32 × 32 silicon electro-optic switch with built-in monitors and balanced-status units

To construct large-scale silicon electro-optical switches for optical interconnections, we developed a method using a limited number of power monitors inserted at certain positions to detect and determine the optimum operating points of all switch units to eliminate non-uniform effects arising from fabrication errors. We also introduced an optical phase bias to one phase-shifter arm of a Mach–Zehnder interferometer (MZI)-type switch unit to balance the two operation statuses of a silicon electro-optical switch during push–pull operation. With these methods, a 32 × 32 MZI-based silicon electro-optical switch was successfully fabricated with 180-nm complementary metal–oxide–semiconductor (CMOS) process technology, which is the largest scale silicon electro-optical switch to the best of our knowledge. At a wavelength of 1520 nm, the on-chip insertion losses were 12.9 to 16.5 dB, and the crosstalk ranged from −17.9 to −24.8 dB when all units were set to the ‘Cross’ status. The losses were 14.4 to 18.5 dB, and the crosstalk ranged from −15.1 to −19.0 dB when all units were in the ‘Bar’ status. The total power consumptions of the 32 × 32 switch were 247.4 and 542.3 mW when all units were set to the ‘Cross’ and ‘Bar’ statuses, respectively.

of power monitors to every switch unit 12,19 . We also propose a method for setting the optical phase bias in the MZI phase-shift arm to eliminate the difference between the switch's 'Cross' and 'Bar' statuses with push-pull operation control and no extra heat power added for phase tuning. With these methods, we experimentally demonstrated a 32 × 32 MZI-based silicon EO switch fabricated with 180-nm CMOS process technology. To the best of our knowledge, this is the largest scale silicon EO switch worldwide. At a wavelength of 1520 nm, the on-chip insertion losses were from 12.9 to 16.5 dB, and the crosstalk ranged from −17.9 to −24.8 dB when all units were in the 'Cross' status with a power consumption of 247.4 mW. Moreover, the on-chip insertion losses were from 14.4 to 18.5 dB, and the crosstalk ranged from −15.1 to −19.0 dB when all units were in the 'Bar' status with a power consumption of 542.3 mW.

Results
Monitoring switch of a Benes network. A Benes network is an iterative network based on 2 × 2 and 4 × 4 basic units, which are helpful for finding a common rule for inserting limited numbers of monitors at defined positions. A crosstalk analysis of a Benes switch can be easily understood starting with 2 × 2 and 4 × 4 switches. Ignoring the propagating loss of the switch units, the crosstalk of a 2 × 2 switch can be defined as X = (P in − P out )/ P in = 1 − T, where the transmittance T is the ratio of the output power P out of the selected port to the input power P in , as shown in Fig. 1(a). The operation point of a 2 × 2 switch unit can be determined and adjusted by detecting its minimal crosstalk power, which is more sensitive to the operation status than the transmission.
The model of light propagation for a 4 × 4 switch can be defined as shown in Fig. 1(b), where S ij stands for the switch unit in stage i and line j, and the input port of a switch is labelled with a binary number, as described elsewhere 20 . X ij and T ij are the crosstalk and transmission of unit S ij , respectively. The crosstalk and insertion loss of waveguide crossings were ignored for a simpler analysis 21,22 . Assuming that all of the switch units are working in the 'Cross' status and light is input from left-side port 01 with a power of P 0 , the output power of each switch unit is listed in Table 1.
In Table 1, the crosstalk X 11 of S 11 can be detected by measuring the outputs of the 1 st stage, P 0 X 11 , and 2 nd stage, P 0 X 11 T 22 . These outputs have a simple proportional relationship with the crosstalk X 11 without being affected by the switch's extinction, which allows for the minimum value of X 11 to be measured easily. The power monitor could be a contactless optical probe or tuneable coupler 23,24 , which has less influence on the switch's performance. As shown in Fig. 1(b), the crosstalk of S 11 and S 21 could be obtained by detecting the output of the 2 nd stage, P 0 X 11 T 22 and P 0 T 11 X 21 , by monitors M2 and M1 when light is input from left port 01. However, the 3 rd stage output is the interference result of the main power and crosstalk power. The crosstalk is difficult to detect by measuring this output, even when the maximum output power in Table 1 is considered. By changing the input port to 10, the crosstalk of S 12 and S 22 can also be detected with monitors M1 and M2. The units in the 3 rd stage can be directly detected by the same monitors, M2 and M1, with the light input from the right-side port of the switch. The crosstalk of all switch units can also be measured when all units are set to the 'Bar' status, as summarised in 0 (00) P 0 T 11 P 0 T 11 X 21 P 0 T 11 X 21 X 31 + P 0 X 11 X 22 T 31 P 0 (01) P 0 X 11 P 0 T 11 T 21 P 0 T 11 X 21 T 31 + P 0 X 11 X 22 X 31 0 (10) 0 P 0 X 11 X 22 P 0 T 11 T 21 X 32 + P 0 X 11 T 22 T 32 0 (11) 0 P 0 X 11 T 22 P 0 T 11 T 21 T 32 + P 0 X 11 T 22 X 32    Fig. 2(a), the 8 × 8 switch has 20 switch units and four 10% bi-directional power taps used as monitors. The bi-directional coupler in Fig. 2(b) consists of a directional coupler and two grating fibre couplers and detects the light in the waveguide from both the left and right sides. The bi-directional coupler had a 250-nm-wide gap between a straight waveguide and a bent waveguide with a 20-μm bend radius. The grating coupler was designed with a 610-nm period and 60% duty, which has a −5 dB coupling efficiency at 1550 nm with a 10° fibre angle. The switch is built up by multimode interferometers (MMIs) and waveguide crossings. The dimensions and test result are shown in Fig. 2(c). The insertion loss of the MMI is from 0.12 to 0.56 dB from 1500 to 1570 nm, which was obtained by a linear fitting of the measured losses of the cascade MMIs. The crosstalk of the waveguide crossing is less than −30 dB, and the insertion loss is from 0.07 to 0.15 dB from 1515 to 1570 nm. Figure 2(d) shows an 8 × 8 Benes switch consisting of two 4 × 4 Benes switches embedded with power monitors in complementary positions. In a Benes network, the transmission light power and crosstalk power of the 1 st stage unit propagate to the upper and lower 4 × 4 switches, respectively. Taking the output of the 1 st stage in the 8 × 8 switch as the 4 × 4 switch input, the first two stages of the 4 × 4 switch units can be detected with the methods described before. The crosstalk light in the 1 st stage passes through the first two stages of a 4 × 4 switch and can also be detected by the monitors. For example, by setting all of the switch units to the 'Cross' status with light input from left-side port 000, the crosstalk light of S 11 passing through the path S 11 → S 21 → S 32 → M2 is detected by monitor M2, as indicated by the red line in Fig. 2(d). The units on the right side of the monitors can be detected with the incident light input from the right-side ports. For example, for incident light from right-side port 001' , the crosstalk power of S 51 propagates through the route S 51 → S 43 → M4, as indicated by the blue line. Similarly, all other units can be detected by four monitors and adjusted to their optimum operating points according to the method in Table 3.  N × N switch evolution. The applicability of the monitoring method can be extended with induction to a 2 n × 2 n Benes network switch as a universal method, which is shown in Fig. 3(a). For a 2 n × 2 n network, the monitor positions in the upper 2 n−1 × 2 n−1 Benes network are complementary to the lower 2 n−1 × 2 n−1 Benes network. Since this method is applicable to 4 × 4 and 8 × 8 Benes switches, we may assume that the method is also applicable to a 2 n−1 × 2 n−1 Benes switch. Light passes through a switch unit in the input stage and is input into two 2 n−1 × 2 n−1 subnetworks; the inputs of the 2 n−1 × 2 n−1 networks contain the status information of the unit in the input stage. Since the monitor positions in these two 2 n−1 × 2 n−1 networks are complementary to each other, the crosstalk power of the unit in the 1 st stage can always be monitored by the monitor in the upper or lower 2 n−1 × 2 n−1 networks. All units in the input and output stages of the 2 n × 2 n network can be detected by the monitors in two 2 n−1 × 2 n−1 networks. Thus, we can conclude that the monitoring method is applicable to a 2 n × 2 n Benes network switch.
Optical Phase Bias. Usually, an MZI switch has equally long phase-shift arms. It needs a phase shift of π to switch between the 'Cross' and 'Bar' statuses. The switch we designed has two 200-μm-long PIN phase-shift arms with the cross section shown in Fig. 4(a). Phase shifting in a EO switch uses the plasma dispersion effect caused by the injected carriers, which also causes optical absorption. The optical absorption loss introduces an imbalance in the two arms and increases crosstalk. Here, we propose a method that sets a phase bias of π/2 in one MZI arm and uses push-pull operation to decrease the phase shift needed for switching by changing the waveguide structure. A lower phase shift results in less optical loss and less crosstalk. As shown in Fig. 4(a), the phase bias is provided by decreasing the length of the lower output taper in the left MMI and filling with a straight waveguide called an L-shifter. The optimised bias length of the L-shifter of 0.9 μm was obtained by measuring the ratio of the two outputs, as shown in Fig. 4(b). A comparison of a phase-bias switch and an origin unit without a bias is shown in Fig. 4(c). The difference in the crosstalk between the 'Cross' and 'Bar' was improved from 18.1 to 3.6 dB. The power consumption was also reduced from 6.24 to 1.9 mW in the 'Bar' status. The switching time of the switch unit was measured with the falling and rising times of 1.0 and 1.2 ns, respectively, as shown in Fig. 4(d).

× switch experiment.
With the methods proposed above, the 32 × 32 silicon EO switch with dimensions of 12.1 mm × 5.2 mm in Fig. 5(a) was fabricated. It has 144 phase-bias 2 × 2 MZI-switch units and 288 electrode pads; 16 bi-directional power taps were implemented in the switch to detect the operation statuses of all 144 units, as shown in Fig. 3(b). First, the operation points for all units in the 'Cross' and 'Bar' statuses were investigated with the 16 power taps. The total power consumption of the 32 × 32 switch was measured to be 247. 4

Discussion
We have demonstrated a method using limited numbers of power monitors inserted at certain positions to detect and determine the optimum operating points of all switch units in a large-scale silicon optical switch. This method more effectively realises reductions in both the chip size and measurement cost compared to the addition of power monitors to each switch unit. We also introduced an optical phase bias to one arm of an MZI to balance the performance of two operation statuses of a silicon EO switch by push-pull operation. With these methods, a 32 × 32 MZI-based EO switch was successfully demonstrated on a 180-nm CMOS process platform, which is the largest scale silicon EO switch to the best of our knowledge. The total power consumptions of the 32 × 32 switch were 247.4 and 542.3 mW when all units were in the 'Cross' and 'Bar' statuses, respectively. At a wavelength of 1520 nm, the on-chip insertion losses were 12.9 to 16.5 dB, and the crosstalk ranged from −17.  would be needed to build larger silicon EO switches for reducing the insertion loss of the switch. The polarization dependence of the silicon optical switch using a submicron waveguide is also a significant issue for use in practical optical communication. Polarization control devices 25 and polarization-independent silicon waveguides 26 could be used to solve this problem.

Methods
Fabrication. All of the silicon optical switches mentioned in this paper were fabricated with the 180-nm CMOS process line of Semiconductor Manufacturing International Corporation (SMIC) on the basis of a 340 nm silicon-on-insulator (SOI) platform. The connecting rib waveguide has a width of 450 nm with an 80-nm-high slab, which has an average propagation loss of 0.53 dB/mm at 1550 nm wavelength. All of the passive components such as the MMIs, waveguide crossings, and grating couplers were designed for the TE mode. The P++ and N++ doping densities are approximately 2 × 10 20 cm −3 . The metal wires are 0.45-μm-thick aluminium wires. Trenches with an area of 240 μm × 4.5 μm were inserted between adjacent PIN phase shifters to avoid thermal and electrical crosstalk.
Measurement. The crosstalk of one port of an optical switch is defined as the ratio between the output power from the target input port and that from all other ports when light was launched into all input ports. The on-chip insertion loss excluded the grating coupling loss spectrum. The switch chip mounted on the heat sink was connected to the test board using wire bonding. A tuneable laser (SANTEC TSL-510), an optical power meter (YOKOGAWA AQ2200), an optical spectrum analyser (YOKOGAWA AQ6370C), a high-speed sampling oscilloscope (Tektronix DSA 8300), a pulse pattern generator (Agilent 81130 A), and direct-current power supplies (Agilent E3631A and B2901A) were used to test the static and dynamic characters of the switches.