1. Introduction

High-bandwidth, low-power optical interconnects, integrated in close proximity to digital logic and memory are increasingly being considered a viable interconnect strategy for enabling further scaling of the processing power of massively parallel computing systems. Short-range optical-interconnect approaches based on silicon photonic components have drawn considerable interest from the research community over the past ten years, owing in part to the promise of seamless integration with CMOS circuits [1–4]. On the inter- and intra-chip interconnect levels, optical network-on-chip (ONoC) architectures capable of providing reconfigurable communication paths between the processor cores and memory systems on a chip multiprocessor have been proposed [5,6].

A key device for an ONoC based on wavelength-insensitive optical routing scheme, is a broadband and noise-tolerant optical switch, capable of routing an optical data stream encoded in multiple parallel wavelength channels [7–10]. In [11], we proposed and analyzed an ultra-broadband digital switch design, based on a multi-stage Mach-Zehnder lattice (MZL), with strongly improved noise tolerance as compared to a single-stage Mach-Zehnder (MZ) switch. Such an improved noise tolerance not only helps to enable robust switching operation in an on-chip environment, which is known to be prone to substantial supply-voltage noise [12], but it also helps to improve the thermal stability of the ‘on’-state of the device, cancelling out any fluctuations of the injected-carrier density induced by temperature variations in the forward-biased p-i-n diode phase shifters [11].

In this paper, we report an experimental demonstration of the digital MZL-switch concept proposed in [11]. Robust crosstalk levels of lower than −15 dB are demonstrated in an apodized four-stage MZL switch, for drive-voltage noise levels as high as 300 mV_pp. In contrast, a single-stage MZ switch is shown to have a drive-voltage noise tolerance of only 45 mV_pp. Also, greater ‘on’-state extinction levels (below −26 dB) are demonstrated as well as lower free-carrier-induced ‘on’-state insertion loss (less than 0.45 dB). As the improved switching performance comes at the expense of increased power consumption (compared to a MZ switch), we will briefly discuss the outlook for low-power operation of these digital switches.

2. Design and fabrication of a Mach-Zehnder lattice switch in SOI

The design of a broadband digital electro-optic switch based on a Mach-Zehnder lattice was extensively discussed in our earlier paper [11]. The MZL switch consists of a series of balanced interferometric stages, in which each lower branch of the stages includes a forward-biased p-i-n diode phase shifter, as illustrated in Fig. 1(a) for the specific case of a four-stage lattice. In such multi-stage structures, a digital switching response originates from properly engineered multiple-path interference at the output ports, and can be realized by optimizing the coupling coefficients κ_i of the individual coupling sections. An optimized MZL switch design is expected to result in lower ‘on’-state crosstalk, lower ‘on’-state insertion loss, as well as a drastically improved tolerance to thermal and drive-voltage noise, as compared to a conventional single-stage MZ switch [11].

 

Fig. 1 (a) Schematic representation of the four-stage MZL switch, featuring Hamming-apodized coupling coefficients (κ₁ = 3.2%, κ₂ = 13.1% and κ₃ = 20.5%). (b) Microscope image of the fabricated device, which occupies a total footprint of 160 μm x 75 μm. The directional-coupler lengths are respectively L₁ = 4.1 μm, L₂ = 8.5 μm, and L₃ = 10.8 μm. Each of the four p-i-n diode phase shifters in the lower active branch is 50-μm long. The top branch contains shorted dummy p-i-n diodes. The nodes connected to the signal and ground electrodes are illustrated by dashed lines and black dots. The scale bar is 30-μm long.

Download Full Size | PDF

Here, we report the switching characteristics of a fabricated four-stage MZL switch, shown in Fig. 1(b). The switch is fabricated on a silicon-on-insulator (SOI) wafer using a subset of processing modules from a standard IBM front-end CMOS process flow [7]. The cross-sectional dimensions of the silicon waveguides are 500 × 220 nm². Each of the four p-i-n diode phase shifters in the bottom (active) branch of the stages are 50-μm long. The top branch contains shorted dummy p-i-n diodes, which are included to ensure balance of optical losses between branches. The coupling coefficients κ_i are apodized according to a Hamming window function (κ₁ = 3.2%, κ₂ = 13.1% and κ₃ = 20.5%), resulting in respective directional-coupler lengths of L₁ = 4.1 μm, L₂ = 8.5 μm, and L₃ = 10.8 μm. The device occupies a total footprint of 160 μm x 75 μm.

The expected switching characteristics of the fabricated MZL switch are obtained by simulation [11,13] as a function of the carrier density N injected into the p-i-n diode phase shifters, and are shown in Fig. 2(a) . The switching response of this four-port structure is given by the four transmittance curves T_ij(λ) = |S_ij(λ)|², where S_ij(λ) are the complex transfer functions of the optical field from input port a_i to output port b_j, with i, j = 1, 2. The Hamming apodization of the coupling coefficients results in a switching response where any sidelobes in the T₁₂ transmittance are suppressed to well below −20 dB. As a result, if the switch is biased to an ‘on’-state level sufficiently above N = 2.8 × 10¹⁸ cm⁻³ (Δn_Si ~7.3 × 10⁻³), it is expected to have a substantial tolerance against noise on the injected-carrier density. In addition, at a larger bias of N = 5.6 × 10¹⁸ cm⁻³ (Δn_Si ~1.34 × 10⁻²), T₁₂ extinction levels as large as −30 dB can be obtained. The insertion loss of the T₁₁ transmission path is predicted to remain under 0.4 dB throughout the full operating range of the switch despite the substantial optical absorption loss (on the order of Δα_Si ~350 dB/cm) induced by the free carriers injected into the waveguide core. However, the T₂₂ transmission path suffers from increasing insertion loss with carrier density, as the free-carrier absorption (FCA) is highest for this path. As such, the T₂₂ transmission path is not as useful in practical applications, and the MZL switch is preferably used as a 1 × 2 switch.

 

Fig. 2 (a) Calculated switching response of the four-stage MZL switch as shown in Fig. 1(a), as a function of injected-carrier density. The T₁₂ switching sidelobes are suppressed to less than −20 dB. (b) Calculated switching response of the reference MZ switch.

Download Full Size | PDF

Figure 2(a) illustrates the principal of digital performance demonstrated by the MZL design. When switched sufficiently deep into the ‘on’-state, the T₁₂ (T₁₁) transmittance will robustly remain low (high), even when appreciable variations in the injected free carrier density (or equivalently the p-i-n diode drive voltage) are introduced.

For comparison, the simulated switching response of a balanced, single-stage MZ switch with a one 200-μm-long p-i-n diode phase shifter and two 50% directional couplers is shown in Fig. 2(b). In contrast to the digital behavior of the MZL, the dual-path interference in the MZ switch type results in an analog switching response, with only limited tolerance against noise in the ‘on’-state, which is reached at N = 1.4 × 10¹⁸ cm⁻³ (Δn_Si ~4 × 10⁻³). In addition, the ‘on’-state T₁₂ extinction is limited to −21 dB, while the ‘on’-state T₁₁ insertion loss is as high as 0.8 dB, as 50% of the input light is sent through the lossy free-carrier-based phase-shifter (Δα_Si ~90 dB/cm).

Finally, as both the MZL not the MZ switch designs are built from balanced interferometric stages without any built-in phase delay (in the ‘off’ state), they can be considered optically temperature insensitive, under the assumption that any temperature difference between the upper and lower branches in each stage is insignificant. Temperature variations within the footprint of the device (160 μm x 75 μm for the MZL) would obviously still affect the obtained crosstalk and extinction ratio, as they would induce a non-intentional phase delay. However, the impact of these effects can be minimized through layout optimization and further reduction of the MZL device footprint.

In conclusion, when considering optical switches appropriate for on-chip CMOS-integrated photonic networks, in which supply-voltage noise and thermal variations are common [11,12], the multi-stage MZL switch has several advantages over the traditional MZ design.

3. Measured switching characteristics

3.1 Optical bandwidth

The optical bandwidth of the fabricated MZL switch is characterized by coupling TE-polarized light from a broadband LED source to one of the input ports of the switch and analyzing the intensity spectrum of the transmitted light at both output ports. This analysis is performed first without an applied bias voltage, in order to obtain the transmission spectra T₁₁(λ) and T₁₂(λ) for the ‘off’-state, shown in Fig. 3(a) . The intensity spectra measured at the respective output ports are normalized against the sum of the intensity spectra of both output ports, with the input signal at the same input port. As a result, the obtained T₁₁(λ) and T₁₂(λ) spectra do not include ‘off’-state optical losses of a passive nature, such as those due to optical scattering or bending. The additional passive insertion loss of the MZL switch is separately measured to be 1.25 ± 0.25 dB at 1479 nm. As seen in Fig. 3(a), the obtained ‘off’-state crosstalk is lower than −15 dB over an optical bandwidth Δλ_BW of 40 nm centered on a wavelength of 1479 nm. The poor ‘off’-state extinction of T₁₁ is the result of a small fabrication-induced optical path mismatch between the nominally identical phase-delay sections, which arises from variations in waveguide width and etch depth.

 

Fig. 3 (a) T₁₁ and T₁₂ transmittance spectra of the four-stage MZL switch, measured for a continuous-wave applied voltage V_D = 0 V (‘off’) and V_D = 1.3 V (‘on’). Crosstalk levels lower than −17 dB are obtained in the ‘off’-state, whereas better than −26 dB crosstalk levels are obtained in the ‘on’-state (at 1478 nm). (b) Transmittance spectra of the reference MZ switch. Crosstalk levels of about −20 dB are obtained for both the ‘off’-state (V_D = 0 V), and the ‘on’-state (V_D = 1.1 V), each at 1530 nm.

Download Full Size | PDF

Subsequently, the same measurements are performed while applying a continuous-wave voltage V_D that maximizes the T₁₂ extinction at 1479 nm (V_D = 1.3 V). The resulting normalized ‘on’-state T₁₁(λ) and T₁₂(λ) spectra include only the insertion losses from FCA loss. These spectra are also shown in Fig. 3(a). The best-case ‘on’-state crosstalk is lower than −26 dB at 1479 nm and lower than −20 dB over an optical bandwidth of more than 80 nm around a wavelength of 1479 nm. The excess ‘on’-state T₁₁ insertion loss is approximately 0.45 ± 0.2 dB for the MZL switch, which agrees very well with the simulations presented in section 2.

For comparison, similar measurements are performed for the MZ switch, the results being shown in Fig. 3(b). A crosstalk of less than −20 dB is obtained at 1530 nm for both ‘on’ and ‘off’ switching states. The excess ‘on’-state insertion loss due to FCA loss is 0.8 ± 0.2 dB. The additional passive insertion loss of the MZ switch is measured to be 0.9 ± 0.25 dB at 1530 nm. Lower than −15 dB crosstalk is obtained over an optical bandwidth of 40 nm centered on a wavelength of 1530 nm. As such, the MZL and MZ switches have essentially the same optical bandwidth Δλ_BW. This bandwidth is determined by the wavelength sensitivity of the employed directional couplers [11]. For applications requiring wider optical bandwidths, wavelength-insensitive couplers, each consisting of an individual MZ structure, could be utilized [7,11].

The blue shift of the optimum operating wavelength of the MZL switch (1479 nm) with respect to that of the MZ switch (1530 nm) results from additional coupling in the waveguide bend sections which are immediately adjacent to the straight directional-coupling sections in each stage. This additional coupling was not accounted for in the original MZL design, but is found to be substantial for the employed partially etched waveguides (see [7] for a cross-sectional image) and the implemented bending radius of 6.5 μm.

3.2 Intrinsic switching response

Under steady-state biasing conditions, the constant current flowing through the p-i-n diodes leads to self-heating effects, which partially cancel out the free-carrier electro-optic effect and complicate the device analysis [7]. In order to assess the intrinsic, free-carrier-induced switching response of the fabricated four-stage MZL switch, fully decoupled from thermo-optic effects, a series of time-resolved transmittance measurements are made. A drive signal consisting of 80-ns long pulses with a base voltage of 0 V and a peak voltage V_D, and a 2-% duty cycle is applied to the switch. Light from a transverse-electric (TE) polarized continuous-wave laser operating at a wavelength of 1.48 μm is launched into one of the input ports of the device. The transmitted light is collected at the output port of interest, and subsequently amplified using an erbium-doped fiber amplifier (EDFA) followed by an optical bandpass filter with 1 nm bandwidth. The time-varying optical signal is then detected using a high-speed detector, and the resulting waveforms are recorded on a high-speed oscilloscope. In these waveforms, the transmitted optical power level is evaluated 60 ns after arrival of each pulse, at which point steady-state electrical conditions of the measurement setup are reached.

In order to compare the measured switching response with the simulated transmittance spectra in Fig. 2(a), the output optical power levels measured for the T₁₂ and T₂₁ transmission paths are normalized against the maximum value of the transmitted optical power, measured for each transmittance path at zero applied voltage. For the T₁₁ and T₂₂ transmission paths, the obtained optical power levels are normalized against the maximum power level measured for these transmission paths within the considered range of applied voltages (0-1.6 V). A direct comparison of the output power levels for the T₁₁ and T₂₂ transmission paths on one hand and the T₁₂ and T₂₁ transmission paths on the other hand proves difficult in these time-resolved measurements, owing to a difference in optical gain provided by the EDFA. The difference in gain originates from the large difference in average optical power fed into the EDFA for both sets of transmission paths, owing to the low duty cycle of the driving signal, in combination with saturation mechanisms in the EDFA. Therefore, the relative offset between and the T₁₂ and T₂₁ transmission curves, and the T₁₁ and T₂₂ transmission curves is estimated from their ‘on’-state transmittance values measured using continuous-wave bias voltages as discussed in section 3.1. This approach leads to a worst-case estimation for the intrinsic, FCA-related insertion loss of each switching state. Again, only optical losses originating from FCA losses are represented using this normalization scheme.

The resulting response of the MZL switch is shown as a function of applied peak voltage in Fig. 4(a) . It can be seen that in the ‘off’-state, obtained for voltages below 0.6 V, the obtained crosstalk level is lower than −18 dB, whereas the T₁₂ and T₂₁ insertion losses are relatively flat at approximately 0 dB. For voltages exceeding 1.16 V, the T₁₂ and T₂₁ transmittance are suppressed to a level lower than −15 dB. For voltages between 1.18 V and 1.28 V, the extinction of the T₁₂ transmittance is larger than −20 dB. While a switching sidelobe is apparent for V_D = 1.37 V, the T₁₂ extinction is below −16 dB for voltages up to 1.6 V, and crosstalk levels in this voltage range are lower than −15 dB. For voltages higher than 1.18 V, the excess T₁₁ insertion loss remains relatively flat and is lower than 0.45 dB. The T₂₂ insertion loss increases monotonically in this voltage range as predicted by simulation, due to increasing FCA optical losses.

 

Fig. 4 (a) Switching response of a four-stage MZL switch as a function of applied voltage, measured for 80-ns-long pulses with a 2-% duty cycle. More than −15 dB extinction of T₁₂ is obtained for V_D > 1.16 V, and more than −20 dB for 1.18 V < V_D < 1.28 V. Despite a switching sidelobe with only −16 dB extinction at V_D = 1.37 V, a digital switching response is obtained. (b) Analog switching response of the reference MZ switch. The ‘on’-state, showing more than −20 dB extinction of T₁₂, is reached at V_D = 1.1 V. The noise floor of the measurements in both (a) and (b) is −21 dB. For voltages in the range 0-0.6 V, the switching response remains essentially flat due to the electrical turn-on characteristics of the p-i-n diode phase-shifter.

Download Full Size | PDF

For comparison, the switching response of a single-stage MZ switch characterized using the same driving conditions, but measured at 1.53 μm wavelength is shown in Fig. 4(b). In the ‘off’-state, the crosstalk levels are lower than −20 dB, whereas the T₁₂ and T₂₁ insertion loss are relatively flat at approximately 0 dB. The ‘on’-state is reached at V_D = 1.1 V (estimated N ~1.4 × 10¹⁸ cm⁻³), with a T₁₂ and T₂₁ extinction of more than −20 dB. With increasing applied voltage, the oscillatory analog switching response of the MZ switch becomes apparent.

The general shape of the switching response of both the MZL as well as that of the MZ switch show excellent qualitative agreement with the switching response calculated as function of injected-carrier density shown in Fig. 2. For the MZL switch, the higher level of the measured switching sidelobe (−16 dB) as compared to the simulation (−22 dB) can be explained in part by a deviation of the fabricated coupling apodization from the ideal simulated case. This deviation originates from additional coupling in the curved waveguide sections adjacent to the straight directional coupling sections, as previously discussed in section 3.1.

From the obtained voltage-dependent switching curves in Fig. 4, the voltage-noise tolerance of both switches can be estimated. For the ‘off’-state, the voltage-noise tolerance ΔV_off is on the order of 700 mV (peak-to-peak) for both MZ and MZL switches, owing to the rectifying behavior of the forward-biased p-i-n diode [11]. For the ‘on’-state, the noise tolerance ΔV_on depends on the considered ‘on’-state bias voltage V_on. For the digital (MZL) switch, a trade-off can be made between V_on and noise tolerance ΔV_on, with a high bias voltage typically resulting in a higher voltage-noise tolerance, as evident from Fig. 4(a). As an example, if we specify the ‘on’-state voltage to be near the center of the first dip in T₁₂ transmittance (V_on = 1.23 V, estimated N ~3 × 10¹⁸ cm⁻³), the peak-to-peak voltage-noise tolerance ΔV_on for obtaining at least −15 dB and −20 dB extinction of T₁₂ is respectively 160 mV and 100 mV. At V_on = 1.31 V, ΔV_on for obtaining −15 dB extinction of T₁₂ can be as high as 300 mV. In contrast, the noise tolerance of the MZ switch at V_on = 1.1 V for the −15 dB and −20 dB extinction levels is only 45 mV and 25 mV, respectively. The data in Fig. 4 illustrates that the noise tolerance of the MZL switch is more than a factor three larger than that of the MZ switch. This advantage comes at the expense of an increased ‘on’-state power consumption of 13.9 mW for the MZL switch at V_on = 1.23 V, versus 4.7 mW for the MZ switch. The measured device characteristics of both the MZL and MZ switch are summarized in Table 1 .

Table 1. Comparison of Measured Switching Characteristics of the MZL and MZ Switch

View Table

4. Direct noise-tolerance measurements

In addition to the estimates made possible by Fig. 4, a more explicit demonstration of the noise tolerance of the MZL switch is made by applying a noisy drive signal to the switch and recording the resulting time-resolved T₁₂ transmittance waveform. The applied drive signal consists of 150-ns-long pulses with peak voltage V_D and a low duty cycle (<1%), combined with pseudo-random voltage noise generated by a 500 MHz 2⁷-1 PRBS signal with variable peak-to-peak noise voltage V_pp, as illustrated in Fig. 5(b) .

 

Fig. 5 (a) Time-resolved T₁₂ transmittance waveforms for the four-stage MZL switch, measured for 150-ns-long, drive pulses with V_on = 1.25 V and various amplitudes of added voltage noise. For 300 mV_pp voltage noise, the worst-case T₁₂ extinction is as low as −7 dB. (b) Noisy applied switching waveform, where the noise consists of a 500-MHz 2⁷-1 PRBS signal with variable peak-to-peak amplitude V_pp. (c) Overview of the worst-case T₁₂ transmittance as a function of voltage noise V_pp, comparing the reference MZ switch and the four-stage MZL switch biased at V_on = 1.25 V as well as at V_on = 1.5 V. At −15 dB extinction, the noise tolerance of the MZ is 50 ± 5 mV_pp, and that of the MZL is 140 ± 14 mV_pp, for V_on = 1.25 V. The noise tolerance at −15 dB increases to more than 300 ± 30 mV_pp for V_on = 1.5V.

Download Full Size | PDF

The resulting T₁₂ transmittance waveforms, recorded for the MZL switch with V_D = 1.25 V and voltage noise levels in the range of 0-300 mV_pp, are shown in Fig. 5(a). In the ‘off’-state, i.e. from time t = 0 ns to t = 20 ns, the T₁₂ transmittance is essentially insensitive to the noise level. This results from the rectifying behavior of the p-i-n diode, which minimizes the resulting noise of the injected-carrier density, as is also apparent from the switching response shown in Fig. 4(a). For the ‘on’-state however, the situation is vastly different. It can be seen from Fig. 5(a) that the worst-case T₁₂ extinction reduces significantly with increasing noise amplitude V_pp, from more than −20 dB at V_pp = 0 mV to only −7 dB at V_pp = 300 mV. The results of these noise tolerance measurements are summarized in Fig. 5(c), where the worst-case T₁₂ extinction is plotted as a function of applied voltage noise V_pp. For V_D = 1.25 V, the worst-case T₁₂ extinction for the MZL increases monotonically with increasing V_pp. Better than −15 dB extinction of T₁₂ is obtained for V_pp < 140 ± 14 mV, which is in excellent agreement with the switching curves in Fig. 4(a). The same measurements are repeated using an ‘on’-state voltage V_on = 1.5 V. For this ‘on’-state voltage, the worst-case T₁₂ extinction is relatively flat at about −15 dB, even for noise levels as high as 300 ± 30 mV_pp. The lower extinction level at this larger bias results from the T₁₂ switching sidelobe apparent in Fig. 4(a). However, the sidelobe-degraded performance can be avoided in future designs in which optical coupling in the curved waveguide sections is appropriately compensated. Figure 5(c) also shows the worst-case T₁₂ extinction measured for the MZ switch. As expected with reference to Fig. 4(b), the noise tolerance is much lower, with only about 50 ± 5 mV_pp noise budget for obtaining −15 dB T₁₂ extinction.

5. Bit-error-rate measurements

Finally, in order to demonstrate the ability of the MZL switch to route high-speed data streams, even in the presence of substantial voltage noise, a series of bit-error-rate (BER) measurements are performed, using a 40 Gbps 2⁷-1 PRBS input optical data stream. First, the switch is kept at zero bias voltage, and the data stream is fed into the a₁ input port. The transmitted data stream is collected at output port b₂, and a BER curve is recorded for the T₁₂ transmission path. Subsequently, the switch is biased (continuous-wave) to V_D = 1.3 V, and a BER curve is recorded for the T₁₁ transmittance path. Finally, the switch bias is maintained at 1.3 V, while a substantial amount of voltage noise (200 mV_pp) is added in the form of a 2 GHz 2³¹-1 PRBS signal. The results are shown in Fig. 6 . The ‘off’-state and the ‘on’-state without noise have essentially identical BER curves. The BER curve recorded for T₁₁ transmission in the presence of 200 mV_pp voltage noise shows a power penalty of only 0.7 ± 0.2 dB with respect to the T₁₁ BER curve without added noise. These results indicate the ability of the switch to route 40 Gbps data streams with low error count, even in the presence of substantial voltage noise.

 

Fig. 6 Bit-error-rate (BER) curves measured for the four-stage MZL switch using a 40 Gbps 2⁷-1 PRBS optical data stream, both for T₁₂ transmission in the ‘off’-state (blue), as well as for T₁₁ transmission in the ‘on’-state (V_on = 1.3 V, continuous-wave drive, green). A BER below 10⁻¹⁰ can be obtained for both transmission paths. The orange data set shows the BER curve recorded for T₁₁ transmission at V_on = 1.3 V with 200 mVpp of added noise. A BER below 10⁻¹⁰ can still be obtained, albeit with a 0.7 ± 0.2 dB power penalty. The insets show 40 Gbps 2³¹-1 PRBS eye diagrams for the respective transmitted optical data streams, recorded for 3 mW of received optical power. The black scale bar represents a 20 ps time interval.

Download Full Size | PDF

6. Discussion

The switching characteristics of the MZL and MZ switches presented in sections 3, 4 and 5 agree very well with the simulated switching performance presented in section 2, and more generally confirm the design methodology for digital switches presented in our earlier paper [11]. However, the experimental data obtained here enables us to refine the p-i-n diode model, which in [11] was assumed to be an ideal diode, featuring an infinite carrier lifetime and negligible series resistance. These assumptions led to an underestimation of the noise-tolerance of both the MZ switch (estimated to be only 5 mVpp for −20 dB crosstalk in [11]) and the MZL switch.

In the present work, the noise tolerance measured for the fabricated MZ switch is as high as 25 mV for obtaining less than −20 dB crosstalk. This higher voltage-noise tolerance results from the voltage drop across the series resistance of the diode, which in turn increases with the amount of current flowing through the device. This current is inversely proportional to the free-carrier lifetime, which itself strongly depends upon the carrier recombination rate in the intrinsic region of the p-i-n diode, primarily resulting from surface trap states and other defects in the silicon lattice [14]. As such, a short carrier lifetime in the diode actually helps to improve the voltage-noise tolerance of a carrier-injection-based interferometric optical switch. However, a short carrier lifetime also results in higher power consumption and the associated parasitic self heating of the diode, which can strongly degrade the switching performance for long ‘on’-state durations [7]. Therefore, carrier-injection-based switches preferably make use of p-i-n diode phase shifters with long carrier lifetimes, which can be obtained by electrically passivating the surface of the diode’s intrinsic volume [14]. The electro-optic behavior of such a well-passivated p-i-n diode is expected to more strongly resemble that of an ideal diode, in which case, adoption of a digital design as presented in this paper will be even more important for obtaining switches with sufficient voltage-noise tolerance.

Finally, even with the relatively high power consumption of about 14 mW currently obtained for the MZL switch, a low switching energy per bit could still be obtained for a WDM-encoded data stream. Assuming that a 15 × 40 Gbps WDM data stream (i.e. 15 channels with coarse 2 nm spacing) is simultaneously switched by the device, the switching energy per bit is only 23 fJ/bit.

7. Conclusion

We have demonstrated noise-tolerant operation of a broadband silicon electro-optic switch, based upon a multi-stage Mach-Zehnder lattice design. The multi-path interference at the output ports of the switch is engineered to produce a digital switching response, which results in a voltage-noise budget as high as 300 mV_pp for obtaining lower than −15 dB crosstalk levels. This is more than six times better than a conventional single-stage MZ switch. A 40 Gbps data stream can be switched error-free by the device in the presence of 200 mV_pp voltage noise, with a power penalty of only 0.7 dB. Other attractive features of the MZL switch are the higher obtainable ‘on’-state extinction levels and lower FCA-induced ‘on’-state insertion loss. The demonstrated switching characteristics are highly desired for switches that route network traffic at the nodes of proposed optical-networks–on-chip.