Ciena's multi-format 400G coherent QSFP-DD pluggable
Ciena showcased a working 400-gigabit Universal coherent pluggable module at the ECOC 2022 conference and exhibition in Basel, Switzerland.
Ciena is using its WaveLogic 5 Nano coherent digital signal processor (DSP) for the Universal QSFP-DD coherent pluggable module.
“We call it universal because it supports many transmission modes – interoperable and high performance; the most in the industry,” says Helen Xenos, senior director of portfolio marketing at Ciena.
The pluggable has custom extended-performance modes and supports three industry formats: the 400ZR interoperable standard, the 400ZR+ multi-source agreement (MSA), and the OpenROADM MSA. (See tables below).
IP over DWDM
Communications service providers (CSPs) want to add pluggable coherent modules to their IP routers, removing the need for a separate transponder card or box linking the router to the optical line system.
The advent of coherent QSFP-DD pluggables has meant the same form factor can be used for client-side and line-side optics, ensuring efficient use of the router ports.
The CSPs want the coherent QSFP-DD module to have sufficient optical performance to meet their demanding networking requirements. For example, the module’s output signal can pass through the filtering stages of reconfigurable optical add-drop multiplexers (ROADMs) along the optical link.
Optical amplification and filtering
Ciena’s coherent QSFP-DD adds a fibre-based optical amplifier and a tunable optical filter to the coherent photonics and electronic ICs.
The optical amplification enables the high-performance mode and the launching of a 4dBm output signal. In contrast, 400ZR and 400ZR+ have a launch power of -10dBm.
“This is the industry’s highest [QSDP-DD] transmit power,” says Xenos.
The tunable optical filter improves the optical performance of the coherent receiver.
In an optical line system with colourless ROADMs, the Erbium-doped fibre amplifiers (EDFAs) generate out-of-band transmission noise – amplified spontaneous emission (ASE). The noise sources superimpose and become significant, impairing wavelength and system performance dramatically.
The tunable optical filter eliminates this effect and simplifies deployment over any photonics line system. In addition, Ciena says the pluggables can now work alongside high-baud rate transponders in existing ROADM applications.
The QSFP-DD’s tunable optical filter means its optical performance closely matches that of the CFP2-DCO, aiding the two classes of pluggables working together.
Modes of operation
400ZR defines the module’s baseline coherent performance. The OIF developed the 400ZR standard so hyperscalers can link their equipment in two separate data centres up to 120km apart.
The 400ZR specification delivers just enough optical performance to meet the optical link budget. The OIF produced a low-cost, interoperable, pluggable coherent specification.
400ZR supports a single baud rate – 60 gigabaud (GBd), and modulation scheme – dual-polarisation 16-QAM, and carries Ethernet frames.
Google, Meta, Microsoft and Alibaba were all involved in the OIF development, with the 400ZR Implementation Agreement published in early 2020.
400ZR supports two-channel widths: 75GHz and 100GHz, while the forward error correction scheme used is CFEC.
The 400ZR+ MSA enhances the performance by supporting other data rates – 100, 200 and 300 gigabits-per-second (Gbps) – as well as 400Gbps. In addition, it uses several modulation schemes and the enhanced O-FEC error correction scheme that extends reach.
Ciena’s module also meets the OpenROADM MSA, supporting Ethernet and OTN and an enhanced reach at 400Gbps.
Ciena’s Universal module’s extended performance modes up the symbol rate to 65 and 70 gigabaud (GBd) and uses probabilistic constellation shaping (PCS).
PCS maps the bitstream onto the constellation to maximise the data recovery at the coherent receiver, thereby improving overall optical performance. The scheme also allows the fine-tuning of the data rate sent.
At ECOC, Ciena showed the module implementing the high-performance mode at 70GBd and PCS.
ECOC innovation award
The ECOC Exhibition Market Focus Advisory Committee awarded the most innovative product award to Ciena’s WaveLogic 5 Nano 400G Universal QSFP-DD.
The era of 400G coherent pluggables finally emerges
Part 1: 7nm coherent DSPs, ZR and ZR+
The era of 400-gigabit coherent pluggable modules has moved a step closer with Inphi’s announcement of its Canopus coherent digital signal processor (DSP) and its QSFP-DD ColorZ II optical module.
NeoPhotonics has also entered the fray, delivering first samples of its 400-gigabit ClearLight CFP2-DCO module that uses the Canopus DSP.
The ColorZ II and ClearLight modules support the 400ZR OIF standard used to link data centres up to 120km apart. They also support extended modes, known as ZR+, that is not standardised.
ZR+’s modes include 400 Gigabit-per-second (Gbps) over distances greater than 400ZR’s 120km and lower data rates over metro-regional and long-haul distances.
The announcements of the Canopus DSP and 400-gigabit pluggable coherent modules highlight the approaches being taken for ZR+. Optical module vendors are aligning around particular merchant DSPs such that interoperability exists but only within each camp.
The first camp involves Inphi and three other module vendors, one being NeoPhotonics. The second camp is based on the OpenZR+ specification that offers interoperability between the DSPs of the merchant players, Acacia Communications and NTT Electronics (NEL). Cisco is in the process of acquiring Acacia.
Market analysts, however, warn that such partial interoperability for ZR+ harms the overall market opportunity for coherent pluggables.
“ZR+ should be interoperable like ZR, and come along with the hard decisions the ZR standard required,” says Andrew Schmitt, founder and directing analyst at research form, Cignal AI.
The optical module vendors counter that only with specialist designs – designs that are multi-sourced – can the potential of a coherent DSP be exploited.
Applications
The advent of 400-gigabit coherent optics within compact client-side form factors is a notable development, says Inphi. “The industry has been waiting for this inflextion point of having, for the first time, 400-gigabit coherent pluggables that go on router and switch interfaces,” says Pranay Aiya, vice president of product marketing and applications engineering at Inphi.
“IP over DWDM has never happened; we have all heard about it till the cows come home,” says Aiya.
IP-over-DWDM failed to take off because of the power and space demands of coherent optics, especially when router and switch card slots come at a premium. Using coherent optics on such platforms meant trading off client-side faceplate capacity to fit bulkier coherent optics. This is no longer the case with the advent of QSFP-DD and OSFP coherent modules.
“If you look at the reasons why IP-over-DWDM – coloured optics directly on routers – failed, all of those reasons have changed,” says Schmitt. The industry is moving to open line systems, open network management, and more modular network design.
“All of the traffic is IP and layer-1 switching and grooming isn’t just unnecessary, it is more expensive than low-feature layer-2 switching,” says Schmitt, adding that operators will use pluggables wherever the lower performance is acceptable. Moreover, this performance gap will narrow with time.
The Canopus DSP also supports ZR+ optical performance and, when used within a CFP2-DCO module with its greater power enveloped than OSFP and QSFP-DD, enables metro and long-haul distances, as required by the telecom operators. This is what Neophotonics has announced with its ClearLight CFP2-DCO module.
Canopus
Inphi announced the Canopus DSP last November and revealed a month later that it was sampling its first optical module, the ColorZ II, that uses the Canopus DSP. The ColorZ II is a QSFP-DD pluggable module that supports 400ZR as well as the ZR+ extended modes.
Inphi says that, given the investment required to develop the 7nm CMOS Canopus, it had to address the bulk of the coherent market.
“We were not going after the ultra-long-haul and submarine markets but we wanted pluggables to address 80-90 per cent of the market,” says Aiya.
This meant developing a chip that would support the OIF’s 400ZR, 200-gigabit using quadrature phased-shift keying (QPSK) modulation for long haul, and deliver 400-gigabit over 500-600km.
The Canopus DSP also supports probabilistic constellation shaping (PCS), a technology that until now has been confined to the high-end coherent DSPs developed by the leading optical systems vendors.
With probabilistic shaping, not all the constellation points are used. Instead, those with lower energy are favoured; points closer to the origin on a constellation graph. The only time all the constellation points are used is when sending the maximum data rate for a given modulation scheme.
Choosing the inner, lower-energy constellation points more frequently than the outer points to send data reduces the average energy and improves the signal-to-noise ratio. To understand why, the symbol error rate at the receiver is dominated by the distance between neighbouring points on the constellation. Reducing the average energy keeps the distance between the points the same, but since a constant signal power level is used for DWDM transmission, applying gain increases the distance between the constellation points. The result is improved optical performance.
Probabilistic shaping also allows an exact number of bits-per-symbol to be sent, even non-integer values.
For example, using standard modulation schemes such as 64-QAM with no constellation shaping, 6 bits-per-symbol are sent. Using shaping and being selective as to which constellation points are used, 5.7 bits-per-symbol could be sent, for example. This enables finer control of the sent data, enabling operators to squeeze the maximum data rate to suit the margins on a given fibre link.
“This is the first time a DSP with probabilistic shaping has been available in a size and power that enables pluggables,” says Aiya.
The resulting optical performance using the Canopus is up to 1,500km at 300Gbps signals and up to 2,000km for 200Gbps transmissions (see Table above). As for baud rates, the DSP ranges from 30+ to the mid-60s Gigabaud.
Inphi also claims a 75 per cent reduction in power consumption of the Canopus compared to 16nm CMOS DSPs found in larger, 4×5-inch modules.
Several factors account for the sharp power reduction: the design of the chip’s architecture and physical layout, and the use of 7nm CMOS. The Canopus uses functional blocks that extend the reach, and these can be turned off to reduce the power consumption when lower optical performance is acceptable.
The architectural improvements and the physical layout account for half of the overall power savings, says Aiya, with the rest coming from using a 7nm CMOS.
The result is a DSP a third the size of 16nm DSPs. “It [pluggables] requires the DSP to be very small; it’s not a paperweight anymore,” says Aiya.
400ZR and ZR+
The main challenge for the merchant coherent DSP camps is the several, much larger 400ZR eco-systems from Ciena, Cisco and Huawei.
“Each one of these eco-systems will be larger than the total merchant market of 400ZR,” says Vladimir Kozlov, CEO and founder of LightCounting. The system vendors will make sure that their products offer something extra if plugged into their equipment while maintaining interoperability. “This could be some simple AI-like features monitoring the link performance and warning customers of poor operation of devices on the other side of the link if these are made by another supplier,” says Kozlov.
LightCounting says that ZR+ units will be half to a third of the the number of 400ZR units shipped. However, each ZR+ module will command a higher selling price.
Regarding the ZR+ camps, one standardisation effort is OpenZR+ that adopts the forward-error correction (oFEC) scheme of the OpenROADM MSA, supports multiplexing of 100 Gigabit Ethernet (GbE) and 200GbE client signals, and different line rates – 100-400Gbps – to achieve greater reaches.
The backers of OpenZR+ include the two merchant DSP vendors, Acacia and NEL, as well as Fujitsu Optical Components, Lumentum, and Juniper Networks.
The second ZR+ camp includes four module-makers that are adopting the Canopus: Inphi, Molex Optoelectronics, NeoPhotonics and an unnamed fourth company. According to Schmitt, the unnamed module maker is II-VI. II-VI declined to comment when asked to confirm.
Schmitt argues that ZR+ should be interoperable, just like 400ZR. “I think NEL, Acacia, and Inphi should have an offsite and figure this out,” he says. “These three companies are in a position to nail down the specs and create a large, disruptive coherent pluggable market.”
Simon Stanley, founder and principal consultant at Earlswood Marketing Limited, expects several ZR+ solutions to emerge but that the industry will settle on a common approach. “You will initially see both ZR+ and OpenZR+,” says Stanley. “ZR+ will be specific to each operator but over time I expect OpenZR+ or something similar to become the standard solution.”
But the optical vendors stress the importance of offering differentiated designs to exploit the coherent DSP’s full potential. And maximising a module’s optical performance is something operators want.
“We are all for standards where it makes sense and where customers want it,” says Inphi’s Aiya. “But for customers that require the best performance, we are going to offer them an ecosystem around this DSP.”
“It is always a trade-off,” adds Ferris Lipscomb, vice president of marketing at NeoPhotonics. “More specialised designs that aren’t interoperable can squeeze more performance out; interoperable has to be the lowest common denominator.”
Next-generation merchant DSPs
The next stage in coherent merchant DSP development is to use a 5nm CMOS process, says Inphi. Such a state-of-the-art [CMOS] process will be needed to double capacity again while keeping the power consumption constant.
The optical performance of a 5nm coherent DSP in a pluggable will approach the high-end coherent designs. “It [the optical performance of the two categories] is converging,” says Aiya.
However, demand for such a device supporting 800 gigabits will take time to develop. Several years have passed for demand for 400-gigabit client-side optics to ramp and there will be a delay before telecom operators need 400-gigabit wavelengths in volume, says Inphi.
LightCounting points out that it will take Inphi and its ecosystem of suppliers at least a year to debug their products and demonstrate interoperability.
“And keep in mind that we are talking about the industry that is changing very slowly,” concludes Kozlov.
WaveLogic 5: Packing a suitcase of ideas in 7nm CMOS
-
Ciena’s WaveLogic 5 coherent digital signal processor family comprises the Extreme and Nano chips
-
The WaveLogic 5 Extreme maximises optical capacity and transmission reach while the WaveLogic 5 Nano is targeted at compact, power-conservative applications
Kim Roberts
Advancing coherent optical transmission performance; targeting the emerging coherent pluggable market; selling modules directly, and the importance of being more vertically integrated. All these aspects were outlined by Cisco to explain why it intends to buy the coherent optical transmission specialist, Acacia Communications; a deal that is set to be completed in the spring of 2020.
But such strategic thinking is being pursued by Ciena with its next-generation WaveLogic 5 family of coherent DSPs.
The WaveLogic 5 continues Ciena’s tradition of issuing a coherent digital signal processor (DSP) family approximately every three years: Ciena announced the WaveLogic 3 in 2012 and the WaveLogic Ai in 2016. (Add links).
The company has managed to maintain its three-yearly cadence despite the increasing sophistication of each generation of coherent DSP. For example, the WaveLogic 5 Extreme will support 800 gigabits-per-wavelength, double Ciena’s WaveLogic Ai that has been shipping for nearly two years.
Kim Roberts, vice president of WaveLogic science, says Ciena has managed to deliver its coherent DSPs in a timely manner since much of the algorithmic development work was done 5-6 years ago. The issue has been that certain features developed back then could not be included within the WaveLogic Ai.
WaveLogic 5 is implemented using a 7nm FinFET CMOS process whereas the WaveLogic Ai uses a 28nm specialist CMOS process known as fully-depleted silicon-on-insulator (FD-SOI).
“Seven-nanometer CMOS, due to its density and low heat, allows us to implement things that didn’t make the cut for the WaveLogic Ai,” says Roberts.
The company has a ‘suitcase of ideas’, he says, but not all of the concepts make it into any one generation of chip. “They have to justify performance versus schedule versus heat [generated],” says Roberts. “As we improve the technology, more features make the cut.”
And there are developments that will be included in future designs: “We keep refilling the suitcase,” says Roberts.
NAMING
Ciena first used the Extreme and Nano nomenclature with the WaveLogic 3. In contrast, the WaveLogic Ai, when launched in 2016, was a single-chip targeting the high-end. Ciena chose to change the naming scheme with the Ai since the chip signified a shift with features such as network monitoring.
However, Ciena highlights a key difference between the WaveLogic 3 and WaveLogic 5 families. The WaveLogic 3 Extreme and the WaveLogic 3 Nano could talk to each other on appropriate spans. In contrast, the two WaveLogic 5 chips are distinct. “They are not designed to interwork,” says Roberts.
NETWORKING TRENDS
Telecom service providers are investing in their networks to make them more adaptive. They want their networks to be scalable and programmable, says Ciena.
The operators also want to better understand what is happening in their networks and that requires collecting data, performing analytics and using software to configure their networks in an automated way.
“How do you get there? It is all about coherent technology,” says Helen Xenos, senior director, portfolio marketing at Ciena. “It is a critical element that is helping operators scale their networks.”
By enhancing the traffic-carrying capacity of fibre, coherent technology enables operators to reduce transport costs. “It allows them to be more competitive as they can do more with the hardware they deploy,” says Xenos.
Helen Xenos
Both telcos and cable operators are also applying coherent technology to new applications in their networks such as access.
These transport needs are causing a divergence in requirements.
One is to keep advancing optical performance in terms of the spectral efficiency and the traffic-carrying capacity of links. This is what the WaveLogic 5 Extreme tackles.
The second requirement - producing a compact coherent design for the network edge - is addressed by the WaveLogic 5 Nano.
For access designs, what is important is a compact design where the optics and the DSP can operate over an extended temperature range.
The Nano also addresses the hyperscalers’ need to connect their distributed data centres across a metro. “They need high capacity - 400 gigabits - and short-reach connectivity,” says Xenos. “It really needs to be the smallest footprint to maximise density.”
VERTICAL INTEGRATION
In addition to unveiling the WaveLogic 5 Extreme and Nano ICs, Ciena has outlined how it is more vertically integrated after investing in optics. In 2016, Ciena acquired the high-speed photonics division of Teraxion, gaining expertise in indium phosphide and silicon photonics expertise. {add link}.
Ciena is also now selling coherent optical modules. Gazettabyte revealed last year that Ciena was planning to sell modules using its own optics and WaveLogic technologies. {add link}
The company has no preference regarding indium phosphide and silicon photonics and uses what is best for a particular design.
“Silicon photonics buys you ease-of-manufacturing and cost; indium phosphide is what you need for 800 gigabits,” says Xenos.
Ciena stresses, however, that there is no simple formula as to when each is preferred. In terms of size and heat, silicon photonics has a strong advantage. “In terms of performance, you get better performance in some instances with indium phosphide and then there are overlaps because you bring in cost and other constraints,” says Roberts. “So there is no simple divide.”
“As we move forward, we are going to see an increasing percent of Ciena-custom components in WaveLogic coherent modems,” says Xenos.
Source: Gazettabyte
EXTREME
The WaveLogic 5 Extreme introduces several developments. It operates at specific baud rates ranging from 60 to 95 gigabaud. The baud rates are chosen so that both fixed-grid 100GHz channels and flexible grid ones are supported.
“For the best performance, you have flexible grid when 95 gigabaud is the primary baud rate,” says Roberts.
It is also Ciena’s first coherent DSP that uses probabilistic constellation shaping, a coding scheme used to achieve granular capacity increments. {add link}
“From 200 gigabits to 800 gigabits [in 25-gigabit increments], optimised over any path or the available margin,” says Roberts. “But what is unique about this is that it is optimised for non-linear propagation.”
Initially, the products using the WaveLogic 5 Extreme will use 50-gigabit increments. “This is what is required to service customers’ client requirements today: ten gigabits and multiples of 100-gigabit clients,” says Xenos.
>
“With 25-gigabit steps in client rate, the customer can choose to spend the margin on sending more bits”
The DSP uses four-wave frequency-division multiplexing to mitigate non-linear impairments, particularly beneficial for sub-sea systems.
Ciena says the four-wave frequency-division multiplexing is achieved electrically, reducing the optics to a minimum. “One laser and one modulator are used, so all the [cost-saving] economics of a single optical wavelength,” says Roberts. “But it has the non-linear performance of four tightly-coupled electrical systems.”
Ciena has also added an improved forward-error correction (FEC) scheme - a ‘throughput-optimised FEC’ - that uses variable overhead bits depending on the client rate.
“It will handle 8.6 percent errors compared to what we used in the WaveLogic Ai which handles 3.5 percent errors,” says Roberts. “So it is a decibel better.”
The Extreme chip also has improved link-monitoring capabilities. It monitors the signal-to-noise per channel as well as quantifies the non-linear contributions. “It helps people to understand what is happening in the network and create algorithms to optimise the capacity across the network,” says Xenos.
PROBABILISTIC CONSTELLATION SHAPING
Probabilistic shaping is used to improve the optical performance by lowering the signal energy by not using all the constellation points. Unless, that is, the full data rate is used and then all the constellation points are needed.
The degree of probabilistic shaping used is determined for each link. The parameters used to determine the probabilistic shaping are the amount of dispersion on the link, the span’s reach, and the transmitted client rate.
“The modem will measure what is going on in the link and the customer or some higher-level software will say what the client rate is,” says Roberts. “The modem will then figure out how to do the best non-linear probabilistic shaping to support that rate on the link.”
Roberts says other firms’ probabilistic shaping use one symbol at a time whereas Ciena use blocks, each comprising 128 symbols. “A bigger number would be better but I'm limited by my hardware,” says Roberts.
The 128 symbols equate to 1024 bits: four magnitude bits using 64-ary quadrature amplitude modulation (64-QAM) multiplied by two, one for each polarisation.
This means there are a total of 2^1024 combinations of 1024-bit sequences that could be sent. However, when sending a 400 Gigabit Ethernet (GbE) client signaland, for the benefit of explanation, assuming that 555 bits are needed to carry the data payload and the overhead, the number of possible bit sequences is trimmed to 2^555.
This is still a fantastically huge number but the DSP can work out which are the best 555-bit sequences to send based on them having the most tolerance to linear and non-linear interference.
“The ones that play nicely with their neighbours such that they cause the minimum non-linear degradation on the neighbouring wavelengths and on the other symbols,” explains Roberts.
Ciena is not forthcoming as to how it calculates the best sequences. “Ciena’s algorithms decide which ones are best,” says Xenos. “This is one of our key differentiators.”
The result is that, depending on the fibre type, a 1.5dB performance improvement is achieved for the non-linear characteristics.
“It allows more capacity to be chosen by the customer on that same link,” says Roberts. “With 25-gigabit steps in client rate, the customer can choose to spend the margin on sending more bits.”
Operating the Extreme at 95GBd, a reach of 4,000 km is possible at 400 gigabits and at 600 gigabits, the reach is 1,000 km (see table).
WAVELOGIC 5 NANO
The WaveLogic Nano supports 100-gigabit to 400-gigabit wavelengths and is aimed at applications that need compact designs that generate the least heat.
One application is to enable cable operators to move optics closer to the user and that must operate over an extended temperature range. Here, a packet platform is used that will support line interworking as equipment from different vendors may be at each end of the link.
Another requirement is operating over multiple spans in a metro. Here, compact equipment and low power are more important than spectral efficiency but it is still a challenging environment, says Ciena. Hundreds of nodes may be talking to each other and there may be cascaded reconfigurable optical add-drop multiplexers (ROADMs) with different fibre types making up the network.
A third application is single-span data centre interconnect where achieving the highest density on routers is key. This is the application the 400-gigabit, at least 80km 400ZR specification developed by the Open Internetworking Forum will address.
“The design that we are doing for the WaveLogic 5 Nano for 400ZR is to fit into a QSFP-DD,” says Xenos. “If there is a need for an OSFP [pluggable module], we will offer OSFP.”
Ciena also expects to offer a Nano-based CFP2-DCO module, which will outperform the ZR in terms of reach and features, for more demanding metro applications.
Another new segment requiring coherent optics is 4G and 5G access. “It is to be determined what type of platform is the winning solution in this environment,” says Xenos.
MAKING MODULES
Ciena first made its coherent DSP available to third parties in 2017 when it signed an agreement with Lumentum, NeoPhotonics and at the time Oclaro (since acquired by Lumentum) to use its WaveLogic Ai in their modules.
Now Ciena is selling directly the full coherent modem: the DSP and the optics. This is why Ciena created its Optical Microsystems unit in late 2017.
CMOS PROCESS
Moving to a 7nm FinFET CMOS process delivers several benefits.
It generates much lower heat than the WaveLogic Ai’s 28nm FD-SOI process. It also has a lower quiescent current, the current dissipated independent of whether the chip’s logic is active or not. And 7nm CMOS delivers much greater circuit density: the functionality that can be crammed into a square micrometre of silicon.
“So, a low power [consumption] on features you are not using, and we can include features that if you can't afford the heat, you can turn them off,” says Roberts.
It will offer its Nano in the form of pluggable modules, the WaveLogic Ai as a 5x7-inch module, and the WaveLogic 5 Extreme in another module form factor that will have its own interface. “These would all be viable optics,” says Xenos.
Availability
The first Wave Logic 5 Nano products will appear in the second half of this year while the first Extreme-based products will be available at the end of this year. The 400ZR coherent pluggable module is expected to be available in the first half of 2020.
Infinera’s ICE flow
Infinera’s newest Infinite Capacity Engine 5 (ICE5) doubles capacity to 2.4 terabits. The ICE, which comprises a coherent DSP and a photonic integrated circuit (PIC), is being demonstrated this week at the OFC show being held in San Diego.
Infinera has also detailed its ICE6, being developed in tandem with the ICE5. The two designs represent a fork in Infinera’s coherent engine roadmap in terms of the end markets they will address.
Geoff BennettThe ICE5 is targeted at data centre interconnect and applications where fibre in being added towards the network edge. The next-generation access network of cable operators is one such example. Another is mobile operators deploying fibre in preparation for 5G.
First platforms using the ICE5 will be unveiled later this year and will ship early next year.
Infinera’s ICE6 is set to appear two years after the ICE5. Like the ICE4, Infinera’s current Infinite Capacity Engine, the ICE6 will be used across all of Infinera’s product portfolio.
Meanwhile, the 1.2 terabit ICE4 will now be extended to work in the L-band of optical wavelengths alongside the existing C-band, effectively doubling a fibre’s capacity available for service providers.
Infinera’s decision to develop two generations of coherent designs follows the delay in bringing the ICE4 to market.
“The fundamental truth about the industry today is that coherent algorithms are really hard,” says Geoff Bennett, director, solutions and technology at Infinera.
By designing two generations in parallel, Infinera seeks to speed up the introduction of its coherent engines. “With ICE5 and ICE6, we have learnt our lesson,” says Bennett. “We recognise that there is an increased cadence demanded by certain parts of the industry, predominately the internet content providers.”
ICE5
The ICE5 uses a four-wavelength indium-phosphide PIC that, combined with the FlexCoherent DSP, supports a maximum symbol rate of 66Gbaud and a modulation rate of up to 64-ary quadrature amplitude modulation (64-QAM).
Infinera says that the FlexCoherent DSP used for ICE5 is a co-development but is not naming its partners.
Using 64-QAM and 66Gbaud enables 600-gigabit wavelengths for a total PIC capacity of 2.4 terabits. Each PIC is also ‘sliceable’, allowing each of the four wavelengths to be sent to a different location.
Infinera is not detailing the ICE5’s rates but says the design will support lower rates, as low as 200 gigabit-per-second (Gbps) or possibly 100Gbps per wavelength.
Bennett highlights 400Gbps as one speed of market interest. Infinera believes its ICE5 design will deliver 400 gigabits over 1,300km. The 600Gbps wavelength implemented using 64-QAM and 66Gbaud will have a relatively short reach of 200-250km.
“A six hundred gigabit wavelength is going to be very short haul but is ideal for data centre interconnect,” says Bennett, who points out that the extended reach of 400-gigabit wavelengths is attractive and will align with the market emergence of 400 Gigabit Ethernet client signals.
Probabilistic shaping squeezes the last bits of capacity-reach out of the spectrum
Hybrid Modulation
The 400-gigabit will be implemented using a hybrid modulation scheme. While Infinera is not detailing the particular scheme, Bennett cites several ways hybrid modulation can be implemented.
One hybrid modulation technique is to use a different modulation scheme on each of the two light polarisations as a way of offsetting non-linearities. The two modulation schemes can be repeatedly switched between the two polarisation arms. “It turns out that the non-linear penalty takes time to build up,” says Bennett.
Another approach is using blocks of symbols, varying the modulation used for each block. “The coherent receiver has to know how many symbols you are going to send with 64-QAM and how many with 32-QAM, for example,” he says
A third hybrid modulation approach is to use sub-carriers. In a traditional coherent system, a carrier is the output of the transmit laser. To generate sub-carriers, the coherent DSP’s digital-to-analogue converter (DAC) applies a signal to the modulator which causes the carrier to split into multiple sub-carriers.
To transmit at 32Gbaud, four sub-carriers can be used, each modulated at 8Gbaud, says Bennett. Nyquist shaping is used to pack the sub-carriers to ensure there is no spectral efficiency penalty.
“You now have four parallel streams and you can deal with them independently,” says Bennett, who points out that 8Gbaud turns out to be an optimal rate in terms of minimising non-linearities made up of cross-phase and self-phase modulation components.
Sub-carriers can be described as a hybrid modulation approach in that each sub-carrier can be operated at a different baud rate and use a different modulation scheme. This is how probabilistic constellation shaping - a technique that improves spectral efficiency and which allows the data rate used on a carrier to be fine-tuned - will be used with the ICE6, says Infinera.
For the ICE5, sub-carriers are not included. “For the applications we will be using ICE5 for, the sub-carrier technology is not as important,” says Bennett. “Where it is really important is in areas such as sub-sea.”
Silicon photonics has a lower carrier mobility. It is going to be harder and harder to build such parts of the optics in silicon.
Probabilistic constellation shaping
Infinera is not detailing the longer-term ICE6 beyond highlights two papers that were presented at the ECOC show last September that involved a working 100Gbaud sub-carrier-driven wavelength and probabilistic shaping applied to a 1024-QAM signal.
The 100Gbaud rate will enable higher capacity transponders while the use of probabilistic shaping will enable greater spectral efficiency. “Probabilistic shaping squeezes the last bits of capacity-reach out of the spectrum,” says Bennett.
“In ICE6 we will be doing different modulation on each sub-carrier,” says Bennett. “That will be part of probabilistic constellation shaping.” And assuming Infinera adheres to 8Gbaud sub-carriers, 16 will be used for a 100Gbaud symbol rate.
Infinera argues that the interface between the optics and the DSP becomes key at such high baud rates and it argues that its ability to develop both components will give it a system design advantage.
The company also argues that its use of indium phosphide for its PICs will be a crucial advantage at such high baud rates when compared to silicon photonics-based solutions. “Silicon photonics has a lower carrier mobility,” says Bennett. “It is going to be harder and harder to build such parts of the optics in silicon.”
ICE4 embraces the L-band
Infinera’s 1.2 terabit six-wavelength ICE4 was the first design to use Nyquist sub-carriers and SD-FEC gain sharing, part of what Infinera calls its advanced coherent toolkit.
At OFC, Infinera announced that the ICE4 will add the L-band in addition to the C-band. It also announced that the ICE4 has now been adopted across Infinera’s platform portfolio.
The first platforms to use the ICE4 were the Cloud Xpress 2, the compact modular platform used for data centre interconnect, and the XT-3300, a 1 rack-unit (1RU) modular platform targeted at long-haul applications.
A variant of the platform tailored for submarine applications, the XTS-3300, achieved a submarine reach of 10,500km in a trial last year. The modulation format used was 8-QAM coupled with SD-FEC gain-sharing and Nyquist sub-carriers. The resulting spectral efficiency achieved was 4.5bits/s/Hz. In comparison, standard 100-gigabit coherent transmission has a spectral efficiency of 2bits/s/Hz. The total capacity supported in the trial was 18.2 terabits.
Since then, the ICE4 has been added the various DTN-X chassis including the XT-3600 2.4 terabit 4RU platform.
Coherent gets a boost with probabilistic shaping
Nokia has detailed its next-generation PSE-3 digital signal processor (DSP) family for coherent optical transmission.
The PSE-3s is the industry’s first announced coherent DSP that supports probabilistic constellation shaping, claims Nokia.
Probabilistic shaping is the latest in a series of techniques adopted to improve coherent optical transmission performance. These techniques include higher-order modulation, soft-decision forward error correction (SD-FEC), multi-dimensional coding, Nyquist filtering and higher baud rates.
Kyle Hollasch
“There is an element here that the last big gains have now been had,” says Kyle Hollasch, director of product marketing for optical networks at Nokia.
Probabilistic shaping is a signal-processing technique that squeezes the last bit of capacity out of a fibre’s spectrum, approaching what is known as the non-linear Shannon Limit.
“We are not saying we absolutely hit the Shannon Limit but we are extremely close: tenths of a decibel whereas most modern systems are a couple of decibels away from the theoretical maximum,” says Hollasch.
Satisfying requirements
Optical transport equipment vendors are continually challenged to meet the requirements of the telcos and the webscale players.
One issue is meeting the continual growth in IP traffic: telcos are experiencing 25 percent yearly traffic growth whereas for the webscale players it is 60 percent. Vendors must also ensure that their equipment keeps reducing the cost of transport when measured as the cost-per-bit.
Operators also want to automate their networks. Technologies such as flexible-grid, reconfigurable optical add/drop multiplexers (ROADMs), higher-order modulation and higher baud rates all add flexibility to the optical layer but at the expense of complexity.
There is an element here that the last big gains have now been had
“It is easy to say software-defined networking will hide all that complexity,” says Hollasch. “But hardware has an important role: to keep delivering capacity gains but also make the network simpler.”
Satisfying these demands is what Nokia set out to achieve when designing the PSE-3s.
Capacity and cost
Like the current PSE-2 coherent DSPs that Nokia launched in 2016, two chips make up the PSE-3 family: the super coherent PSE-3s and the low-power compact PSE-3c.
The PSE-3s is a 1.2-terabit chip that can drive two sets of optics, each capable of transmitting 100 to 600 gigabit wavelengths. This compares to the 500-gigabit PSE-2s that can drive two wavelengths, each up to 250Gbps.
The low-power PSE-3c also can transmit more traffic, 100 and 200-gigabit wavelengths, twice the capacity of the 100-gigabit PSE-2c.
Nokia has used a software model of two operators’ networks, one an North America and another in Germany, to assess the PSE-3s.
The PSE-3s’ probabilistic shaping delivers 70% more capacity while using a third fewer line cards when compared with existing commercial systems based on 100Gbps for long haul and 200Gbps for the metro. When the PSE-3s is compared with existing Nokia PSE-2s-based platforms on the same networks, a 25 percent capacity gain is achieved using a quarter fewer line cards.
Hollasch says that the capacity gain is 1.7x and not greater because 100-gigabit coherent technology used for long haul is already spectrally efficient. “But it is less so for shorter distances and you do get more capacity gains in the metro,” says Hollasch.
Probabilistic shaping
The 16nm CMOS PSE-3s supports a symbol rate of up to 67Gbaud. This compares to the 28nm CMOS PSE-2s that uses two symbol rates: 33Gbaud and 45Gbaud.
The PSE-3s’ higher baud rate results in a dense wavelength-division multiplexing (DWDM) channel width of 75GHz. Traditional fixed-grid channels are 50GHz wide. With 75GHz-wide channels, 64 lightpaths can fit within the C-band.
The PSE-3s uses one modulation format only: probabilistic shaping 64-ary quadrature amplitude modulation (PS-64QAM). This compares with the PSE-2s that supports six modulations ranging from binary phase-shift keying (BPSK) for the longest spans to 64-QAM for a 400-gigabit wavelength.
Using probabilistic shaping, one modulation format supports data rates from 200 to 600Gbps. For 100Gbps, the PSE-3s uses a lower baud rate in order to fit existing 50GHz-wide channels.
In current optical networks, all the constellation points of the various modulation formats are used with equal probability. BPSK has two constellation points while 64-QAM has 64. Probabilistic shaping does not give equal weighting to all the constellation points. Instead, it favours those with lower energy, represented by those points closer to the origin in a constellation graph. The only time all the constellation points are used is at the maximum data rate - 600Gbps for the PSE-3s.
Using the inner, lower energy constellation points more frequently than the outer points reduces the overall average energy and this improves the signal-to-noise ratio. That is because the symbol error rate at the receiver is dominated by the distance between neighbouring points on the constellation. Reducing the average energy still keeps the distance between the points the same, but since a constant signal power level is used for DWDM transmission, applying gain increases the distance between the constellation points.
“We separate these points further in space - the Euclidean distance between them,” says Hollasch. “That is where the shaping gain comes from.”
Changing the probabilistic shaping in response to feedback from the chip, from the network, we think that is a powerful innovation
Using probabilistic shaping delivers a maximum 1.53dB of improvement in a linear transmission channel. In practice, Nokia says it achieves 1dB. “One dB does not sound a lot but I call it the ultimate dB, the last dB in addition to all the other techniques,” he says.
By using few and fewer of the constellation points, or favouring those points closer to the origin, reduces the data that can be transported. This is how the data rate is reduced from the maximum 600Gbps to 200Gbps.
To implement probabilistic shaping, Nokia has developed an IP block for the chip called the distribution matcher. The matcher maps the input data stream as rates as high as 1.2 terabits-per-second onto the constellation points in a non-uniform way.
Theoretically, probabilistic shaping allows any chosen data rate to be used. But what dictates the actual data rate gradations is the granularity of the client signals. The Optical Internetworking Forum’s Flex Ethernet (FlexE) standard defines 25-gigabit increments and that will be the size of the line-side data rate increments.
Embracing a single modulation format and a 75GHz channel results in network operation benefits, says Hollasch: “It stops you having to worry and manage a complicated spectrum across a broad network.” And it also offers the prospect of network optimisation. “Changing the probabilistic shaping in response to feedback from the chip, from the network, we think that is a powerful innovation,” says Hollasch.
The reach performance of the PSE-3s using 62Gbaud and PS-64QAM. The reach performance of the PSE-2s is shown (where relevant) for comparison purposes.
Product plans
The first Nokia product to use the PSE-3 chips is the 1830 Photonic Service Interconnect-Modular, a 1 rack-unit compact modular platform favoured by the webscale players.
Nokia has designed two module-types or ‘sleds’ for the 1830 PSI-M pizza box. The first is a 400-gigabit sled that uses two sets of optics and two PSE-3c chips along with four 100-gigabit client-side interfaces. Four such 400-gigabit sleds fit within the platform to deliver a total of 1.6 terabits of line-side capacity.
In contrast, two double-width sleds fit within the platform using the PSE-3s. Each sled has one PSE-3 chip and two sets of optics, each capable of up to a 600-gigabit wavelength, and a dozen 100-gigabit interfaces. Here the line-side capacity is 2.4 terabits.
Nokia says the 400-gigabit sleds will be available in the first half of this year whereas the 1.2 terabit sleds will start shipping at the year-end or early 2019. The first samples of the PSE-3s are expected in the second half of 2018. Nokia will then migrate the PSE-3s to the rest of its optical transport platform portfolio.
So has coherent largely run its course?
“In terms of a major innovation in signal processing, probabilistic shaping is completing the coherent picture,” says Hollasch. There will be future coherent DSP chips based on more advanced process nodes than 16nm with symbol rates approaching 100GBaud. Higher data rates per wavelength will result but at the expense of a wider channel width. But once probabilistic shaping is deployed, further spectral efficiencies will be limited.