NeoPhotonics : Toward One Terabit per Second On A Single Wavelength

By Ferris Lipscomb, Ph.D. on February 16, 2019 | Leave a Comment

The next milestone in the quest to pack as much data into a fiber as possible will be 1 Terabit transmission on a single DWDM wavelength. But, you might say, 1.2T Transponders are now beginning to be shipped commercially, so isn't a mere 1T a step backwards? However, the current generation of 1.2T transceivers utilize two separate DWDM wavelengths each running at 600G to achieve a total of 1.2 terabits/sec capacity. This means that while one transponder can support the full 1.2T, it uses two wavelength channels in the fiber, and this means two lasers, two modulators and two receivers. So from a cost perspective, it would be better to transmit 1.2T on a single wavelength rather than two. Since coherent transmission was introduced a decade ago, two factors have allowed the continuing increase in the data rate of a single wavelength. First, Coherent allows the use of higher order modulation to increase the number of bits that are encoded in single 'on-off' cycle or symbol rate, also called the baud rate. For example, non-coherent on-off keying systems have one bit per baud, while initial coherent systems had 4 bits per baud (including the use of two polarizations). Another example is to 64QAM, which gives 6 bits per symbol. The second way to increase the data rate is to increase the baud rate itself. Initial coherent systems utilized 32 Gbaud, but the current 600G/lambda run at 64 Gbaud. If we combine the benefits of both high baud rate and high order modulation, such as 64Gbaud and 64QAM in dual polarizations, then you will note that 64 Gbaud times 6 bits per symbol times 2 polarizations gives 768 Gbps, well the above 600Gbps touted. Coherent systems require some amount of overhead for forward error correction (FEC), so you don't get the full amount of bandwidth for usable data.

So what will it take to get to 1T per wavelength? Either the baud rate or the constellation can be increased. There are advantages and disadvantages to each. Increasing the constellation tends to decrease the distance that can be transmitted, because the distance between the symbols becomes smaller. The advantage of increasing constellation is to have a higher spectral efficiency. For example, increasing to 256 QAM while maintaining the baud rate at 64 Gbaud would yield 1.02T of total bandwidth, but the net data rate would be reduced by overhead required for FEC. The other approach is to increase the baud rate itself. Increasing from 64 Gbaud to 100 Gbaud would give 1.2T of total capacity, to be reduced by FEC overhead. Depending on the actual system requirements, some combination of the two is likely to be used. For example, 90 Gbaud and 128 QAM gives 1.26T total capacity, which can support 1T per wavelength with 20% FEC overhead.

Increasing either of these two parameters requires more computation by the DSP to extract the data, and hence leads to higher power consumption. However, increasing the baud rate also requires increasing the speed of the Analog to Digital converter front end as well as the speed of the DSP itself. These higher speed electronics are now being delivered to us by the inexorable progress of semiconductor chips. As the semiconductor industry moves to more advanced process nodes with smaller features the resultant chips become faster and lower power. The first generation of coherent systems used DSPs based on a 35 nm process, which supported 100G based on 32 Gbaud and QPSK . The current 600G per wavelength systems are based on 16 nm process nodes and utilize 64 Gbaud and 64 QAM. And now DSPs based on the 7 nm node are becoming available. These have gotten most attention due to their low power which will enable compact pluggable modules like OSFP and QSFP-DD. It is believed that yet another generation of 7nm DSP will support 100+ Gbaud and high order modulation for metro and long distances.

With the inevitability of the DSP, although not without extensive effort, the only question about achieving 1T per wavelength is the availability of the optical components that can work with the DSP. Faster modulators and receivers will be required to modulate and demodulate the analog signals for the DSP to convert back to digital form. Fortunately, good progress has been made on the optics as well. Initial systems were shipped with modulators and receivers that supported 32 Gbaud. These were termed Class 20 due to the fact that the 3 dB analog bandwidth was 20GHz. In general the possible digital transmission rate is higher than the analog bandwidth because the roll-off is gradual and some response exists at the higher frequencies. The current 600G per wavelength systems use Class 40 receivers and modulators which support 64 Gbaud. And now, Class 50 receivers and modulators are being introduced, which have 3 dB bandwidths of around 50 GHz and which can support 90-100 Gbaud.

One other element is critical to achieving 1T per wavelength. As mentioned above, when higher order modulation is used, the symbol spots come closer and closer together. If they overlap then the DSP cannot distinguish one from the other. One of the reasons for the spots smearing is phase noise in the lasers used. Therefore, it is critical to use lasers with very low phase noise, or linewidth, when using 64 QAM or 128 QAM. Fortunately, these lasers also exist and are in mass production. The typical linewidth of an external cavity laser is well below 100 kHz and can minimize the OSNR penalty with higher QAM.

With these developments, I expect we will begin to see systems with 1 T per wavelength appear in the 2021 time frame.