Research Article Vol. 12, No.

9 / September 2020 / Journal of Optical Communications and Networking D1

DSP enabled next generation 50G TDM-PON

Borui Li,1, * Kuo Zhang,2 Dechao Zhang,3 Jiale He,1 Xiaolong Dong,2 Qian Liu,4 AND
Shengping Li1
1
Transmission & Access Research Department, Huawei Technologies Co., Ltd, Wuhan 430079, China
2
Transmission & Access Research Department, Huawei Technologies Co., Ltd, Dongguan 523808, China
3
Network and IT Technology Department, Research Institute of China Mobile, Beijing 100053, China
4
China Academy of Information and Communications Technology, Beijing 100191, China
*Corresponding author: [email protected]

Received 2 March 2020; revised 21 April 2020; accepted 24 April 2020; published 18 May 2020 (Doc. ID 391904)

With 10G time division multiplexed passive optical network (TDM-PON) systems ready for massive deploy-
ment worldwide, the next generation PON standards are being intensively discussed in both IEEE and ITU-T.
Compared with the IEEE’s wavelength stacked 2∗25G solution, the ITU-T’s single carrier non-return-to-zero-
based 50G TDM-PON has the advantages of low cost, easy operation, and convenient management. However,
the 50G TDM-PON will also have to face great challenges in real-world operating conditions. In this paper, we
will illustrate the unique role for digital signal processing (DSP) in solving the challenges for the next generation
50G TDM-PON systems, such as device bandwidth limitation, dispersion induced power penalty, and high link
budget. To analyze DSP’s feasibility in this application, potential issues such as cost and interoperability are
analyzed. Through offline experiments and prototype demonstrations, DSP’s role in enabling next generation
single carrier 50G TDM-PON is highlighted. © 2020 Optical Society of America
https://doi.org/10.1364/JOCN.391904

1. INTRODUCTION at least for the OLT parts. This trend is most obvious in China,
The time division multiplexed passive optical network (TDM- which has the largest wireline access market and the most
PON) has been the most popular and important infrastructure positive attitude toward 10G-PON deployment. The huge
to realize fiber to the home/building (FTTH/FTTB) since 10G-PON component shipment volumes will bring down the
it was invented in the late 1980s. Due to the point to mul- price greatly and further accelerate 10G-PON development.
tipoint architecture by passive power splitters at the remote With more and more bandwidth hungry Internet broad-
node, the TDM-PON system is cost-effective and conven- band services like VR/AR applications and ultra-high
ient to maintain. Starting with broadband PON (BPON, definition video, and more and more unpredictable emerg-
622 Mb/s downstream and 155 Mb/s upstream rate), TDM- ing technology development, it is natural for standards bodies
PON is evolving from Ethernet PON (EPON) and Gigabit to consider next generation PON. The IEEE P802.3ca Task
PON (GPON) to the following 10G EPON and 10 Gigabit Force has already been working on the standardization of
asymmetric/symmetric PON (XG-PON/XGS-PON) stand- 25G TDM-PON since 2016. The standard is expected to
ardized by IEEE 802.3 and ITU-T Q2/SG15, respectively. be finished in 2020. It was claimed that the large volume of
The progress in PON standards is shown in Fig. 1. Typically, it 25G optical components in the data center market can drive
takes 6–7 years to enable one generation of PON to be widely the quick deployment of 25G-PON with low cost [2,3].
deployed considering the declining cost of optoelectronic However, the performance requirement gap between the
components and growing demand for broadband Internet 25GE-LR/100BASE-ER4 scenario and 25G-PON is signifi-
services. Currently, Gigabit-PON is no longer satisfying the cant in aspects such as transmitter eye quality, launch power,
increasing residential broadband service consumption and and the receiver sensitivity. These factors will have great impact
business requirements, as well as the converged service’s needs. on the optical chips, component yield rate, and cost. The parts
According to Ovum’s forecast [1], the 10G-PON optical line that may possibly be leveraged are electrical components such
terminal (OLT) port volume will approach more than 90% as transimpedance amplifiers (TIAs) or SerDes. Consider 10G-
of all PON OLT ports in the next 5 years. Also, for optical PON for example: when we look back at its initial deployment
network terminal (ONT) ports (units), 10G-PONs will be in 2016, a high cost had to be borne until the PON ship-
about half of the total volume. That means most of the current ment volume became larger and larger, even though the 10G
Gigabit-PON system will be replaced by a 10G-PON system, Ethernet ecosystem was already mature. The maturity of the

1943-0620/20/0900D1-08 Journal © 2020 Optical Society of America

Authorized licensed use limited to: CMU Libraries - library.cmich.edu. Downloaded on August 17,2020 at 18:24:49 UTC from IEEE Xplore. Restrictions apply.
 D2 Vol. 12, No. 9 / September 2020 / Journal of Optical Communications and Networking Research Article

and the PON system’s challenges, it is very difficult to realize

a practical 50G-PON in the conventional way that uses opto-
electronic transceiver components with enough bandwidth to
transmit and receive the NRZ signal.
When we review developments in the optical fiber commu-
nication industry, it is notable that the digital signal processing
(DSP) technique has already been implemented from the
Fig. 1. Progress in TDM-PON standards in ITU-T and IEEE. long-haul transmission scenario to the short-haul data center
400GE/200GE system [11,12]. As mentioned above, the truly
leveraged part for PON is the electronic part. Therefore, on
latter contributed little to the former’s success. Moreover, in the expectation that 50G optoelectronic transceivers will find
view of some operators who have already been well prepared it difficult to meet the 50G-PON requirement, it is necessary
to widely implement 10G-PON systems, 25G is a small step to implement the 50G-PON system based on current 25G
from 10G, considering the 5–6 year continuous infrastructure commercial optical and electronics components. To combat
investment cycle [4]. these challenges, DSP will be a good catalyst to facilitate the
Although the IEEE task force’s objective also includes 50G 50G-PON.
[5], it is based on a solution that stacks two fixed wavelengths, In this paper, we will illustrate the DSP merits in over-
which will add more cost to the system and more maintenance coming the aforementioned challenges and in enabling 50G
difficulties. As a counterpart, ITU-T also began to standardize TDM-PON deployment. We will illustrate the key algorithms
50G TDM-PON, but with a single carrier in early 2018 [6,7]. in the DSP flow and demonstrate the 50G-PON based on DSP
While 50G-PON is an exciting target, there will be no experimentally in Section 2. In Section 3, we will address some
doubt great challenges to face due to the PON system’s unique concerns in DSP implementation in the PON system such as
requirements. Every time a new PON generation is planned, interoperability, the cost, and the feasibility. Finally, Section 4
the first thing to consider is how it can coexist with legacy concludes this paper.
PONs and reuse the deployed optical distribution networks
(ODNs). That means choosing unused wavelength bands,
which may bring new problems, and overcoming the current 2. DSP BASED 50G TDM-PON FEASIBILITY
ODN power budget, which is as large as 30–32 dB. When As mentioned above, based on the latest standard advancement
the bit rate increases, the large power budget is an obvious and device maturity status, the 50 GBaud NRZ-based TDM-
issue. That is why 50G non-return-to-zero (NRZ) wins against PON operated at 1342 nm with a 25G transceiver device will
25 GBaud four-level pulse amplitude modulation (PAM4) as face three issues: inter-symbol interference (ISI) brought on
the adopted modulation format. Furthermore, considering by a limited bandwidth device, dispersion caused by signal
that the lower O-band wavelengths are occupied by lower degradation, and the power budget requirement. To solve these
speed PON systems, only the O+ band remains for 50G- problems, we implement advanced DSP at the receiver side
PON. ITU-T has agreed on the 1342 nm ±2 nm range as correspondingly as a linear equalizer, a nonlinear equalizer,
the downstream wavelength for 50G-PON, which means and advanced forward error correction (FEC). Furthermore,
up to 80 ps/nm chromatic dispersion (CD) exists in a 20 km apart from the common requirements like performance and
PON link [8]. For a 50 GBaud NRZ signal, the dispersion power consumption for DSP in the modules, there is a unique
induced penalty is also another challenge. In addition, for a feature in the PON system from burst mode (BM) operation
50 GBaud NRZ solution, the maturity of optical and electrical for upstream (US) transmission. The BM receiver in the OLT
components is a third issue. Although 50 GBaud lasers will be receives optical burst signals from different optical network
employed in 400GE, either led by the CDWDM8 multisource units (ONUs) that have different optical power levels and
agreement or 100 Gb/s/wavelength with PAM4 format, they initial phases. As a result, not only is a BM-TIA with high
will still need years to achieve maturity and satisfy the large dynamic range essential to overcome the different power levels,
output power requirement for PONs. At the receiver side, but also fast convergence in both the BM-TIA and the equal-
the challenge is more severe. Currently, the industrial 10G izer is required to reduce the BM overhead and improve the US
avalanche photodiode (APD) is a normal incidence type whose traffic efficiency [13].
responsivity is approximately 0.9 A/W, and the multiplication
gain is about 8. When increasing to 25G, the typical respon-
A. Linear Equalization
sivity and gain decrease to about 0.6 A/W and 5 (limited by
the gain–bandwidth product). The best academic result for a We measured the frequency responses of a 25G O-band
III-V based 50G APD is 0.5 A/W responsivity at 35 GHz with CWDM wavelength externally modulated laser (EML) trans-
M = 3 [9]. The Si-Ge based APD has better performance and mitter optical subassembly (TOSA) and a 25G APD receiver
the best case, using waveguide structure, was shown at OFC optical subassembly (ROSA), which are shown in Fig. 2. The
2019, with 0.36 A/W responsivity and M = 10 at 36 GHz 3 dB bandwidth is about 17 GHz, and the cascaded bandwidth
[10]. In addition to the sensitivity deterioration from increas- for the system is approximately 15 GHz, resulting in a low pass
ing signal baud rate, this high speed APD’s inherent properties property of the frequency domain in both the transmitter and
will further make 50G signal sensitivity worse. Therefore, con- receiver sides. Precompensation and postcompensation are
sidering the combined effects of optical component maturity both adopted to combat high frequency loss in both sides. At

Authorized licensed use limited to: CMU Libraries - library.cmich.edu. Downloaded on August 17,2020 at 18:24:49 UTC from IEEE Xplore. Restrictions apply.
 Research Article Vol. 12, No. 9 / September 2020 / Journal of Optical Communications and Networking D3

B. Nonlinear Equalization
While linear feed-forward equalization is an effective approach
to mitigate the distortions caused by low-bandwidth devices,
it also enhances the high-frequency noise, which limits its
application in systems with severe bandwidth-narrowing
effect. On the other hand, although the downstream trans-
mission operates at an O band wavelength of 1342 nm, the
worst case accumulated chromatic dispersion can be as high as
Fig. 2. Measured frequency response for a 25G (a) transmitter and 80 ps/nm for 20 km G.652 fiber length. Chromatic dispersion
(b) receiver. can be easily compensated for under coherent detection with
an analytical formula, but it becomes nonlinear distortion
with an intensity modulation and direct detection system.
5 Therefore, the maximum likelihood sequence estimation
(MLSE) approach is preferred for its outstanding ability of
4
mitigating nonlinear distortion and its mature deployment
3
in current commercial long-haul and short-reach optical
interconnect systems.
2 To reduce the complexity of MLSE, a whitening filter is
Tap Value

first used between the linear equalizer and MLSE to reduce

1
and whiten the noise but introduce the ISI at the same time
0 [22]. MLSE is then introduced after the whitening filter to
mitigate the impact of the known ISI for optimal performance.
-1 To derive the whitening filter, we take the autoregressive (AR)
-2
spectral estimate of the power spectral density of the noise [23].
The whitening filter taps are coefficients of the AR model of
-3 the noise. Then, the Viterbi algorithm is used in the MLSE
0 50 100 150
Samples(T-Space)
implementation [24]. Figure 4 is the simulated bit error rate
(BER) performance when typical 25G-class optoelectronics
Fig. 3. Impulse response from the experimental setup based on a
are adopted. The transceiver bandwidth is simulated by a 5th-
25G transceiver.
order 18 GHz Bessel filter. It is shown that an obvious BER
floor is observed when only a linear feedforward equalization
the transmitter side, a linear equalizer is utilized to open the (FFE) is adopted. By contrast, MLSE can remarkably improve
signal eye. It is realized by either frequency domain equaliza- the sensitivity by more than 2 dB at the BER threshold of
tion (FDE) or time domain equalization (TDE) with preset 1 × 10−2 . Therefore, MLSE is an effective approach to alle-
typical finite impulse response (FIR) filter taps [14]. At the viate the noise enhancement of FFE in bandwidth-limited
receiver side, the ISI can be compensated for by a similar 50G-PON systems. Figure 5 shows MLSE’s effects in mitigat-
equalizer with either blind or trained approaches to converge ing dispersion induced power penalty in another simulation.
the FIR filter taps. The impulse response from the receiver In the simulation environment of VPItransmissionMaker,
side converged FIR filter tap values can be obtained as shown we set the chirp factor of an EML as 0.8, the fiber dispersion
in Fig. 3. Depending on the degree of device limitation, the coefficient as 3.8 ps/nm/km, and the transceiver bandwidth to
filter with multiple taps is necessary to cover the range of such a typical 18 GHz. Due to the linear equalizer’s property, the
impulse broadening. It is well known that, when the number dispersion penalty can be as large as 2.5 dB. With the help of
of taps exceeds some extent that depends on the equalizer
architectural design, FDE will have less power consumption
than TDE [15,16]. To further reduce the power consumption
of FDE, the fast Fourier transform calculations during the
equalization can be implemented by the split-radix method
rather than the typical radix-2 method [17].
For the US transmission, the equalizer is required to
converge rapidly to support burst reception. Generally, the
precalculation coefficient method is used, which first generates
the FIR filter coefficients for each ONU during the ONU acti-
vation stage and is then configured to each ONU burst during
the burst US transmission [18,19]. Also, to reduce the iteration
steps, the adaptive convergence coefficients algorithm [20] and
sign change detection can be utilized [21]. The performance
of a BM equalizer has a close relationship with the BM-TIA Fig. 4. BER performance versus receiver optical power (ROP)
design. The latest research on this topic is reviewed in [13]. with MLSE.

Authorized licensed use limited to: CMU Libraries - library.cmich.edu. Downloaded on August 17,2020 at 18:24:49 UTC from IEEE Xplore. Restrictions apply.
 D4 Vol. 12, No. 9 / September 2020 / Journal of Optical Communications and Networking Research Article

-1 BTB
10
20km-FFE
20km-MLSE

-2
10
BER

-3
10

-4
10

-28 -27 -26 -25 -24 -23 -22 -21

ROP[dBm]
Fig. 7. Experimentally measured BER performances of SD-LDPC
and HD-LDPC with rate 0.826, in comparison with that reported in
Fig. 5. BER performance versus ROP over fiber transmission with [25].
and without MLSE.

14,392 bits, and a rate of 0.849. For an output BER of 10−12 ,

MLSE’s nonlinear equalizer function, this power penalty can the rate-0.849 reference HD-LDPC code, the rate-0.826 HD-
be reduced to less than 0.5 dB. LDPC, and the rate-0.826 SD-LDPC require the input BER
to be 1.1 × 10−2 , 1.25 × 10−2 , and 2.4 × 10−2 , respectively.
C. Advanced FEC At the same code rate of 0.826, the SD decoder is 1.1 dB better
than the HD decoder in terms of net coding gain (NCG). This
FEC with high coding gain can contribute to achieving higher
is consistent with the theoretical NCG difference between
link budget. The IEEE 802.3ca 25G-EPON standard has
SD and HD for a rate of about 0.8 [28]. Compared to the
adopted high-performance, hard-decision (HD), low-density
reference HD-LDPC with a rate of 0.849, the SD-LDPC with
parity check (LDPC) FEC with a BER threshold of about
a rate of 0.826 provides ∼1.3 dB more gross coding gain at an
10−2 and a rate of 0.849 [25]. For 50G-PON, the link bud-
output BER of 10−12 . This additional coding gain is expected
get is more challenging. To fully exploit the DSP function,
to translate into ∼0.9 dB lower received optical power, and
soft-decision (SD) FEC is preferred, which can utilize the
thus higher link budget [27].
linear equalizer’s output or MLSE output through the BCJR
For the upstream direction, the case is a little more complex
algorithm. By reusing the LDPC mother code approved by
and needs further study [29]. The PON system is operated in
IEEE 802.3ca, we achieved a better coding gain with a slightly
burst mode for upstream so that the channel is no longer an
reduced code rate down to 0.826 in a real-time demonstration
AWGN channel and bit error distribution may be nonuni-
[26]. Figure 6 shows the parity-check matrix of the LDPC
form and more concentrated at the beginning portion of each
mother code, which is a 12 × 69 quasi-cyclic matrix with
upstream burst.
circulant size of 256. The codeword length is 256 × 69 bits or
17,664 bits, the payload length is 256 × 57 bits or 14,592 bits,
and the parity length is 12 × 69 bits or 3072 bits. The resulting D. Experimental Results
code rate is 57/69 or 0.826. The LDPC decoder is imple- To verify the 50G TDM-PON feasibility, we built a 50G
mented in a real-time platform based on a field-programmable TDM-PON system setup using a commercial O-band
gate array (FPGA). The real-time processing speed is about 1330 nm 25G EML TOSA and 25G APD ROSA as shown in
0.1 Gb/s. The 0.75 scaled minsum algorithm is used in the Fig. 8. The 50 GBaud NRZ signal is generated in an arbitrary
decoder, whose details can be found in [27]. waveform generator (AWG, Keysight M8196A) whose sam-
Figure 7 shows the experimentally measured BER per- pling rate is set to 90 GSa/s. After amplification by an electrical
formance of SD-LDPC in comparison with HD-LDPC. The driver, the signal is injected into the 25G EML TOSA through
simulated BER performance of the HD-LDPC chosen by the the bias-T, combined with a dc bias signal. Due to the limited
IEEE 802.3ca is also plotted here as a reference. This reference bandwidth of the AWG, we implement precompensation at the
HD-LDPC code employs the same mother code but has 512 transmitter side to mitigate this effect. The eye diagrams with
parity bits punctured and 200 payload bits shortened, result-
and without precompensation are shown in Fig. 9. Because the
ing in a codeword length of 16,952 bits, a payload length of
operating wavelength should be 1342 nm instead of 1330 nm,
we use a fiber length of 34 km to emulate the CD impairment.
After fiber transmission, an optical attenuator is used to control
the received optical power. The signal is received and captured
by a real time oscilloscope (Keysight DSAZ594A) with a sam-
pling rate of 160 GSa/s. The captured data signal is recovered
by offline DSP. During the DSP, the signal is first normalized
and timing recovered, then linearly equalized by 41 tap LMS-
Fig. 6. Mother code H matrix structure of the LDPC code. FFE, whiten filtered and nonlinear equalized by 1-memory

Authorized licensed use limited to: CMU Libraries - library.cmich.edu. Downloaded on August 17,2020 at 18:24:49 UTC from IEEE Xplore. Restrictions apply.
 Research Article Vol. 12, No. 9 / September 2020 / Journal of Optical Communications and Networking D5

Fig. 8. Experimental setup.

Fig. 11. IEEE adopted 2∗25G wavelength stacked scheme for

50G-PON.

DSP is indeed a new thing in the PON field, but unlike

Fig. 9. Eye diagrams without and with precompensation. the optical components whose craft has to be redeveloped to
satisfy the PON system’s harsh requirements, as mentioned
above, the electronics parts’ craft can be quickly borrowed from
10
-1 the data center market. According to LightCounting’s report
on the market for PAM4 and coherent DSP, the demand for
BER@2e-2
-2
PAM4 DSP used in 400GE/200GE/50GE transceivers was
10
BER@1e-2 very strong in 2019, and the volume will reach 10 million
by 2023 [30]. The high volume will drive the selling price to
10
-3 decline steeply. According to LightCounting’s report, the price
of a chipset with the largest volume is predicted to be as low as
BER

-4
$12 [30]. In spite of not being exactly the same DSP chipset
10
as for 50G-PON DSP, this can be set as a reference for the
oB2B_FFE only
PON market. The DSP is manufactured using mixed-signal
-5
10 oB2B_FFE+MLSE technology. The analog functions for both cases are similar,
34km_FFE only and the difference in the digital functions is small, just like the
34km_FFE +MLSE
10
-6 scale of filter taps and different FEC choices. Therefore, the
-30 -28 -26 -24 -22 -20 -18 -16 scale of the gates and the size would probably be almost the
ROP(dBm)
same, especially using the advanced 7 nm chip manufacturing
Fig. 10. Experimental offline results in the OBTB and 34 km fiber
process. We believe, considering the shipment volumes of
transmission cases.
PON systems, that the DSP chipset price will probably be
even less than $10. This decreased cost will make the bit cost
of single carrier 50G-PON cheaper than that of single carrier
MLSE, and finally the BER is calculated. The result is shown
25G-PON, not to mention 2∗25G PON.
in Fig. 10. Both FFE only and the FFE + MLSE configuration
The IEEE 2∗25G wavelength stacked 50G system includes
are verified in the optical back-to-back (OBTB) and 34 km
fiber transmission cases. In both cases, the FFE + MLSE con- a double 25G component count. That means the system cost
figuration performs better than FFE, which demonstrates the will also be at least doubled compared with 1∗25G PON.
MLSE ability of combating residual ISI, noise, and CD. The Besides that, there is another issue. A multiplexer and demul-
sensitivity can be achieved at –27 and –25.9 dBm at a BER of tiplexer are needed before the transmitter and receiver at both
1 × 10−2 for the OBTB and 34 km fiber transmission cases, the OLT and ONT sides, as shown in Fig. 11. The additional
respectively. insertion loss for one direction can be as big as 2 dB. This will
impose more stringent requirements on either transmitter
output power or the receiver sensitivity, which will add more
3. FEASIBILITY ANALYSIS FOR THE DSP BASED burden and cost to the industry. Also, compared with the
SOLUTION single wavelength case, two wavelengths will make the system’s
Although the fact that DSP can enable 50G NRZ transmis- operation and maintenance more difficult [4].
sion over a physical link is easily accepted, there are still some
concerns about its implementation. The success of the PON B. TDEC Metric for DSP-Enabled 50G-PON
system relies on a healthy industry supply chain, better with
low cost and interoperability among different vendors to ease Transmitter compliance testing is important to enabling a
deployments. The analyses of these concerns are as follows. wide range of equipment from different vendors/suppliers to
operate together. In this compliance test, reliable metrics to
define and quantify the performance of transmitters is the key
A. Cost Analysis technology. Traditionally, an eye mask metric provides a coarse
To implement DSP in the PON system modules, the first method to screen out bad transmitters. However, an eye mask
question is what is the cost of the DSP chipset, because the is not an accurate method and cannot ensure the compliance
PON system, especially at the ONT side, is sensitive to cost. of the transmitter especially in high data rate links when DSP

Authorized licensed use limited to: CMU Libraries - library.cmich.edu. Downloaded on August 17,2020 at 18:24:49 UTC from IEEE Xplore. Restrictions apply.
 D6 Vol. 12, No. 9 / September 2020 / Journal of Optical Communications and Networking Research Article

is adopted in the receiver. Recently in the ITU 50G-PON dis- (typically enhances) the noise that is added to the receiver. In
cussions, a general consensus on a new transmitter metric for this condition, Eq. (3) can be changed to
PON—transmitter and dispersion eye closure (TDEC)—was   
y −Pavg
R
reached [31]. In TDEC, only a minor modification is needed
1 f u (y )Q C eq ·σDUT
dy
to include the impact of the receiver-side equalization. This 
f u (y )dy
R
metric has already been standardized in 100G–400G Ethernet 2
interconnects where it is often referred to as TDECQ (‘Q’   
stands for quaternary) when PAM4 signals are used [32]. The Pavg −y
R
1 f l (y ) Q C eq ·σDUT
dy
basic idea of TDEC is to compare the amount of noise we can +   = BERtarget , (4)
f l (y )dy
R
add to the device under test (DUT) and an ideal signal to reach 2
the BER at the FEC limit. Consequently, the calculation of
TDEC can be divided into three steps as follows. where C eq is the noise enhancement factor caused by the refer-
ence equalizer and is expressed as
1. Estimating the Noise σ DU T That We Can Add to the sZ
Received Signal from the DUT
C eq = N( f ) × |Heq ( f )|2 d f , (5)
Mathematically, σDUT can be estimated by finding a value such f
that the following equation is satisfied:
   where N( f ) is the normalized noise spectrum, and Heq ( f ) is
y −Pavg
R
1  f u (y )Q σDUT dy 
the frequency response of the reference equalizer.
Another modification in 50G-PON for TDEC is to adapt
f u (y )dy
R
2 for APDs, in which shot noise should be considered, because
   the original TDEC method was only based on a simple
Pavg −y thermal-noise-limited PIN receiver. As indicated in [31], for
R
1  f l (y )Q σDUT dy 
+ = BERtarget , (1) the APD receiver, an additional constraint M is defined for the
f l (y )dy
R
2 TDEC calculation such that M = σσ G1 , where σ G0 and σ G1
G0
are the noise for input levels ‘0’ and ‘1’, respectively. The value
where f u (y ) and f l (y ) are the upper and lower distributions, of M depends on many factors, such as the extinction ratio,
respectively, of the scope samples in the eye; Pavg is the average OMA, and APD parameters, and is typically about 1.5.
value of the scope samples; σDUT is the standard deviation of
the noise added to the DUT; BERtarget is the BER of the FEC
limit; and Q(x) is the area under a normal curve larger than x . C. Prototype of DSP-Aided 50G-PON Demonstration
y −P
In this way, Q σ avg can be regarded as the error probability Previously, several demonstrations based on offline signal
DUT
when adding noise to samples with y amplitude level. In this processing show the feasibility of 50G TDM-PON. However,
equation, we can easily obtain the estimated noise by increasing it is even more convincing and valuable to show the industry
σDUT in the software until the left side reaches BERtarget . that a real-time 50G TDM-PON system can work well using
current commercially available optical components and DSP
technologies.
2. Estimating the Noise σid e al That We Can Add to an Ideal
Signal
Figure 12 shows our FPGA based real-time asymmetric
50G-PON system with 50 Gbit/s downstream and 25 Gbit/s
In fact, σideal can be easily represented with the following: burst mode upstream transmission. The downstream transmis-
sion is realized by a 25G EML TOSA and 25G APD ROSA,
σideal = OMA/ (2 · Q) , (2)
while for the upstream, a 10G DML TOSA and 10G APD
where Q is the theoretical Q factor value at the target BER. For ROSA. Each FPGA platform consists of four Virtex UltraScale
example, in 50G-PON, Q is 2.326 at a BER of 1 × 10−2 , and VU440 FPGAs that implement DSP and MAC functions.
Q is 2.054 at a BER of 2 × 10−2 . The interposer board consists of time-interleaved DAC and
ADCs with a sampling rate of 33.33 GSa/s and a resolution of
6 bits. With 128 lane parallel processing, the FPGA clock is
3. TDEC Calculation
260.4 MHz. The four FPGAs are partitioned as follows:
At last, the TDEC is calculated as
• FPGA_A: Transmitter processing;
σideal
 
TDEC = 10 · log . (3) • FPGA_B: Receiver synchronization and equalization;
σDUT • FPGA_C: MAC;
• FPGA_D: Ethernet framing.
Previously, DSP was only considered in TDECQ standards
for PAM4 signals. However, in 50G-PON, due to the tight
link budget, DSP is also required for NRZ signals. When The FPGA has 2880 DSPs and three super logic regions
DSP is included in TDEC, the reference equalizer not only (SLRs). Each SLR has 960 DSPs. The burst mode EQ is imple-
improves the transmitter signal quality, but also influences mented on FPGA_B. After 20 km SSMF transmission, the

Authorized licensed use limited to: CMU Libraries - library.cmich.edu. Downloaded on August 17,2020 at 18:24:49 UTC from IEEE Xplore. Restrictions apply.
 Research Article Vol. 12, No. 9 / September 2020 / Journal of Optical Communications and Networking D7

industry. Although the 50G TDM-PON application will

possibly emerge in about 2023 in accordance with the observed
PON industry cycle, DSP can help accelerate its deployment
like convergence with wireless X-haul transmission or in the
case of emerging services consuming large bandwidth.

Funding. Huawei Technologies.

Fig. 12. FPGA based real-time prototype setup. REFERENCES

1. “Overview—Wireline Access Equipment Market,” Ovum.
2. E. Harstead, “Leveraging the 25G ecosystem for low cost
optical components,” IEEE P802.3ca Meeting (2016). http://
www.ieee802.org/3/ca/public/meeting_archive/2016/03/harstead_
3ca_1a_0316.pdf.
3. E. Harstead, D. van Veen, V. Houtsma, and P. Dom, “Technology
roadmap for time-division multiplexed passive optical networks
(TDM PONs),” J. Lightwave Technol. 37, 657–664 (2019).
4. B. Wang and D. Zhang, “Reconsiderations on 50G and 100G
EPON,” IEEE P802.3ca Meeting (2017), http://www.ieee802.org/3/
ca/public/meeting_archive/2017/07/wangbo_3ca_2_0717.pdf.
5. IEEE P802.3ca 50 Gb/s Ethernet Passive Optical Networks Task
Force, http://www.ieee802.org/3/ca/index.shtml.
6. “PON Transmission Technologies Above 10 Gbit/s Per Wavelength,”
ITU-T G.Sup64 Supplement (2018).
7. “Higher Speed Passive Optical Networks,” ITU-T G.9804.x Series of
Recommendations, G.9804.1 consented (2019).
8. “Characteristics of a single-mode optical fibre and cable,” ITU-T
Recommendation G.652 (2009).
9. M. Nada, T. Yoshimatsu, F. Nakajima, K. Sano, and H. Matsuzaki,
“A 42-GHz bandwidth avalanche photodiodes based on III-V com-
Fig. 13. EQ convergence test result. pounds for 106-Gbit/s PAM4 applications,” J. Lightwave Technol.
37, 260–265 (2019).
10. M. Huang, P. Cai, S. Li, G. Hou, N. Zhang, T. Su, C. Hong, and D.
Pan, “56 GHz waveguide Ge/Si avalanche photodiode,” in Optical
results are shown in Fig. 12. The real-time prototype demon-
Fiber Communication Conference and Exposition (2018).
strates the feasibility of 50G TDM-PON with more than 11. F. Chang, S. Bhoja, J. Riani, I. Hosagrahar, J. Wu, S. Herlekar, A.
29 dB power budget. Tiruvur, P. Khandelwal, and K. Gopalakrishnan, “Link performance
Another advantage of DSP for TDM-PON is its capability investigation of industry first 100G PAM4 IC chipset with real-time
for fast burst mode reception. As shown in Fig. 13, the burst DSP for data center connectivity,” in Optical Fiber Communications
Conference and Exhibition (OFC), Anaheim, CA (2016).
arrives at 3810 ns. At the beginning of the burst, there are
12. Z. Zhang, C. Li, J. Chen, T. Ding, Y. Wang, H. Xiang, Z. Xiao, L. Li, M.
about 20 ns guard bits and a sync word, and the burst mode Si, and X. Cui, “Coherent transceiver operated at 61 GBaud/s,” Opt.
detection is implemented in the pipeline. By presetting the tap Express 23, 18988–18995 (2015).
coefficients of FFE [19,33], the total latency for burst mode 13. G. Coudyzer, P. Ossieur, L. Breyne, A. La Porta, S. Paredes, J.
recovery is about 200 ns (or at 4010 ns in the figure). Note that Bauwelinck, and X. Yin, “Study of burst-mode adaptive equalization
for >25G PON applications [Invited],” J. Opt. Commun. Netw. 12,
the processing latency can be further reduced by using an ASIC
A104–A112 (2020).
(instead of FPGA) due to the resource limitation of FPGA. 14. D. Rafique, T. Rahman, A. Napoli, and B. Spinnler, “Digital pre-
emphasis in optical communication systems: on the nonlinear
performance,” J. Lightwave Technol. 33, 140–150 (2015).
4. CONCLUSIONS 15. B. Spinnler, F. N. Hauske, and M. Kuschnerov, “Adaptive equal-
izer complexity in coherent optical receivers,” in 34th European
Considering the ongoing massive deployment of 10G-PON Conference on Optical Communication (2008).
and infrastructure investment, the next generation PON is 16. M. S. Faruk and K. Kikuchi, “Adaptive frequency-domain equali-
preferred to be 50G instead of 25G. Compared with 2∗25G, zation in digital coherent optical receivers,” Opt. Express 19,
single carrier 50G based on DSP has obvious advantages, espe- 12789–12798 (2011).
cially cost and maintenance. With the aid of DSP to overcome 17. H. Abdollahifakhr, N. Bélanger, Y. Savaria, and F. Gagnon, “Power-
efficient hardware architecture for computing Split-Radix FFTs on
various challenges, 50G-PON can be realized based on more highly sparsed spectrum,” in IEEE 13th International New Circuits
mature 25G components. Although DSP is a brand-new tech- and Systems Conference (2015).
nique for PON application, it is already mature and reliable 18. F. J. Effenberger, H. Zeng, A. Shen, and X. Liu, “Burst-mode error
in many other optical fiber transmission scenarios. Via wide distribution and mitigation in DSP-assisted high-speed PONs,”
implementation in the data center interconnection, the DSP J. Lightwave Technol. 38, 754–760 (2020).
19. R. Koma, M. Fujiwara, J. Kani, K. Suzuki, and A. Otaka, “Burst-
chipset can be quickly realized in the PON field. Furthermore, mode digital signal processing that pre-calculates FIR filter
with tens of millions of shipments of the PON system, the coefficients for digital coherent PON upstream,” J. Opt. Commun.
DSP chipset cost will drop greatly and in return promote the Netw. 10, 461–470 (2018).

Authorized licensed use limited to: CMU Libraries - library.cmich.edu. Downloaded on August 17,2020 at 18:24:49 UTC from IEEE Xplore. Restrictions apply.
 D8 Vol. 12, No. 9 / September 2020 / Journal of Optical Communications and Networking Research Article

20. T. Aboulnasr and K. Mayyas, “A robust variable step-size LMS-type 27. M. Yang, L. Li, X. Liu, and I. Djordjevic, “FPGA-based real-time soft-
algorithm: analysis and simulations,” IEEE Trans. Signal Process. decision LDPC performance verification for 50G-PON,” in Optical
45, 631–639 (1997). Fiber Communication Conference and Exposition (2019).
21. R. Harris, D. Chabries, and F. Bishop, “A variable step (VS) adaptive 28. T. Mizuochi, “Recent progress in forward error correction and its
filter algorithm,” IEEE Trans. Acoust. Speech Signal Process. 34, interplay with transmission impairments,” IEEE J. Sel. Top. Quantum
309–316 (1986). Electron. 12, 544–554 (2006).
22. J. Li, E. Tipsuwannakul, T. Eriksson, M. Karlsson, and P. A. 29. M. Yang, L. Li, X. Liu, and I. Djordjevic, “Real-time verification of
Andrekson, “Approaching Nyquist limit in WDM systems by low- soft-decision LDPC coding for burst mode upstream reception in
complexity receiver-side duobinary shaping,” J. Lightwave Technol. 50G-PON,” J. Lightwave Technol. 38, 1693–1701 (2020).
30, 1664–1676 (2012). 30. “Market for PAM4 and coherent DSPs,” in LightCounting (2019).
23. L. Liu, L. Li, and Y. Lu, “Detection of 56GBaud PDM-QPSK gener- 31. G. Le Cheminant, K. Zhang, V. Houtsma, E. Harstead, and X.
ated by commercial CMOS DAC with 11 GHz analog bandwidth,” in Liu, “TDEC (transmitter dispersion and eye-closure) method for
European Conference on Optical Communication (2014). equalizer-enabled 50G-PON,” in ITU-T Q2 Meeting (2019).
24. O. E. Agazzi, M. R. Hueda, H. S. Carrer, and D. E. Crivelli, 32. “IEEE Standard for Ethernet - Amendment 10: Media Access
“Maximum-likelihood sequence estimation in dispersive optical Control Parameters, Physical Layers, and Management
channels,” J. Lightwave Technol. 23, 749–763 (2005). Parameters for 200 Gb/s and 400 Gb/s Operation,”
25. M. Laubach, S. Yang, Y. Han, R. Hirth, and G. Kramer, https://ieeexplore.ieee.org/document/8207825.
“LDPC Adjustments from Motion #6, Chicago,” IEEE 33. L. Anet Neto, J. Maes, P. Larsson-Edefors, J. Nakagawa, K.
P802.3ca Meeting (2018), http://www.ieee802.org/3/ca/public/ Onohara, and S. J. Trowbridge, “Considerations on the use of digital
meeting_archive/2018/05/laubach_3ca_1_0518.pdf. signal processing in future optical access networks,” J. Lightwave
26. L. Li, X. Liu, and F. Effenberger, “Soft-decision LDPC for 50G-PON,” Technol. 38, 598–607 (2020).
in ITU-T Q2 Interim Meeting (2018), paper D75.

Authorized licensed use limited to: CMU Libraries - library.cmich.edu. Downloaded on August 17,2020 at 18:24:49 UTC from IEEE Xplore. Restrictions apply.