an IMDD system employing real-valued timedomain signals, for each subband, the relationship between thenumber of data-carrying subcarriers NS and the IFFT/FFT sizeN , is NS = (N /2)–1. For an N point radix-2 decimation-in-time IFFT/FFT architecture, the number of complex operationsis N log2N complex additions and (N/2)log2N complex multi-plications, as described in Section V-B. If the total capacity andthus the number of subcarriers in a PON is fixed, then the effectof employing multiple bands is to reduce the number of sub-carriers per subband and also reduce the required bandwidth ofeach subband. If each ONU supports one subband, the requiredIFFT/FFT complexity and DAC/ADC bandwidth for each ONUwill reduce as the number of subbands increases. Table III andFig. 23 illustrate this relationship for a PON with a total of atleast 500 subcarriers.

TABLE IIIRELATIVE SUBBAND PARAMETERS FOR MULTIBAND OOFDM

Fig. 23. Relative IFFT/FFT complexity and DAC/ADC bandwidth per ONUfor varying number of side-bands in a multiband OOFDMA-PON.

If the relative IFFT/FFT processing complexity andDAC/ADC bandwidths, for a single subband ONU, are cal-culated for the case of IQ modulated subbands. The relativeparameter values in Table III and Fig. 23 show almost ex-actly the same trends when considering a PON supporting ∼500subcarriers.

Using the adaptively modulated 4 GS/s real-time OOFDMtransceivers, the feasibility of multiband OOFDM has beendemonstrated by real-time dual-band OOFDM experimentaldemonstrations employing baseband (0–2 GHz) and passband(∼4–8 GHz) OFDM signals. OOFDM signal transmission at>19 Gb/s over 25 km of SSMF has been successfully demon-strated [51], [52]. Furthermore, by using this approach, real-timeOFDM signal generation and detection at 30 Gb/s has beendemonstrated [53].

VII. CONCLUSION

This paper has provided an overview of the implementationaspects associated with DSP-based optical transceivers for fu-ture access networks by examining the optical transceiver struc-ture and the key transceiver constituent elements. This paper fo-cuses on DSP functionality and architecture of OOFDM-basedoptical transceivers. Real-time OOFDM transceivers experi-mentally demonstrated by Bangor University are also analyzedin depth.

In summary, today’s digital processing and DAC/ADC devicetechnologies are sufficiently mature to support the application ofDSP in high speed optical access networks. Exploiting DSP not


only allows future network performance requirements to be met,but can also enable cost-effective technology by allowing theutilization of low cost, simple IMDD optical components. Fur-thermore, the high equipment volumes associated with opticalaccess networks can inevitably lead to cost-effective electron-ics. OOFDM is one of the leading DSP-based optical accesstechnologies which is perceived by many as one of the maincontenders for future optical access networks due to its poten-tial for high cost-effectiveness, data capacity per wavelengthfar beyond 10 Gb/s, adaptiveness to varying network charac-teristics, and flexibility in terms of bandwidth allocation. It isalso envisioned that adopting a multiband OOFDM-based PONwill maximize total access network cost-effectiveness. Giventhe exponentially growing demand for data capacity and theoperators’ need for flexible cost-efficient access networks, it isbelieved that DSP-based optical access networks will emerge inthe future.

REFERENCES

[1] Z. Kostic and S. Seefharaman, “Digital signal processors in cellular radiocommunications,” IEEE Commun. Mag., vol. 35, no. 12, pp. 22–35, Dec.1997.

[2] S. Chaturvedi, “The role of digital signal processors (DSP) for 3 G mobilecommunication systems,” Int. J. Emerging Technol., vol. 1, no. 1, pp. 23–26, Feb. 2010.

[3] A. Gatherer, T. Stetzler, M. McMahan, and E. Auslander, “DSP-based ar-chitectures for mobile communications: Past, present, and future,”. (2000,Jan.). IEEE Commun. Mag., vol. 38, pp. 84–90, 2000.

[4] “Coherent DWDM technologies.” (2012). White Paper Infinera. [On-line] Available: www.infinera.com/pdfs/whitepapers/Infinera_Coherent_Tech.pdf

[5] T. Koonen, “Trends in optical access and in-building networks,” in Eur.Conf. Opt. Commun., Brussels, Belgium, 2008, pp. 1–31.

[6] P. Vetter, “Next generation optical access technologies,” presented at Euro-pean Conf. Optical Communication, Amsterdam, The Netherlands, 2012.

[7] Y. Luo, X. Zhou, F. Effenberger, X. Yan, G. Peng, Y. Qian, and Y. Ma,“Time- and wavelength-division multiplexed passive optical network(TWDM-PON) for next-generation PON stage 2 (NG-PON2),” J. Lightw.Technol., vol. 31, no. 4, pp. 587–593, Feb. 2013.

[8] G. Engel, D. E. Fague, and A. Toledano, “RF digital-to-analog convertersenable direct synthesis of communications signals,” IEEE Commun. Mag.,vol. 50, no. 10, pp. 108–116, Oct. 2012.

[9] C. Cox, E. Ackerman, R. Helkey, and G. E. Betts, “Techniques and per-formance of intensity-modulation direct-detection analog optical links,”IEEE Trans. Mircowave Theory Tech., vol. 45, no. 8, pp. 1375–1383, Aug1997.

[10] A. Villafranca, J. Lasobras, and I. Garces, “Precise characterization ofthe frequency chirp in directly modulated DFB lasers,” in Spanish Conf.Electron. Devices, 2007, pp. 173–176.

[11] M. Muller and M. C. Amann, “State-of-the-art and perspectives for long-wavelength high speed VCSELs,” in 13th Int. Conf. Transp. Opt. Netw.,Stockholm, Sweden, 2011, pp. 1–4, Paper Mo.C5.2.

[12] C. H. Yeh and C. W. Chow, “10-Gb/s upstream TDM-PON based on fourWDM signals with OFDM-QAM remodulation,” in 13th OptoElectron.Commun. Conf., Hong Kong, 2009, pp. 1–2, Paper ThLP46.

[13] M. F. Huang, D. Qian, and N. Cvijetic, “A novel symmetric lightwave cen-tralized WDM-OFDM-PON architecture with OFDM-remodulated ONUsand a coherent receiver OLT,” presented at Europeon Conf. Optical Com-munication, Geneva, Switzerland, 2011, Paper Tu.5.C.1.

[14] N. Cvijetic, “OFDM for next-generation optical access networks,” J.Lightw. Technol., vol. 30, no. 4, pp. 384–398, Feb. 2012.

[15] J. L. Wei, J. D. Ingham, D. G. Cunningham, R. V. Penty, and I. H. White,“Performance and power dissipation comparisons between 28 Gb/s NRZ,PAM, CAP and optical OFDM systems for data communication applica-tions,” J. Lightw. Technol., vol. 30, no. 20, pp. 3273–3280, Oct. 2012.

[16] “Stratix V device overview,” (2013). [Online]. Available: http://www.altera.co.uk/literature/hb/stratix-v/stx5_51001.pdf

[17] R. Bouziane, P. Milder, R. Koutsoyannis, Y. Benlachtar, J. C. Hoe,M. Puschel, M. Glick, and R. I. Killey, “Design studies for ASIC imple-mentations of 28 GS/s optical QPSK- and 16-QAM-OFDM transceivers,”Opt. Exp., vol. 19, no. 21, pp. 20857–20864, Oct. 2011.

[18] R. W. Chang, “Synthesis of band-limited orthogonal signals for multi-channel data transmission,” Bell Syst. Technol. J., vol. 45, pp. 1775–1796,1966.

[19] R.W Chang and R. A. Gibby, “A theoretical study of performance of an or-thogonal multiplexing data transmission scheme,” IEEE Trans. Commun.Technol., vol. 16, no. 4, pp. 529–540, Aug. 1968.

[20] M. Zimmerman and A. Kirsch, “The AN/GSC-10/KATHRYN vari-able rate data modem for HF radio,” IEEE Trans. Commun. Technol.,vol. COM-15, no. 2, pp. 197–205, Apr. 1967.

[21] J. Salz and S. B. Weinstein, “Fourier transform communication system,”in Proc. ACM Symp. Problems Optim. Data Commun. Syst., New York,NY, USA, 1969, pp. 99–128.

[22] W. L. Briggs and V. E. Henson, The DFT An Owner’s Manual for theDiscrete Fourier Transform. Philadelphia, PA: SIAM, 1995.

[23] J. Armstrong, “OFDM for optical communications,” J. Lightw. Technol.,vol. 27, no. 3, pp. 189–204, Feb. 2009.

[24] A. Ragheb and H. Fathallah, “Candidate modulation schemes for nextgeneration-passive optical networks (NG-PONs),” presented at the HighCapacity Optical Networks and Enabling Technol., Istanbul, Turkey, 2012.

[25] C. Milion, T. Duong, N. Genay, E. Grard, V. Rodrigues, B. Charbonnier,J. Le Masson, M. Ouzzif, P. Chanclou, and A. Gharba, “High bit rate trans-mission for NG-PON by direct modulation of DFB laser using discretemulti-tone,” presented at the European Conf. Optical Communication,Vienna, Austria, 2009, Paper 7.5.4.

[26] L. A. Neto, A. Gharba, P. Chanclou, N. Genay, B. Charbonnier, M. Ouzzif,C. Aupetit-Berthelemot, and J. Le Masson, “High bit rate burst modeoptical OFDM for next generation passive optical networks,” presentedat the European Conf. Optical Communication, Torini, Italy, 2010, PaperTu.3.B.5.

[27] E. Wong, “Next-generation broadband access networks and technologies,”J. Lightw. Technol, vol. 30, no. 4, pp. 597–608, Feb. 2012.

[28] N. Cvijetic, “Recent progress in DSP-based optical access and OFDMA-PON,” presented at the Opto-Electronic and Communication Conf., Busan,Korea, 2012, Paper 4A4-1.

[29] B. Inan, S. Adhikari, O. Karakaya, P. Kainzmaier, M. Mocker, H. vonKirchbauer, N. Hanik, and S. L. Jansen, “Realization of a real-time 93.8-Gb/s polarization-multiplexed OFDM transmitter with 1024-point IFFT,”in Proc. Eur. Conf. Opt. Commun., Geneva, Switzerland, 2011, PaperWe.9.A.7.

[30] S.G. Wilson, Digital Modulation and Coding. Englewood Cliffs, NJ:Prentice-Hall, 1995.

[31] J. M. Tang, P. M. Lane, and K. A. Shore, “Transmission performance ofadaptively modulated optical OFDM signals in multimode fiber links,”IEEE Photon. Technol. Lett., vol. 18, no. 1, pp. 205–207, Jan. 2006.

[32] J. M. Tang and K. Alan Shore, “30-Gb/s signal transmission over 40-kmdirectly modulated DFB-laser-based single-mode-fiber links without opti-cal amplification and dispersion compensation,” J. Lightw. Tech., vol. 24,no. 6, pp. 2318–2327, Jun. 2006.

[33] E. Giacoumidis, A. Kavatzikidis, A. Tsokanos, J. M. Tang, and I. Tomkos,“Adaptive loading algorithms for IMDD optical OFDM PON systemsusing directly modulated lasers,” J. Opt. Commun. Netw., vol. 4, no. 10,pp. 769–778, Oct. 2012.

[34] X. Q. Jin, J. L. Wei, R. P. Giddings, T. Quinlan, S. Walker, and J. M. Tang,“Experimental demonstrations and extensive comparisons of end-to-endreal-time optical OFDM transceivers with adaptive bit and/or power load-ing,” IEEE Photon. J., vol. 3, no. 3, pp. 500–511, Jun. 2011.

[35] R. P. Giddings, X. Q. Jin, E. Hugues-Salas, E. Giacoumidis, J. L. Wei,and J. M. Tang, “Experimental demonstration of a record high 11.25 Gb/sreal-time optical OFDM transceiver supporting 25 km SMF end-to-endtransmission in simple IMDD systems,” Opt. Exp., vol. 18, no. 6, pp. 5541–5555, Mar. 2010.

[36] Y. Benlachtar, P. M. Watts, R. Bouziane, P. Milder, D. Rangaraj,A. Cartolano, R. Koutsoyannis, J. C. Hoe, M. Puschel, M. Glick, andR. I. Killey, “Generation of optical OFDM signals using 21.4 GS/s realtime digital signal processing,” Opt. Exp., vol. 17, no. 20, pp. 17658–17668, Sep. 2009.

[37] Y. Benlachtar, P. M. Watts, R. Bouziane, P. Milder, R. Koutsoyannis,J. C. Hoe, M. Puschel, M. Glick, and R. I. Killey, “21.4 GS/s real-time DSP-based optical OFDM signal generation and transmission over1600 km of uncompensated fibre,” presented at the 35th European Conf.Optical. Communication, Vienna, Austria, 2009, Paper PD2.4.


[38] F. Buchali, R. Dischler, A. Klekamp, M. Bernhard, and D. Efinger, “Reali-sation of real-time 12.1 Gb/s optical OFDM transmitter and its applicationin a 109 Gb/s transmission system with coherent reception,” presented atthe 35th European Conf. Optical Communication, Vienna, Austria, 2009,Paper PD2.1.

[39] R. Schmogrow, M. Winter, B. Nebendahl, D. Hillerkuss, J. Meyer,M. Dreschmann, M. Huebner, J. Becker, C. Koos, W. Freude, andJ. Leuthold, “101.5 Gbit/s real-time OFDM transmitter with 16QAMmodulated subcarriers,” presented at the Optical Fiber CommunicationConf., Los Angeles, CA, USA, 2011, Paper OWE5.

[40] R. Schmogrow, R. Bouziane, D. Hillerkuss, P. A. Milder,R. J. Koutsoyannis, Y. Benlachtar, P. M. Watts, P. Bayvel, R. I. Killey,C. Koos, W. Freude, and J. Leuthold, “85.4 Gbit/s real-time OFDM signalgeneration with transmission over 400 km and preamble-less reception,”presented at the Conf. Lasers and Electro-Optics, San Jose, CA, USA,2012, Paper CTh1H.4.

[41] Q. Yang, S. Chen, Y. Ma, and W. Shieh, “Real time reception of multi-gigabit coherent optical OFDM signals,” Opt. Exp., vol. 17, no. 10,pp. 7985–7992, May 2009.

[42] S. Chen, Y. Ma, and W. Shieh, “Multiband real-time coherent opticalOFDM reception up to 110 Gb/s with 600-km transmission,” IEEE Pho-ton. J., vol. 2, no. 3, pp. 454–459, Jun. 2010.

[43] N. Kaneda, Q. Yang, X. Liu, S. Chandrasekhar, W. Shieh, and Y. Chen,“Real-time 2.5 GS/s coherent optical receiver for 53.3-Gb/s sub-bandedOFDM,” J. Lightw. Tech., vol. 28, no. 4, pp. 494–501, Feb. 2010.

[44] X. Xiao, F. Buchali, M. Berhard, S. Chen, and H. Bulow, “Implementationof a 10.5 Gb/s real-time CO-OFDM receiver,” presented at the 16th Op-toeElectronics and Communications Conf. (OECC), Kaohsiung, Tiawan,2011.

[45] D. Qian, T. T. Kwok, N. Cvijetic, J. Hu, and T. Wang, “41.25 Gb/s real-timeOFDM receiver for variable rate WDM-OFDMA-PON transmission,” pre-sented at the Optical Fiber Communication Conf., San Diego, CA, USA,2010, Paper PDPD9.

[46] R. P. Giddings, X. Q. Jin, H. H. Kee, X. L. Yang, and J. M. Tang, “Real-time implementation of optical OFDM transmitters and receivers for prac-tical end-to-end optical transmission systems,” Electron. Lett., vol. 45,no. 15, pp. 800–802, 2009.

[47] R. P. Giddings, X. Q. Jin, and J. M. Tang, “Experimental demonstration ofreal-time 3 Gb/s optical OFDM transceivers,” Opt. Exp., vol. 17, no. 19,pp. 16654–16665, Sep. 2009.

[48] R. P. Giddings, X.Q. Jin, and J.M. Tang, “First experimental demonstra-tion of 6 Gb/s real-time optical OFDM transceivers incorporating chan-nel estimation and variable power loading,” Opt. Exp., vol. 17, no. 22,pp. 19727–19738, Oct. 2009.

[49] R. P. Giddings, E. Hugues-Salas, X.Q. Jin, J. L. Wei, and J. M. Tang,“Experimental demonstration of real-time optical OFDM transmission at7.5 Gb/s over 25-km SSMF using a 1-GHz RSOA,” Photon. Technol. Lett.,vol. 22, no. 11, pp. 745–747, Jun. 2010.

[50] R.P. Giddings, E. Hugues-Salas, B. Charbonnier, and J. M. Tang, “Exper-imental demonstration of real-time optical OFDM transmission at 11.25Gb/s over 500 m MMFs employing directly modulated DFB lasers,” IEEEPhoton. Technol. Lett., vol. 23, no. 1, pp. 51–53, Jan. 2011.

[51] R.P. Giddings, E. Hugues-Salas, and J.M. Tang, “Experimental demon-stration of record high 19.125 Gb/s real-time end-to-end dual-band opticalOFDM transmission over 25 km SMF in a simple EML-based IMDD sys-tem,” Opt. Exp., vol. 20, no. 18, pp. 20666–20679, Aug. 2012.

[52] Q. W. Zhang, E. Hugues-Salas, R. P. Giddings, M. Wang, and J. M. Tang,“Experimental demonstrations of record high REAM intensity modulator-enabled 19.25 Gb/s real-time end-to-end dual-band optical OFDM color-less transmissions over 25 km SSMF IMDD systems,” Opt. Exp., vol. 21,no. 7, pp. 9167–9179, Apr. 2013.

[53] R.P. Giddings, E. Hugues-Salas, and J. Tang, “30 Gb/s real-time triple sub-band OFDM transceivers for future PONs beyond 10 Gb/s/λ,” presented atthe European Conf. Optical Communication, London, U.K., 2013, PaperP.6.7.

[54] S. Cho, K. W. Doo, J. H. Lee, J. Lee, S. I. Myong, and S. S. Lee, “Demon-stration of a real-time 16 QAM encoded 11.52 Gb/s OFDM transceiver forIM/DD OFDMA-PON systems,” in 18th OptoElect. Comm. Conf. (OECC)and Int. Conf. Photon. Switching (OECC/PS), Kyoto, Japan, 2013, PaperWP2-3.

[55] IEEE Standard for Test Access Port and Boundary-Scan Architecture,IEEE Standard 1149.1-2013, 2013.

[56] X.Q. Jin, E. Hugues-Salas, R.P. Giddings, J.L. Wei, J. Groenewald, andJ.M. Tang, “First real-time experimental demonstrations of 11.25 Gb/soptical OFDMA PONs with adaptive dynamic bandwidth allocation,”Opt. Exp., vol. 19, no. 21, pp. 20557–20570, Oct. 2011.

[57] E. Hugues-Salas, R. P. Giddings, X. Q. Jin, Y. Hong, T. Quinlan, S. Walker,and J.M. Tang, “REAM intensity modulator-enabled 10 Gb/s colorlessupstream transmission of real-time optical OFDM signals in a single-fiber-based bidirectional PON architecture,” Opt. Exp., vol. 20, no. 19,pp. 21089–21100, Sep. 2012.

[58] E. Hugues-Salas, Q. Zhang, R.P. Giddings, M. Wang, and J.M. Tang,“Adaptability-enabled record-high and robust capacity-versus-reach per-formance of real-time dual-band optical OFDM signals over variousOM1/OM2 MMF systems [Invited],” J. Opt. Comm. Netw., vol. 5, no. 10,pp. A1–AA11, Jun. 2013.

[59] J. W. Cooley and J. W. Tukey, “An algorithm for the machine calculationof complex fourier series,” Math. Comput., vol. 19, no. 90, pp. 297–301,Apr. 1965.

[60] E. J. Kim and M. H. Sunwoo, “High speed eight-parallel mixed-radix FFTprocessor for OFDM systems,” presented at the IEEE Int. Symposium onCircuits and Systems, Rio de Janeiro, Brazil, 2011.

[61] X.Q. Jin, R.P. Giddings, and J.M. Tang, “Real-time transmission of 3 Gb/s16-QAM encoded optical OFDM signals over 75 km SMFs with negativepower penalties,” Opt. Exp., vol. 17, no. 17, pp. 14574–14585, Aug. 2009.

[62] R.P. Giddings and J. M. Tang, “Real-time experimental demonstration ofa versatile optical OFDM symbol synchronisation technique using low-power DC offset signalling,” in Eur. Conf. Exhibit. Optical Communica-tion. (ECOC), Geneva, Switzerland, 2011, Paper We9A3.

[63] R. P. Giddings and J. M. Tang, “Symbol alignment in high speed opticalorthogonal frequency division multiplexing transmission systems,” U.K.Patent 1105808.8, 2011.

[64] X. Q. Jin, R. P. Giddings, E. Hugues-Salas, and J. M. Tang, “Real-timeexperimental demonstration of optical OFDM symbol synchronization indirectly modulated DFB laser-based 25 km SMF IMDD systems,” Opt.Exp., vol. 18, no. 20, pp. 21100–21110, Sep. 2010.

[65] J. M. Tang and X. Q. Jin, “Synchronization process in optical frequencydivision multiplexing transmission,” U.K. Patent 0919057.0, 2009.

[66] X. Q. Jin and J. M. Tang, “Optical OFDM synchronization with symboltiming offset and sampling clock offset compensation in real-time IMDDsystems,” IEEE Photon. J., vol. 3, no. 2, pp. 187–196, Apr. 2011.

[67] D. K. Kim, S. H. Do, H. B. Cho, H. J. Choi, and K. B. Kim, “A new jointalgorithm of symbol timing recovery and sampling clock adjustment forOFDM systems,” Trans. Consumer Electron., vol. 44, no. 3, pp. 1142–1149, 1998.

[68] R. P. Giddings and J. M. Tang, “World-first experimental demonstration ofsynchronous clock recovery in an 11.25 Gb/s real-time end-to-end opticalOFDM system using directly modulated DFBs,” presented at the OpticalFiber Communication Conf., Los Angeles, CA, USA, 2011, Paper OMS4.

[69] R. P. Giddings and J. M. Tang, “Experimental demonstration and optimi-sation of a synchronous clock recovery technique for real-time end-to-endoptical OFDM transmission at 11.25 Gb/s over 25 km SSMF,” Opt. Exp.,vol. 19, no. 3, pp. 2831–2845, Jan. 2011.

[70] R. P. Giddings and J. M. Tang, “Synchronous clocking for optical orthog-onal frequency division multiplexing transmission systems,” U.K. Patent1 008 018.2, 2010.

[71] N. Cvijetic, M. Cvijetic, M. F. Huang, E. Ip, Y. K. Huang, and T. Wang,“Terabit optical access networks based on WDM-OFDMA-PON,” J.Lightw. Technol., vol. 30, no. 4, Feb. 2012.

Roger Giddings received the B.Eng. (Hons.) degree in engineering science andtechnology from Loughborough University, Loughborough, U.K., in 1989, andthe Ph.D. degree in optical communications in 2011 from Bangor University,Bangor, U.K.

He joined Nokia Networks UK in 1990, where he worked as a HardwareDesign Engineer and Senior Hardware Design Engineer. In 1998, he joinedNokia Networks, Finland. In 2000, he joined Nokia’s Research Center, Finland,as a Senior Research Engineer. In 2004, he joined the Nokia Ventures Unit,Finland, as a Senior Hardware Specialist.

In 2007, he joined the Optical Communications Research Group, School ofElectronic Engineering, Bangor University, as an Experimental Researcher andstudying for the Ph.D. degree in the area of real-time DSP for OOFDM-basednext generation optical networks. In 2012, he was awarded a lectureship and iscontinuing to research optical communication systems based on optical OFDM.

During his work at Bangor he has (co)authored more than 60 papers and hasfiled 2 patents.

real-time digital signal processing for optical ofdm-based future optical access networks

Documents