Our company is always inclined to look towards historical precedent for guidance, and electronics have continually been leveraged throughout the years to bail out optics, such as with its imperfections including nonlinearities, as well as with its cost in enabling greater simplification. So, our initial response would be to look in favor towards a solution like DSP-based PAM -4 for that relatively small segment of the data center market, which is gradually moving to higher data rates using single-mode fiber. Nevertheless, although these optics are being advertised to reach as long as two kilometers, it appears that technical restrictions could confine it to only a quarter of that length.
The biggest advantage of using the sophisticated electronics in PAM-4 is that it allows for fewer lasers. Consequently, it facilitates the use of a less complicated optical multiplexer.
PAM-4 suppliers, such as Inphi, are stating that they are looking forward to the opportunity to provide 4x25G QSFP modules. The server to ToR switch application on the access side is viewed as the sweet spot for that gear.
Despite the apparent optimism by these vendors concerning the new modulation tool, there is another long-established truth from the past based in immutable physics. A severe Signal-to-Noise Ratio (SNR) impairs optics. The higher order modulation scheme of PAM-4 by itself results in a 4.8dB inherent penalty in its link budget.
As a result, the only time the utilization of PAM-4 would appear to be logical would be in cases in which the distances are shorter. Apps in the neighborhood of 500 meters would allow engineers to have SNR to burn. Yet, this limitation on not being able to be deployed in a ubiquitous manner could exclude it totally from consideration in favor of other solutions – especially, if one recognizes the SNR problem hampering increasing margins even as rates get higher.
There is also the matter of the apprehension around the fact that PAM-4 has never been accomplished in a production sense. Even if it works, there is the question of whether it will do so in a robust fashion – it needs to be determined whether users will have a satisfactory experience or some periodic issues.
It is also not inconceivable that the industry could develop a low-cost 100G per lambda in a serial interface before PAM-4 is ready for volume shipments. Perhaps this new technique of using multiple signaling levels will ultimately find a place in markets outside of optics.
Lastly, we believe that Finisar has made a compelling case that if possible, NRZ is the optimal way to go, for single-mode fiber client interfaces based on margin potential, etc.
[written by Mark Lutkowitz]
(To read about Finisar's outspokenness about silicon photonics, please click here.)