The growing bandwidth demands in high-performance computing and datacenter networks are currently addressed by Moore’s law scaling of CMOS-based electronic switch chips, where each generation of merchant silicon, fabricated using the latest CMOS node, provides an increase in bandwidth and/or port count over the previous generation. As fabrication constraints predict the end of Moore’s law, switch chip manufacturers are already looking toward power-hungry multichip implementations to continue bandwidth expansion [1]. In contrast, optical switching technologies enable the reconfiguration of high-bandwidth WDM data streams at a much better power efficiency than electrical switches. They do this by diminishing the need for inefficient O/E/O conversions, ser/des, and transmission lines at the switch. Furthermore, they enable bandwidth scaling through increasing channel counts in WDM systems in addition to standard evolution of line rates. 3D-MEMS-based switches [2] are highly scalable and therefore well suited for routing very large bandwidth signals. However, their slow reconfiguration time is not sufficient for latency-bound applications that require reconfiguration of the optical network at the packet granularity [3]. Silicon photonics is a promising technology for realizing low-power and nanosecond-scale switch fabrics using fast electro-optic (EO) phase shifters. However, to date, the number of input and output ports has been limited to $\le 16$ [4,5] due to both high insertion loss and high optical crosstalk. While there are methods to mitigate these effects (e.g., with the integration of optical gain [3] or the use of dilated topologies [6]), optical crosstalk remains an obstacle that ultimately limits the scalability of the technology [7,8]. In order to increase the number of ports, the design of the $1\times 2$, $2\times 1$, or $2\times 2$ elementary switch unit is critical. Mach–Zehnder-based and ring-resonator-based EO switches have typically been reported with optical crosstalk values in the range of $-15$ to $-20\mathrm{dB}$. In [9], we demonstrated a measured loss and crosstalk of 1.2 dB and $-27\mathrm{dB}$, respectively, using a Mach–Zehnder switch (MZS) driven in push–pull mode. In [10], the authors demonstrate a two-stage (dilated) switch used as a $2\times 2$ unit, improving the crosstalk to $-31\mathrm{dB}$, which is the lowest reported crosstalk for a $2\times 2$ EO silicon photonic switch that we are aware of. Unfortunately, the switch had an insertion loss of 9 dB. This approach also increases the power consumption, and the footprint of the switch unit since four MZSs are required to perform the function of one. In this Letter, we present a novel single-stage $2\times 2$ nested-MZS (NMZS) design that can, theoretically, completely suppress the optical crosstalk by balancing the power inside the MZ interferometer using a variable optical attenuator. We demonstrate the design by fabricating a switch in a 90 nm photonics-enabled CMOS process, complete with device drivers and CMOS interface logic. We show a measured record crosstalk value as low as $-34.5\mathrm{dB}$ with an insertion loss of only 2 dB, while also verifying the nanosecond-scale switching time of the device.

The free-carrier dispersion (FCD) effect in silicon is commonly used for the design of low-power, compact, and fast optical phase shifters. For optical switch applications, the structure of the phase shifter is usually a p-i-n diode operated in current injection mode. While the efficiency is very good, the carrier injection also produces some inherent insertion loss due to the free-carrier absorption (FCA). Figure 1(a) presents a schematic of a $2\times 2$ MZS using one FCD-based phase shifter in one arm. The MZS is operated by turning the phase shifter on and off, enabling light entering one port to be routed either to a through port ($\varphi =\pi$, on-state) or to a crossover port ($\varphi =0$, off-state). In this simple design, the on-state is problematic because the carrier injection that generates the phase shift creates FCA and a power imbalance appears inside the interferometer, leading to light leakage into the cross port and also creating insertion loss. While this effect can be mitigated using a phase shifter in each arm and driving the switch push–pull [9], there is always a fundamental limit on the best crosstalk achievable. In this Letter, we propose a novel switch design that can completely suppress the optical crosstalk without any insertion loss penalty. Figure 1(b) presents the design. It consists of a $2\times 2$ MZS in which the FCD-based straight-line phase shifter is replaced by a FCD-based MZ phase shifter. It also has a (possibly tunable) loss section in the other arm. The MZ phase shifter could use a $1\times 1$ structure but having a $2\times 2$ structure allows one to integrate a photodetector for monitoring the power. In order to generate the required $\pi$ phase change and actuate the switch, the MZ phase shifter is symmetrically driven around maximum transmission in push–pull mode by using a pair of EO phase shifters that swing from 0 to $\pi$. This operation mode is impervious to drive voltage imperfections because it produces an exact digital $\pi$ phase shift at the output of the MZ phase shifter. More importantly here, it also delivers an equivalent insertion loss for the two states. The major advantage of using a MZ phase shifter is that having constant insertion loss, the power can be exactly balanced inside the interferometer by attenuating in the lower arm, which eventually completely suppresses the crosstalk of the device. It should be noted that the MZ phase shifter design requires a twofold EO power increase compared with the simple straight-line phase shifter. However, the power consumption of these devices is generally very low, less than $2\mathrm{mW}/\pi$. Figure 2 shows calculations of the insertion loss and crosstalk of a NMZS as a function of the attenuation in the lower arm of the interferometer. As a reference, we also present calculations in the case of a MZS driven single-ended (dashed green lines). In the simulations, we considered a 250 μm long phase shifter with a ridge waveguide structure that has an optical confinement factor of 0.7 at the input wavelength of 1310 nm. More details on the simulation framework can be found in [9]. When there is no attenuation, the insertion loss is around 0.3 dB and the crosstalk is around $-28\mathrm{dB}$. These starting values are better than in the MZS case because the FCA-induced insertion loss of the MZ phase shifter is half of the straight-line phase shifter. As seen in Fig. 2, when tuning the loss section, the crosstalk is effectively decreasing until a point, around 0.7 dB attenuation in our design, where the power balance inside the interferometer is perfect, leading to infinitely small crosstalk. After that point, the insertion loss of the lower arm is larger than the MZ phase shifter loss, and the crosstalk increases again. The insertion loss of the device linearly increases with the lower arm attenuation and, as expected, is exactly equal to the MZS case at the optimum attenuation point.

 

Fig. 1. Switch designs. (a) Schematic of a $2\times 2$ MZS driven single-ended (se) with one EO phase shifter in one arm. (b) Schematic of a $2\times 2$ NMZS having one push–pull-driven MZ-based EO phase shifter in one arm, a loss section in the other arm, and a power monitor port with a photodetector (PD).

Download Full Size | PDF

 

Fig. 2. Insertion loss (top) and crosstalk (bottom) of a NMZS (solid blue lines) when balancing the power inside the interferometer using the loss section from the lower arm. Dashed green lines are the insertion loss and crosstalk for the MZS (se) in the on-state. The calculations assume optical couplers with no insertion losses and a perfect 3 dB splitting ratio. The input wavelength is 1310 nm, and the optical confinement factor in the phase shifters is 0.7.

Download Full Size | PDF

We designed and fabricated the photonic switch in IBM’s CMOS integrated nanophotonics technology CMOS9WG [11]. Figure 3(a) presents a schematic of the NMZS photonic components. The NMZS has in the upper arm a MZ phase shifter that uses a pair of 250 μm long forward-biased p-i-n diodes and, in the lower arm, a 250 μm long variable optical attenuator (VOA), also implemented as a carrier injection p-i-n diode. The optical couplers are directional couplers designed for 1310 nm. In the present design, the NMZS has a $\sim 50\mathrm{\mu m}$ extra path length which limits the optical bandwidth. If using a length-balanced MZ, just as any MZS, the optical bandwidth of the NMZS primarily depends on the bandwidth of the couplers. The switch also integrates thin silicide thermo-optical (TO) tuners to trim the phase in the MZ phase shifter and in the NMZS. The thermal heaters, the p-i-n diode phase shifters, as well as the waveguide and optical coupler dimensions are the same as those reported in [9]. The switch is monolithically integrated with a digital inverter-based CMOS driver. The driver employs a five-stage digital buffer with a fanout of two. Each stage is comprised of two inverters with the same transistor dimensions. The input to the driver is accessed through a serial-to-parallel interface that consists of a 28-bit shift register. The chip receives the serial electrical inputs (DATA, CLOCK, and ENABLE) which control the switch driver. Figure 3(b) presents a schematic diagram of the CMOS logic and driver monolithically integrated with the photonic switch. More details on the CMOS driver and logic can be found in [12,13]. Figures 4(a) and 4(b) show die photographs of the test site. For testing, the chips were mounted on a custom printed circuit board (PCB) for edge-coupled optical access [13]. The serial electrical inputs, the logic and driver supplies, and the heater biases were all wire bonded to the PCB. The supply voltage for the output stage ($V_{\mathrm{OS}}$) of the switch driver was $\sim 1.1\mathrm{V}$ and the diode current was $\sim 1.5\mathrm{mA}$, leading to an EO efficiency of $\sim 1.65\mathrm{mW}/\pi$. The main source of electrical power consumption originates from the silicide heaters which had a TO efficiency of $\sim 25\mathrm{mW}/\pi$ but could be reduced by using other types of heating structures [14].

 

Fig. 3. Design and fabrication of the monolithic switch in the CMOS9WG process. (a) NMZS schematic showing the interferometer structure, the EO phase shifters (PS1 and PS2), the VOA, and the TO phase trimmers (H1 and H2). (b) Schematic diagram of the CMOS logic and switch driver connected to the photonic switch.

Download Full Size | PDF

 

Fig. 4. Die photographs. (a) NMZS showing details on the EO phase shifters, the VOA, and the heaters. (b) Test site showing an array of NMZS devices monolithically integrated with the CMOS electronics.

Download Full Size | PDF

Figure 5 presents static characterizations of the switch. For all of the optical measurements, the setup included one 1310 nm tunable laser, one polarization controller, the device under test (DUT), and an optical power meter. The DUT was accessed via cleaved fibers, and the gap between the chip facets and fibers was filled with index-matched fluid. Figure 5(a) shows the measured optical crosstalk of the NMZS as a function of the VOA bias for each input/output (IO) configuration in the on- and off-state of the switch at a wavelength of 1319.05 nm. For each configuration, the transmission was normalized to the transmission of a straight silicon waveguide. In the figure, $T_{ij}$ refers to the normalized transmission from port $i$ to port $j$, and the port number designation is the same as presented in Fig. 1(b). The VOA current ($I_{\text{voa}}$) varies from 0 to 0.85 mA, which corresponds to an optical attenuation of 0 to $\sim 0.5\mathrm{dB}$ as measured separately on a stand-alone VOA. For each $I_{\text{voa}}$, the phase needed to be readjusted with the heaters in order to maintain the correct bias point of the NMZS. When increasing $I_{\text{voa}}$, the power balance inside the NMZS improves which leads to a strong decrease of the optical crosstalk. Comparing $T_{11}$, $T_{12}$, $T_{21}$, and $T_{22}$, we see that the lowest crosstalk achievable for each transmission appears at different VOA biases. The optimum $I_{\text{voa}}$ is 0.4 mA for $T_{12}$, 0.6 mA for $T_{11}$, 0.8 mA for $T_{22}$, and $>0.8\mathrm{mA}$ for $T_{21}$. These slightly different operating points downgrade the best optical crosstalk achievable, and the VOA setting should be selected in order to optimize the port isolation for each IO configuration of the switch. The best setting for using the switch as a $2\times 2$ is a VOA current of 0.6 mA which provides an optical crosstalk smaller than $-34.5\mathrm{dB}$ and an insertion loss of $\sim 2\mathrm{dB}$, as seen in Fig. 5(b) spectra. If using a VOA current of 0.45 mA, the device could also be used as a $1\times 2$ switch ($T_{11}$ and $T_{12}$ on/off lines) with a crosstalk of $-37\mathrm{dB}$. The cause of these different operating points is attributed to a slight variation of the power coupling coefficient of the directional couplers which creates some asymmetry in the MZ that cannot be completely corrected using our design. In the future, we could use more process-tolerant optical coupler designs [15,16] in order to improve the 3 dB power split ratio uniformity. Figure 5(b) presents the transmission spectra of the switch at a VOA bias of 0.6 mA. The optical bandwidth is here limited by the MZ wavelength dependence due to the extra path length.

 

Fig. 5. Static characterizations of the NMZS. (a) Optical crosstalk for all configurations of the $2\times 2$ switch as a function of the VOA bias at $\lambda =1319.05\mathrm{nm}$. (b) Optical spectra measured with a tunable laser. The scan resolution is 0.05 nm in the main plot and 0.005 nm in the inset.

Download Full Size | PDF

We also measured the switching transient of the NMZS. To measure the switching time, the CMOS serial-to-parallel interface was connected to a data generator producing a squared signal with a 130 ns period. Figure 6 presents the transient results. In these measurements, we excited port 1 of the NMZS using a tunable laser at 1318.8 nm and measured the optical response from both outputs using a photodetector connected to a sampling scope. As seen in the Fig. 6 inset, the 10%–90% transient time is smaller than 4 ns. Finally, the total electrical power of the switch is $<34\mathrm{mW}$ which would result in 0.34 pJ/bit added to the link when both ports are populated with 50 Gbps signals.

 

Fig. 6. Transient measurements of the NMZS showing the normalized optical response at both outputs of the switch. The input wavelength is 1318.8 nm.

Download Full Size | PDF

In conclusion, we demonstrated a novel single-stage Mach–Zehnder switch unit that can dramatically decrease the optical crosstalk ($<-34\mathrm{dB}$) while maintaining very reasonable insertion losses ($\sim 2\mathrm{dB}$). By improving the uniformity of the power splitting ratio of the optical couplers, and/or using more process-tolerant coupler designs, we believe the optical crosstalk could be further reduced. This achievement can bring about a major breakthrough in the scalability of integrated nanosecond-scale silicon photonic switch fabrics for energy efficient computing networks.