PAM4 vs NRZ: Growing Irrelevance of Standards Bodies

At the OFC 2021 Rump Session, there will be two sides debating the following propositions: “Did the Optics Industry Blunder by Switching Intra-Datacenter Links from NRZ to PAM4? Will More DSP like PAM6 and Coherent Follow, or Will WDM and Parallel Save the Day?” There are certainly opportunities for both teams to be correct. We tend to be at least somewhat sympathetic to the NRZ side, partially because of the assertion that “analog direct detection technology went extinct” is not paying attention to the 5G fronthaul optics space, nor the large amount of very low-cost PSM4 and CWDM4 transceivers at the 100G rate continuing to be shipped. Even if the PAM proponents happen to be the big winners in the future, such a statement ignores the historically slow rate of obsolescence of any technology in the telecom/datacom space. We also would argue that the “once-in-a-generation transition,” as stated by Marvell Technology Group/Inphi, when the pending acquisition was announced, came about unnaturally based on the IEEE Ethernet Group’s premature push for 400GbE, particularly by Cisco Systems, a supplier, which has even, at least in the past, pressed for PAM16. Additionally, other than for the vendor itself, it was hardly in the interest of the marketplace for Inphi to gain such a headlock position in the PAM4 space, offering extremely complicated and costly optical PHYs. The most critical point is that such poor judgement by the IEEE group would only be another reason for large operators around the world to increasingly bypass standards bodies and adopt their own proprietary solutions – and our expectations for further consolidation of public network operators out of necessity (such as the recently announced proposed buyout of Shaw Communications by Rogers Communications) will allow more of them individually to make or break a solution.

Of course, customers, especially the hyperscalers, tend to care less about the type of solution, only that it is cheap and uses the least amount of power – although in answering one of the inquiries from the “DSP team,” the second characteristic is obviously more important than the first to the bean counters at these entities as power consumption is their number-one concern. Moreover, we believe that it is hasty to call PAM4 a revolution yet, in that the bulk of components manufactured will likely remain at 50G wavelengths for quite a while. For example, all of the webscalers in China will transition to 4x50G PAM4 200GbE QSFP56 modules for their next upgrade cycle because they: 1) will last the longest in the network; 2) support multimode fiber well; 3) can leverage the most mature 100G-SR4 and CWDM4 manufacturing ecosystem across the industry; 4) are much easier to knock down the price; and most importantly, 5) permit the large 25.6T radix switch ASIC. (Only with 51.2T switching ASICs will 400G become mandatory.)

Incredibly, the IEEE group managed to skip over an entire generation of gear. The organization also committed a major error in not picking up 25G before it became pervasive, and not permitting the telcos to play its usual role of paying for the low-volume phase as optical plants are able to operate four or more times faster using the same fiber and amplifiers. This oversight has been compounded by accepting multiple technologies, effectively mismanaging the standardization process.

Such market segmentation gets in the way of economies of scale as well as allows for a new solution to drive down the price of an older one. The hyperscalers are adept enough at getting costs to the point of destroying the datacom optics ecosystem. They do not need any additional help.

The fact that the IEEE is being driven more by data center and systems vendors and less by optics component suppliers is the downfall for low-cost, less problematic optical solutions compared to what is being designed in China. As a single player, Huawei Technologies, solely drives the network architectures for the entire nation with no evident favoritism to either electronics or optics. Chinese suppliers should continue to excel at avoiding the least amount of complexity, such as in providing 200GbE transceivers in volume that come close to the prices of 100GbE, as well as being expected to remain lean and mean in dominating the 400GbE space with the simplest possible optical modules.

Regarding wireless, China Mobile, which has been the biggest buyer of 5G fronthaul optics in the world, decided to totally ignore the ITU-T standards and use a CWDM4 variant at the O-band. Even though China had taken control of the organization from Europe, the former either decided to skip a more formal step and/or wished to establish a worldwide competitive advantage sooner. To the extent that Huawei is allowed to sell outside of China, these countries are also getting new NRZ-type devices as well. (Development work is being done by European suppliers on extending the length of these 25G types of gear at the optical C-band, which would make it more accommodating to mobile network operators in other parts of the world.)

Before even enabling mmWave in any meaningful way, China Telecom and Unicom decided to create a common 5G network. The handwriting is on the wall that a global shakeout of wireless operators will gradually occur as there will ultimately be no other way of getting an adequate return on investment. Regrettably, in the very long-term, there will likely be a single common carrier in each country with a lot of purchasing power.

Concerning nine other points on the upcoming NRZ/PAM4 contest:

1. Our expectation of the sustainability of pluggables at least until 200T switches is an endorsement of the idea that only so much DSP can be thrown at the problem. (Certainly, there will be a need for new DSPs, but predicting the timing is very difficult.)

2. Coherent will eventually creep into everything, but it will seem like an eternity (a relatively safe bet is that coherent-lite will be the big winner in the foreseeable future).

3. VCSELs using NRZ modulation over parallel multimode fiber will replace copper at some point in the future (do not hold your breath on copper being eliminated, as indicated at a previous Rump Session).

4. PAM4 is already used in parallel fiber, and it is hard to envision placing additional fibers. (So, PSM4 is here to stay for structured cables.)

5. It remains to be seen whether using lots of 50G NRZ lanes with WDM will be embraced by the Ethernet community (perhaps the machine-learning folks will do so).

6. Apparently, Intel had introduced a CWDM8, which was initially canceled, but it is being proposed again.

7. On the optics side, 100G PAM4 lanes are already shipping (400G-DR4, -FR4) – although clearly not conforming to 802.3db.

8. 50G PAM4 may hit a dead end on the switch electrical channel side at 51.2T.

9. Those advocates that maintain that ICs are cheaper than optics will point out that the very high gross margins will not be sustainable with expected competition.

The upcoming session is the classic direct detect versus DSP argument — PAM using DSP to reconstitute a signal is every bit a coherent receiver technology implementation and the issue is the cost for a processor to deal with phase and amplitude versus a simple on-off design, as involves higher baud rates and parallel signals for those data rates. There are engineers who believe that NRZ will not come back unless the industry goes backwards on the interconnect speeds. The contention is that one cannot get a laser to turn on and off quickly enough, along with the receiver being able to detect the changes of state to be effective.

Network planners involved in both the electrical and optic channels for switch system design will be the most likely to currently find themselves kind of in the middle of this debate. While PAM is preferred because the PCB loss and connector discontinuity at higher Nyquist frequency is desirable for the optic link, NRZ or low-order PAM4 are considered good choices. In the future for higher speed links, such as 224G lambda, there is a compelling reason to use PAM6 or PAM8 for the electrical channel (from switch chip to optic module) and use low-order PAM4 for the optic link, with the necessity of a gearbox. (The higher-order optical DSP is still much more expensive and power-hungry.)

Unlike the anti-NRZ extremists, fibeReality is hardly stating that standards bodies will cease to exist. Too many careers are on the line. Furthermore, there will continue to be a requirement to establish common technical goals that will be in the interests of all players.