800G Optical Communications: Market Trends and Specifications Roundup
Added 2023-09-26: An Introduction to Co-Packaged Optics (CPO)
For more industry insights, follow me on Instagram: @simple_tech_trend
This article consolidates the references below into a status report on 800G optical communications market trends and specifications.
Contents
You may also be interested in my other technical articles:
800G Market and Application Trends
The spread of cloud computing, AR/VR, AI and 5G is making data center demand ever stronger. The chart shows that interconnect bandwidth demand grew by about 48% from 2017 to 2021. Estimates at the time had 400G optical modules entering use in 2020, with 800G adoption starting around 2022-2023. As for the long-discussed co-packaged technology, Vladimir Kozlov, CEO of LightCounting Market Research, believes traditional optical modules will remain mainstream in the 800G era, while co-packaged technology will just begin to be deployed.
Samples of 800G QSFP-DD/OSFP modules began appearing in 2021. On the SerDes side, 100G SerDes is being introduced in 2021 and will become mainstream from 2023. 200G SerDes has a higher and more complex technical barrier, so 100G SerDes may be around for a relatively long time.
LightCounting released its latest forecast in July 2021. The top 5 companies, Alibaba, Amazon, Facebook, Google and Microsoft, spent US$1.4 billion on Ethernet transceivers in 2020, and their spending will rise to more than US$3 billion by 2026.
By the end of 2026, 800G transceivers will dominate the entire DC market.
Google plans to start deploying 1.6T modules in 4-5 years; these are not included in this forecast.
The forecast includes co-packaged optics, which will start replacing pluggable optics in data centers in 2024-2026.
Several factors led to the higher forecast:
An improved outlook for AI-driven data traffic growth, based on the latest data Google shared at OFC 2021.
Good R&D progress by 800G optical module suppliers.
Higher-than-expected bandwidth demand from data center clusters.
Ethernet switch ICs will also release 25.6 Tb/s prototypes in 2021-2022.
In data center applications, 400G and 100G will be replaced by 800G and 200G respectively, while the optical module form factor will remain QSFP-DD/OSFP.
AI cluster data centers have fewer architectural tiers, and the existing 200G/400G will be replaced by 400G/800G.
8x100G Short-Reach Application Scenarios
Short Reach (SR) scenarios typically require lengths of 1~100m, but at 800G bandwidth, multimode technology may only support a maximum transmission distance of 30~50m.
The main limitations are the bandwidth of the VCSEL itself and the dispersion of multimode fiber. Reaching 100m with this architecture may be possible by increasing DSP performance, but this can lead to the drawbacks of:
High cost
High latency
High power consumption
For this reason, the 800G MSA recommends that the 800G-SR8 specification use Parallel Single Mode (PSM) technology, which was previously used only for 500m and above.
Considering cost and power, single-mode technology should be the best solution for SR short-reach transmission. For the laser on the Tx side, DML, EML and silicon photonics are all possible options.
At present, the more mature EML and SiPh technologies show better BER performance. Although the DML option has poorer BER, it may still be compensated by the DSP's powerful equalization, though possibly at the cost of higher power consumption and cost.
8x100G Active Component Progress
Lumentum's 100G PAM4 DML has been verified over a 2km transmission distance, maintaining an extinction ratio (ER) above 3.0 dB from 25 to 85 degC.
AOI verified that its DML achieves TDECQ = 2.54 dB after a 5-tap feed forward equalizer (FFE) filter.
Huawei has already demonstrated 100G/lane eye diagrams at the module level, with TDECQ reaching 3.0 dB.
4x200G Long-Reach Application Scenarios
In FR scenarios, 200G/lane PAM4 signaling will be the most important technology, and it can also be applied to future 1.6T interconnect platforms. To meet this specification, using more advanced, smaller-node processes to achieve lower power will be key.
The CWDM4 power budget needs to account for:
Link Insertion loss
Multipath interference (MPI)
Differential group delay (DGD)
Transmitter dispersion penalty (TDP)
According to the IEEE reference model, MPI and DGD are about 0.4 dB. For TDP, given the doubled data rate, a reasonable value should be around 3.9 dB. Taking all of this into account, the 800G MSA puts Rx sensitivity at around -5 dBm. To meet this specification, stronger FEC must be introduced.
800G FR4 mainly uses CWDM4 technology. One option is 2xCWDM4, which requires eight lasers and PDs and therefore carries a relatively high cost. A single CWDM4 uses fewer active components and could lower the price, but each channel must then support 200G, doubling the bandwidth requirement. Coherent technology can of course also cover 2km applications, but in my view both its power and its price are too high to be practical.
Choosing a 4x200G Modulation Format
To increase bandwidth, PAM4 was successfully introduced at 400G; it delivers twice the data throughput under the same bandwidth constraint. Higher-order modulation such as PAM6 is also being considered, but the technology is not yet mature, and PAM4 looks set to remain mainstream.
For the 800G FR4 specification, CWDM4 4x200G should be the most cost-effective option. As for progress on 112 Gbaud active components:
Sumitomo's PD and EML S21 can both meet the 112 Gbaud bandwidth requirement.
50 GHz RF-ICs have also received initial validation.
So as the technology evolves, 200G/lane transmission is achievable.
Through mathematical budgeting with a target BER of 2E-3, the 800G MSA concludes that the 800G TDECQ specification should be defined as 3.9 dB (3.4 dB for 400G).
TDECQ calculations must incorporate FFE with a higher tap count.
In addition, because there is more noise, it is recommended that Tx optical power be raised by 0.5 dB to compensate for the DGD penalty.
Rx sensitivity benefits from the improved FEC, so for now the recommendation is to leave it unchanged.
Evaluating 4x200G Packaging Approaches
To achieve higher bandwidth (56 GHz), packaging approaches Solution A and Solution B have been proposed.
Solution A is the conventional packaging approach, while Solution B packages the driver using flip chip.
Solution B theoretically has a better S21 curve, while Solution A leaves room for further tuning of the driver-side matching. Both packaging approaches can therefore potentially reach 56 GHz bandwidth.
4x200G FEC Options
To successfully correct a 2E-3 bit error rate, FEC technology must also advance. There are currently two options:
Option 1: use a new FEC technology on the optical module side
Option 2: introduce an entirely new FEC technology across the overall system
Possible 800G DR Application Scenarios
Putting the SR and FR scenarios together, the possible options for DR (500 m) applications are as follows.
PSM8, 100G/lane: extending the 100m approach, reaching 500m is possible.
2xCWDM4, 100G/lane
PSM4, 200G/lane
CWDM4, 200G/lane: considering future 1.6T applications, technology maturity and the ratio of components to fibers, this is the 800G MSA's preferred option, though power and cost still need work.
Simply put, the higher the data-rate requirement, the more likely it is that products built to long-reach specifications can also cover shorter-reach specifications.
IEEE Status on 200G~1.6T DAC and Optical Module Specifications
The 50 Gbps/lane specifications are essentially complete, while 100 Gbps/lane is still being defined. Although Ethernet 1.6 Tb/s is still some way from deployment, IEEE has already begun evaluating the relevant specifications.
Blue text = standardized
Red text = specification in progress
* Note: subject to change as of publication
Possible 800G Optical Module Options
The table below lists possible 800G optical module options based on the 800G MSA and IEEE specifications. They are largely similar to 400G; the biggest difference is that whether single-mode or multimode technology will be used for links under 100m remains an open question.
QSFP-DD800 Form Factor Updates
Higher power requires a better heat dissipation path, so QSFP-DD800 adds a Type 2B module, the main difference being a longer heat sink.
Optical Connector Interface Updates
With so many different packaging technologies, optical connector interfaces have also become increasingly diverse.
Comments