Skip to main content

The Path to 100G Single Lambda in the Data Center - Part 3

Published Date

Maury Wood, AFL Test & Inspection Product Line Manager

Part 3 (part 1 - part 2 - part 4)

Part three of this blog series is on the technical and commercial dynamics that will lead the semiconductor and electro-optical components industries to enable ubiquitous 100G single lambda system designs over the next several years.

Just like in radio frequency spectrum (one of many areas where RF and optical communications share useful characteristics), and to put into very simple terms: the higher the channel bandwidth, the lower the channel range/reach. And the notion of spectral efficiency is equally salient for both RF spectrum and fiber "spectrum." For example, a key parameter of different multimode fiber types is effective modal bandwidth, and the units of this parameter tell us what we need to know–MHz*km, in other words, in optical communications, the product of bandwidth and reach is fixed (at a given optical carrier wavelength, such as 850 nm or 1310 nm). It’s worth a moment's appreciation for the fact that optical carrier frequencies range in the hundreds of TeraHertz, for example 850 nm is 353 THz.

So, the key takeaway is that if we want to maintain the same reach (say 100 m for "200GBASE-SR4," not-as-yet an IEEE standard or "400GBASE-SR4," also not-as-yet an IEEE standard), then we'll need to use the same effective bandwidth or symbol rate (aka baud rate). This gives a path forward, and the industry is moving ahead with a tremendous level of consensus to utilize PAM-4 modulation (2 bits per symbol) with a 25/28 GBaud symbol rate and 50/56 Gbps bit rate. More to discuss here later, but we can say with certainty that short reach 50G single lambda will use PAM-4 (Four Level Pulse Amplitude Modulation) optical modulation, and this same multi-level modulation technology is likely to be utilized for 100G single lambda systems.

Back to the signal chain. We've started on the transmit (Tx) side, as signal processing engineers know that receivers are always tougher; it’s easier to talk than to listen I guess. So for 50G single lambda, we can envision Optical Internetworking Forum CEI-56G-VSR “CAUI-2” compliant Tx SerDes channels coming out of a theoretical next generation Broadcom switch ASIC, traversing perhaps 10 cm of 75 ohm differentially balanced RF microstrip/stripline, possibly using a Teflon-based PCB substrate material into a pluggable transceiver SerDes, with sufficient channel equalization to receive the encoded bit stream with low (perhaps 10e-14) bit error rate. There may be algorithmic Forward Error Correction on this electrical link to provide coding gain and more bit error resiliency.

Once on board the theoretical 200GBASE-SR4 transceiver module, the electrical signals go into a clock/data recovery retimer chip. If the electrical lane configuration doesn't match the optical lane configuration, then a so-called gearbox function is needed, adding cost, power and area to the design, so matched electrical and optical rates are heavily preferred in data center and other cost-pressured transceiver designs.

Since this conceptual next generation transceiver is using 25 Gbaud PAM-4 to achieve single lambda 50G, we need a "mixed-signal Digital Signal Processor" (a DSP with on-chip high speed data converters) to create the four-level PAM-4 modulation analog waveform.

As mentioned earlier, short reach 25G lambda systems universally use Non-Return to Zero (NRZ) On-Off Keying (OOK) intensity modulation in which the supply current to the VCSEL (Vertical Cavity Surface-Emitting Laser) diode is switched above the lasing threshold to create the digital ones (light emitted) and zeros (essentially no light emitted). Note that LED lasers have a finite “extinction ratio” between on and off, but still plenty of Signal-to-Noise Ratio (SNR) to achieve robust 25 Gbps short range links. This on/off bit sequence is a continuous amplitude, continuous time signal, and for our purposes can be considered a binary signal when sampled by the receiver correctly. On the other hand, the PAM-4 signal, with four valid signal/switching levels and three eye openings per symbol, can only be treated (from a Tx synthesized signal and Rx sampled signal perspective) as a continuous amplitude, continuous time "analog" or "linear" signal by both the transmitter and the receiver.

PAM-4 Signal

PAM-4 Signal with Three Eye Openings per Symbol Period (image courtesy NeoPhotonics Corp)

This is important, among other reasons, because in current NRZ/OOK 25G lambda system, the electrical/optical (EO) interface between the VCSEL driver function (downstream of the Tx Clock Data Recovery function) and the VCSEL light source is conventionally defined as "limiting" (discrete levels), whereas the optical modulator driver function in the 50G lambda system is defined as "linear" (continuous levels), and as noted above, a DSP with a simple but very fast two bit Digital-to-Analog Converter is required to synthesize the PAM-4 modulation linear (i.e., analog) signals.

Stay tuned for part four of this series for a discussion on Mach-Zehnder optical modulators.