Merits and challenges of optical transmission at 64 Gbaud
u2t Photonics announced recently a balanced detector that supports 64Gbaud. This promises coherent transmission systems with double the data rate. But even if the remaining components - the modulator and DSP-ASIC capable of operating at 64Gbaud - were available, would such an approach make sense?
Gazettabyte asked system vendors Transmode and Ciena for their views.
Transmode:
Transmode points out that 100 Gigabit dual-polarisation, quadrature phase-shift keying (DP-QPSK) using coherent detection has several attractive characteristics as a modulation format.
It can be used in the same grid as 10 Gigabit-per-second (Gbps) and 40Gbps signals in the C-band. It also has a similar reach as 10Gbps by achieving a comparable optical signal-to-noise ratio (OSNR). Moreover, it has superior tolerance to chromatic dispersion and polarisation mode dispersion (PMD), enabling easier network design, especially with meshed networking.
The IEEE has started work standardising the follow-on speed of 400 Gigabit. "This is a reasonable step since it will be possible to design optical transmission systems at 400 Gig with reasonable performance and cost," says Ulf Persson, director of network architecture in Transmode's CTO office.
Moving to 100Gbps was a large technology jump that involved advanced technologies such as high-speed analogue-to-digital (A/D) converters and advanced digital signal processing, says Transmode. But it kept the complexity within the optical transceivers which could be used with current optical networks. It also enabled new network designs due to the advanced chromatic dispersion and PMD compensations made possible by the coherent technology and the DSP-ASIC.
For 400Gbps, the transition will be simpler. "Going from 100 Gig to 400 Gig will re-use a lot of the technologies used for 100 Gig coherent," says Magnus Olson, director of hardware engineering.
So even if there will be some challenges with higher-speed components, the main challenge will move from the optical transceivers to the network, he says. That is because whatever modulation format is selected for 400Gbps, it will not be possible to fit that signal into current networks keeping both the current channel plan and the reach.
"From an industry point of view, a metro-centric cost reduction of 100Gbps coherent is currently more important than increasing the bit rate to 400Gbps"
"If you choose a 400 Gigabit single carrier modulation format that fits into a 50 Gig channel spacing, the optical performance will be rather poor, resulting in shorter transmission distances," says Persson. Choosing a modulation format that has a reasonable optical performance will require a wider passband. Inevitably there will be a tradeoff between these two parameters, he says.
This will likely lead to different modulation formats being used at 400 Gig, depending on the network application targeted. Several modulation formats are being investigated, says Transmode, but the two most discussed are:
- 4x100Gbps super-channels modulated with DP-QPSK. This is the same as today's modulation format with the same optical performance as 100Gbps, and requires a channel width of 150GHz.

- 2x200Gbps super-channels, modulated with DP-16-QAM. This will have a passband of about 75GHz. It is also possible to put each of the two channels in separate 50GHz-spaced channels and use existing networks The effective bandwidth will then be 100GHz for a 400GHz signal. However, the OSNR performance for this format is about 5-6 dB worse than the 100Gbps super-channels. That equates to about a quarter of the reach at 100Gbps.

As a result, 100Gbps super-channels are more suited to long distance systems while 200Gbps super-channels are applicable to metro/ regional systems.
Since 200Gbps super-channels can use standard 50GHz spacing, they can be used in existing metro networks carrying a mix of traffic including 10Gbps and 40Gbps light paths.
"Both 400 Gig alternatives mentioned have a baud rate of about 32 Gig and therefore do not require a 64 Gbaud photo detector," says Olson. "If you want to go to a single channel 400G with 16-QAM or 32-QAM modulation, you will get 64Gbaud or 51Gbaud rate and then you will need the 64 Gig detector."
The single channel formats, however, have worse OSNR performance than 200Gbps super-channels, about 10-12 dB worse than 100Gbps, says Transmode, and have a similar spectral efficiency as 200Gbps super-channels. "So it is not the most likely candidates for 400 Gig," says Olson. "It is therefore unclear for us if this detector will have a use in 400 Gigabit transmission in the near future."
Transmode points out that the state-of-the-art bit rate has traditionally been limited by the available optics. This has kept the baud rate low by using higher order modulation formats that support more bits per symbol to enable existing, affordable technology to be used.
"But the price you have to pay, as you can not fool physics, is shorter reach due to the OSNR penalty," says Persson.
Now the challenges associated with the DSP-ASIC development will be equally important as the optics to further boost capacity.
The bundling of optical carriers into super-channels is an approach that scales well beyond what can be accomplished with improved optics. "Again, we have to pay the price, in this case eating greater portions of the spectrum," says Persson.
Improving the bandwidth of the balanced detector to the extent done by u2t is a very impressive achievement. But it will not make it alone into new products, modulators and a faster DSP-ASIC will also be required.
"From an industry point of view, a metro-centric cost reduction of 100Gbps coherent is currently more important than increasing the bit rate to 400Gbps," says Olson. "When 100 Gig coherent costs less than 10x10 Gig, both in dollars and watts, the technology will have matured to again allow for scaling the bit rate, using technology that suits the application best."
Ciena:
How the optical performance changes going from 32Gbaud to 64Gbaud depends largely on how well the DSP-ASIC can mitigate the dispersion penalties that get worse with speed as the duration of a symbol narrows.
BPSK goes twice as far as QPSK which goes about 4.5 times as far as 16-QAM
"I would also expect a higher sensitivity would be needed also, so another fundamental impact," says Joe Berthold, vice president of network architecture at Ciena. "We have quite a bit or margin with the WaveLogic 3 [DSP-ASIC] for many popular network link distances, so it may not be a big deal."
With a good implementation of a coherent transmission system, the reach is primarily a function of the modulation format. BPSK goes twice as far as QPSK which goes about 4.5 times as far as 16-QAM, says Berthold.
"On fibres without enough dispersion, a higher baud rate will go 25 percent further than the same modulation format at half of that baud rate, due to the nonlinear propagation effects," says Berthold. This is the opposite of what occurred at 10 Gigabit incoherent. On fibres with plenty of local dispersion, the difference becomes marginal, approximately 0.05 dB, according to Ciena.
Regarding how spectral efficiency changes with symbol rate, doubling the baud rate doubles the spectral occupancy, says Berthold, so the benefit of upping the baud rate is that fewer components are needed for a super-channel.
"Of course if the cost of the higher speed components are higher this benefit could be eroded," he says. "So the 200 Gbps signal using DP-QPSK at 64 Gbaud would nominally require 75GHz of spectrum given spectral shaping that we have available in WaveLogic 3, but only require one laser."
Does Ciena have an view as to when 64Gbaud systems will be deployed in the network?
Berthold says this hard to answer. "It depends on expectations that all elements of the signal path, from modulators and detectors to A/D converters, to DSP circuitry, all work at twice the speed, and you get this speedup for free, or almost."
The question has two parts, he says: When could it be done? And when will it provide a significant cost advantage? "As CMOS geometries narrow, components get faster, but mask sets get much more expensive," says Berthold.
ROADMs and their evolving amplification needs
Technology briefing: ROADMs and amplifiers
Oclaro announced an add/drop routing platform at the recent OFC/NFOEC show. The company explains how the platform is driving new arrayed amplifier and pumping requirements.
A ROADM comprising amplification, line-interfaces, add/ drop routing and transponders. Source: Oclaro
Agile optical networking is at least a decade-old aspiration of the telcos. Such networks promise operational flexibility and must be scalable to accommodate the relentless annual growth in network traffic. Now, technologies such as coherent optical transmission and reconfigurable optical add/drop multiplexers (ROADMs) have reached a maturity to enable the agile, mesh vision.
Coherent optical transmission at 100 Gigabit-per-second (Gbps) has become the base currency for long-haul networks and is moving to the metro. Meanwhile, ROADMs now have such attributes as colourless, directionless and contentionless (CDC). ROADMs are also being future-proofed to support flexible grid, where wavelengths of varying bandwidths are placed across the fibre's spectrum without adhering to a rigid grid.
Colourless and directionless refer to the ROADM's ability to transmit or drop any light path from any direction or degree at any network interface port. Contentionless adds further flexibility by supporting same-colour light paths at an add or a drop.
"You can't add and drop in existing architectures the same colour [light paths at the same wavelength] in different directions, or add the same colour from a given transponder bank," says Bimal Nayar, director, product marketing at Oclaro's optical network solutions business unit. "This is prompting interest in contentionless functionality."
The challenge for optical component makers is to develop cost-effective coherent and CDC-flexgrid ROADM technologies for agile networks. Operators want a core infrastructure with components and functionality that provide an upgrade path beyond 100 Gigabit coherent yet are sufficiently compact and low-power to minimise their operational expenditure.
ROADM architectures
ROADMs sit at the nodes of a mesh network. Four-degree nodes - the node's degree defined as the number of connections or fibre pairs it supports - are common while eight-degree is considered large.
The ROADM passes through light paths destined for other nodes - known as optical bypass - as well as adds or drops wavelengths at the node. Such add/drops can be rerouted traffic or provisioned new services.
Several components make up a ROADM: amplification, line-interfaces, add/drop routing and transponders (see diagram, above).
"With the move to high bit-rate systems, there is a need for low-noise amplification," says Nayar. "This is driving interest in Raman and Raman-EDFA (Erbium-doped fibre amplifier) hybrid amplification."
The line interface cards are used for incoming and outgoing signals in the different directions. Two architectures can be used: broadcast-and-select and route-and select.
With broadcast-and-select, incoming channels are routed in the various directions using a passive splitter that in effect makes copies the incoming signal. To route signals in the outgoing direction, a 1xN wavelength-selective switch (WSS) is used. "This configuration works best for low node-degree applications, when you have fewer connections, because the splitter losses are manageable," says Nayar.
For higher-degree node applications, the optical loss using splitters is a barrier. As a result, a WSS is also used for the incoming signals, resulting in the route-and-select architecture.
Signals from the line interface cards connect to the routing platform for the add/drop operations. "Because you have signals from any direction, you need not a 1xN WSS but an LxM one," says Nayar. "But these are complex to design because you need more than one switching plane." Such large LxM WSSes are in development but remain at the R&D stage.
Instead, a multicast switch can be used. These typically are sized 8x12 or 8x16 and are constructed using splitters and switches, either spliced or planar lightwave circuit (PLC) based .
"Because the multicast switch is using splitters, it has high loss," says Nayar. "That loss drives the need for amplification."

Add/drop platform
With an 8-degree-node CDC ROADM design, signals enter and exit from eight different directions. Some of these signals pass through the ROADM in transit to other nodes while others have channels added or dropped.
In the Oclaro design, an 8x16 multicast switch is used. "Using this [multicast switch] approach you are sharing the transponder bank [between the directions]," says Nayar.
The 8-degree node showing the add/drop with two 8x16 multicast switches and the 16-transponder bank. Source: Oclaro
A particular channel is dropped at one of the switch's eight input ports and is amplified before being broadcast to all 16, 1x8 switches interfaced to the 16 transponders.
It is the 16, 1x8 switches that enable contentionless operation where the same 'coloured' channel is dropped to more than one coherent transponder. "In a traditional architecture there would only be one 'red' channel for example dropped as otherwise there would be [wavelength] contention," says Nayar.
The issue, says Oclaro, is that as more and more directions are supported, greater amplification is needed. "This is a concern for some, as amplifiers are associated with extra cost," says Nayar.
The amplifiers for the add/drop thus need to be compact and ideally uncooled. By not needing a thermo-electrical cooler, for example, the design is cheaper and consumes less power.
The design also needs to be future-proofed. The 8x16 add/ drop architecture supports 16 channels. If a 50GHz grid is used, the amplifier needs to deliver the pump power for a 16x50GHz or 800GHz bandwidth. But the adoption of flexible grid and super-channels, the channel bandwidths will be wider. "The amplifier pumps should be scalable," says Nayar. "As you move to super-channels, you want pumps that are able to deliver the pump power you need to amplify, say, 16 super-channels."
This has resulted in an industry debate among vendors as to the best amplifier pumping scheme for add/drop designs that support CDC and flexible grid.
EDFA pump approaches
Two schemes are being considered. One option is to use one high-power pump coupled to variable pump splitters that provides the required pumping to all the amplifiers. The other proposal is to use discrete, multiple pumps with a pump used for each EDFA.
Source: Oclaro
In the first arrangement, the high-powered pump is followed by a variable ratio pump splitter module. The need to set different power levels at each amplifier is due to the different possible drop scenarios; one drop port may include all the channels that are fed to the 16 transponders, or each of the eight amplifiers may have two only. In the first case, all the pump power needs to go to the one amplifier; in the second the power is divided equally across all eight.
Oclaro says that while the high-power pump/ pump-splitter architecture looks more elegant, it has drawbacks. One is the pump splitter introduces an insertion loss of 2-3dB, resulting in the pump having to have twice the power solely to overcome the insertion loss.
The pump splitter is also controlled using a complex algorithm to set the required individual amp power levels. The splitter, being PLC-based, has a relatively slow switching time - some 1 millisecond. Yet transients that need to be suppressed can have durations of around 50 to 100 microseconds. This requires the addition of fast variable optical attenuators (VOAs) to the design that introduce their own insertion losses.
"This means that you need pumps in excess of 500mW, maybe even 750mW," says Nayar. "And these high-power pumps need to be temperature controlled." The PLC switches of the pump splitter are also temperature controlled.
The individual pump-per-amp approach, in contrast, in the form of arrayed amplifiers, is more appealing to implement and is the approach Oclaro is pursuing. These can be eight discrete pumps or four uncooled dual-chip pumps, for the 8-degree 8x16 multicast add/drop example, with each power level individually controlled.
Source: Oclaro
Oclaro says that the economics favour the pump-per-amp architecture. Pumps are coming down in price due to the dramatic price erosion associated with growing volumes. In contrast, the pump split module is a specialist, lower volume device.
"We have been looking at the cost, the reliability and the form factor and have come to the conclusion that a discrete pumping solution is the better approach," says Nayar. "We have looked at some line card examples and we find that we can do, depending on a customer’s requirements, an amplified multicast switch that could be in a single slot."
VDSL2 vectoring explained
Several system vendors including Adtran, Alcatel-Lucent and ZTE have announced vectoring technology that boosts the performance of very-high-bit-rate digital subscriber line (VDSL2) broadband access technology. Vectoring is used to counter crosstalk - signal leakage between the telephony twisted wire pairs that curtails VDSL2's bit rate performance – as is now explained.
Technology briefing
There is a large uncertainty in the resulting VDSL2 bit rate for a given loop length. With vectoring this uncertainty is almost removed
Paul Spruyt, Alcatel-Lucent
Two key characteristics of the local loop limit the performance of digital subscriber line (DSL) technology: signal attenuation and crosstalk.
Attenuation is due to the limited spectrum of the telephone twisted pair, designed for low frequency voice calls not high-speed data transmission. Analogue telephony uses only 4kHz of spectrum, whereas ADSL uses 1.1MHz and ADSL2+ 2.2MHz. The even higher speed VDSL2 has several flavours: 8b is 8.5MHz, 17a is 17.6MHz while 30a spans 30MHz.
The higher frequencies induce greater attenuation and hence the wider the spectrum, the shorter the copper loop length over which data can be sent. This is why higher speed VDSL2 technology requires the central office or, more commonly, the cabinet to be closer to the user, up to 2.5km away - although in most cases VDSL2 is deployed on loops shorter than 1.5km.
The second effect, crosstalk, describes the leakage of the signal in a copper pair into neighbouring pairs. “All my neighbours get a little bit of the signal sent on my pair, and vice versa: the signal I receive is not only the useful signal transmitted on my pair but also noise, the contributed components from all my active VDSL2 neighbours,” says Paul Spruyt, xDSL technology strategist at Alcatel-Lucent.
Typical a cable bundle comprises several tens to several hundred copper pairs. The signal-to-noise ratio on each pair dictates the overall achievable data rate to the user and on short loops it is the crosstalk that is the main noise culprit.
Vectoring boosts VDSL2 data rates to some 100 megabits-per-second (Mbps) downstream and 40Mbps upstream over 400m. This compares to 50Mbps and 20Mbps, respectively, without vectoring. There is a large uncertainty in the resulting VDSL2 bit rate for a given loop length. "With vectoring this uncertainty is almost removed," says Spruyt.
Vectoring
The term vectoring refers to the digital signal processing (DSP) computations involved to cancel the crosstalk. The computation involves multiplying pre-coder matrices with Nx1 data sets – or vectors – representing the transmit signals.
The crosstalk coupling into each VDSL2 line is measured and used to generate an anti-noise signal in the DSLAM to null the crosstalk on each line.
To calculate the crosstalk coupling between the pairs in the cable bundle, use is made of the ‘sync’ symbol that is sent after every 256 data symbols, equating to a sync symbol every 64ms or about 16 a second.
Each sync symbol is modulated with one bit of a pilot sequence. The length of the pilot sequence is dependent on the number of VDSL2 lines in the vectoring group. In a system with 192 VDSL2 lines, 256-bit-long pilot sequences are used (the next highest power of two).
Moreover, each twisted pair is assigned a unique pilot sequence, with the pilots usually chosen such that they are mutually orthogonal. “If you take two orthogonal pilots sequences and multiply them bit-wise, and you take the average, you always find zero,” says Spruyt. "This characteristic speeds up and simplifies the crosstalk estimation.”
A user's DSL modem expects to see the modulated sync symbol, but in reality sees a modulated sync symbol distorted with crosstalk from the modulated sync symbols transmitted on the neighbouring lines. The modem measures the error – the crosstalk – and sends it back to the DSLAM. The DSLAM correlates the received error values on the ‘victim’ line with the pilot sequences transmitted on all other ‘disturber’ lines. By doing this, the DSLAM gets a measure of the crosstalk coupling for every disturber – victim pair.
The final step is the generation of anti-noise within the DSLAM.
This anti-noise is injected into the victim line on top of the transmit signal such that it cancels the crosstalk signal picked up over the telephone pair. This process is repeated for each line.
Source: Alcatel-Lucent
VDSL2 uses discrete multi-tone (DMT) modulation where each DMT symbol consists of 4096 tones, split between the upstream (from the DSL modem to the DSLAM) and the downstream (to the user) transmissions. All tones are processed independently in the frequency domain. The resulting frequency domain signal including the anti-noise is converted back to the time domain using an inverse fast Fourier transform.
The above describes the crosstalk pre-compensation or pre-coding in the downstream direction: anti-noise signals are generated and injected in the DSLAM prior to transmission of the signal on the line.
For the upstream, the inverse occurs: the DSLAM generates and adds the anti-noise after reception of the signal distorted with crosstalk. This technique is known as post-compensation or post-coding. In this case the DSL modem sends the pilot modulated sync symbols and the DSLAM measures the error signal and performs the correlations and anti-noise calculations.
Challenges
One key challenge is the amount of computations to be performed in real-time. For a fully-vectored 200-line VDSL2 system, some 2,600 billion multiply-accumulates per second - 2.6TMAC/s - need to be calculated. A system of 400 lines would require four times as much processing power, about 10TMAC/s.
Alcatel-Lucent’s first-generation vectoring system that was released end 2011 could process 192 lines. At the recent Broadband World Forum show in October, Alcatel-Lucent unveiled its second-generation system that doubles the capacity to 384 lines.
For larger cable bundles, the crosstalk contributions from certain more distant disturbers to a victim line are negligible. Also, for large vectoring systems, pairs typically do not stay together in the same cable but get split over multiple smaller cables that do not interfere with each other. “There is a possibility to reduce complexity by sparse matrix computations rather than a full matrix,” says Spruyt, but for smaller systems full matrix computation is preferred as the disturbers can’t be ignored.
There are other challenges.
There is a large amount of data to be transferred within the DSLAM associated with the vectoring. According to Alcatel-Lucent, a 48-port VDSL2 card can generate up to 20 Gigabit-per-second (Gbps) of vectoring data.
There is also the need for strict synchronization – for vectoring to work the DMT symbols of all lines need to be aligned within about 1 microsecond. As such, the clock needs to be distributed with great care across the DSLAM.
Adding or removing a VDSL2 line also must not affect active lines which requires that crosstalk is estimated and cancelled before any damage is done. The same applies when switching off a VDSL2 modem which may affect the terminating impedance of a twisted pair and modify the crosstalk coupling. Hence the crosstalk needs to be monitored in real-time.
Zero touch
A further challenge that operators face when upgrading to vectoring is that not all the users' VDSL2 modems may support vectoring. This means that crosstalk from such lines can’t be cancelled which significantly reduces the vectoring benefits for the users with vectoring DSL modems on the same cable.
To tackle this, certain legacy VDSL2 modems can be software upgraded to support vectoring. Others, that can't be upgraded to vectoring, can be software upgraded to a ‘vector friendly’ mode. Crosstalk from such a vector friendly line into neighbouring vectored lines can be cancelled, but the ‘friendly’ line itself does not benefit from the vectoring gain.
Upgrading the modem firmware is also a considerable undertaking for the telecom operators especially when it involves tens or hundreds of thousands of modems.
Moreover, not all the CPEs can be upgraded to friendly mode. To this aim, Alcatel Lucent has developed a 'zero-touch' approach that allows cancelling the crosstalk from legacy VDSL2 lines into a vectored lines without CPE upgrade. “This significantly facilitates and speeds up the roll-out of vectoring” says Spruyt
Further reading:
Boosting VDSL2 Bit Rates with Vectoring
DSL: Will phantom channels become real deployments
