1.

Introduction

Artificial intelligence and machine learning workloads are changing all aspects of the compute infrastructure—from data centers and cloud hyperscalers to high-performance computers (HPC).¹ Data centers are being enhanced with specialized HPC-like hardware to handle machine learning workloads. Hyperscale cloud infrastructure is enhanced with rack-scale and distributed solutions. Optical switching solutions have been proposed to improve the utilization of the computer resources as well as reducing latency and energy consumption.^2–6 Silicon photonic switches are particularly attractive for such applications, and several research groups have developed silicon photonics switches based on cascaded Mach–Zehnder interferometers.^7–14 However, their optical losses increase rapidly with port count, limiting the scalability of these switches.

Our group has previously demonstrated micro-electro-mechanical-system (MEMS)-based silicon photonic switches with moving waveguides. The switches are highly scalable and have several salient features. The $240\times 240$ switch is the largest monolithic switch ever reported.¹⁵ They also have fast switching time ($<1\mu \mathrm{s}$) and broadband operation ($>120\mathrm{nm}$). Polarization insensitivity has been addressed using multi-layer bus waveguides.¹⁶ However, these switches were made with custom fabrication process that are not commonly accessible. On the other hand, simpler single-layer MEMS devices can be made in standard silicon photonics process.¹⁷ $8\times 8$ crossbar switches¹⁸ and $4\times 20$ multicast switches¹⁹ have been reported, though they are not fabricated in foundries. Though the optical loss is not as low as those of more sophisticated multi-layer switches, this simple process is attractive for medium-sized switch and potential integration with other well-established silicon photonic elements including germanium photodetectors for in situ monitoring. In addition, the inclusion of silicon photonic MEMS process in foundry could open up a promising route to build programmable photonics circuits.²⁰ In this paper, we report on a $32\times 32$ silicon photonic MEMS switch fabricated on 200-mm wafers at a commercial complementary metal-oxide-semiconductor (CMOS) foundry (TSI Semiconductors). Switching is accomplished by directional couplers with continuously tunable gaps. The switch exhibits high uniformity, and the power consumption is several orders of magnitude lower than thermo-optic switches.

2.

Switch Design

Figure 1 shows the architecture of the switch. It consists of a crossbar network with 32 input and 32 output waveguides. The switch is designed for a transverse electric (TE) mode operation. The waveguide crossings have been reported in Refs. 21 and 22. Grating couplers are used for input and output for ease of testing. The grating coupler is also designed for TE polarization, has a 640-nm pitch with a 50% duty cycle, and is defined with a 70-nm-deep etch. The bandwidth of the grating coupler is 30 nm. Each switch cell has two directional couplers that are tethered to a common comb-drive actuator moving in the diagonal direction. Similar comb drives were used in Ref. 23. The length, thickness, and width of the coupler waveguides are $20\mu \mathrm{m}$, 220 nm, and 450 nm, respectively. The coupling gap can be continuously varied from 550 nm to nearly physical contact ($\sim 0\mathrm{nm}$) by the comb-drive actuator. To achieve a high-density switch, the footprint of the comb drive should be minimized. At the same time, we need as many comb finger pairs as possible to reduce the switching voltage. We accomplished this using deep submicron comb fingers (300 nm) and spacing (400 nm), which yield 44 pairs of comb fingers. This gives us a switching voltage of 9.45 V. The actuator has a footprint of $88\mu \mathrm{m}\times 88\mu \mathrm{m}$, well within our waveguide pitch of $166\mu \mathrm{m}$.

Fig. 1

The architecture of the silicon photonic MEMS switch with gap-adjustable directional couplers. Light is coupled to the chip using the grating couplers. There are two pairs of the directional couplers and one comb-drive actuator per unit cell. The light paths on the chip are controlled by changing the gap spacing of each directional coupler.

Figure 2(a) shows simulated optical field ($E_x$) along the directional coupler with 550 nm gap. In this state, the coupling between the waveguides is weak, and all light from input port goes to through port. Since the coupling coefficient depends exponentially on the gap spacing, the coupling increase quickly with reducing gap. At 135-nm gap spacing [Fig. 2(b)], nearly 100% of the light goes to drop port.

Fig. 2

Simulated optical performance of the gap-adjustable directional coupler. Unless otherwise stated, the width of the waveguide used for the simulations is 450 nm. (a) Optical field ($E_x$) profile along the directional coupler with 550-nm gap. All light from the input port goes to the through port. (b) Similar profile when the gap spacing of the directional coupler reduces to 135 nm. Light is directed to drop port. (c) Transmission characteristics of the coupler versus gap size. (d) Transmission characteristics of the coupler when the width of the waveguide is changed to 400 nm. (e) Optimal gap versus the waveguide width of directional couplers.

Figure 2(c) shows the simulated transmission of the directional coupler with various gap sizes. The simulation shows that an extinction ratio (for drop port) of 25.8 dB can be achieved by changing the gap size from 550 to 135 nm. Since there are two directional couplers per switch unit cell, the extinction ratio for switching is 51.6 dB. The switching devices using directional couplers^7–13 often suffer from optical power leakage or loss due to the dimensional shift of the directional couplers during fabrication step. However, the gap-adjustable directional coupler allows flexible tuning of the gap and thus allows compensating the dimensional shift. Figure 2(d) shows the simulated transmission characteristics of the gap-adjustable directional coupler when the width of the waveguide changes to 400 nm. It shows that optical switching occurs with 210 nm of gap size. Large extinction ratio (17.9 dB for drop port and 47.5 dB for through port) can still be achieved by changing the gap size between 550 and 210 nm. Figure 2(e) shows optimal gap sizes for maximum power transfer to drop port for various widths of the waveguide. When the width changes from 400 to 500 nm, one can still achieve optimal coupling with the gap sizes in the range from 210 to 90 nm. Therefore, the dimensional accuracy for the waveguide can be relaxed during the fabrication process.

3.

Fabrication

The switch can be fabricated using any standard “passive only” silicon photonics process. Figure 3 illustrates the fabrication flow we used. The fabrication starts with 200-mm SOI wafer with 220-nm-thick silicon device layer and $3\text{-}\mu \mathrm{m}$-thick buried-oxide (BOX) layer. The waveguides, grating couplers, directional couplers, and the comb-drive actuators are patterned by two lithographical steps: a partial and a full silicon etch processes.

Fig. 3

(a) Cross section of the switch and (b) fabrication flow of the switch. (1) The fabrication starts with an SOI wafer. (2) First lithography and partial etching of the silicon layer with 70-nm depth. (3) Second lithography and full etching of silicon layer with 220-nm depth. (4) Third lithography and metal (Au/Cr) deposition. (5) Metal pads patterned by lift-off process. (6) Vapor HF etching to release the moving parts.

A third lithography is used to define metal pads. As a last step, a hydrofluoric (HF) acid vapor is introduced to etch the BOX layer under the waveguide and the actuators for MEMS actuation. Figure 4(a) shows the scanning electron microscopy (SEM) image of the switch array fabricated in the foundry. Figure 4(b) shows the image of the full chip. The $32\times 32$ switch matrix occupies an area of $5.9\mathrm{mm}\times 5.9\mathrm{mm}$. The grating couplers and routing waveguides occupy an additional area of $0.7\mathrm{mm}\times 8.4\mathrm{mm}$. Figures 4(c) and 4(d) show the optical and SEM images of the switch unit cell, respectively. Figure 4(e) shows the SEM image of comb fingers and spring. Figure 5(f) shows the SEM images of the directional coupler. There are 44 pairs of comb fingers in the comb-drive actuator. The width of the comb fingers is 300 nm and the gap between the comb fingers is 400 nm. The suspended comb is supported by four folded springs. The width and the length of the springs are 300 nm and $30\mu \mathrm{m}$, respectively. The thickness of the comb fingers and the springs are the same as the silicon device layer of the SOI wafer, which is 220 nm. The initial gap of the directional coupler is 550 nm. With this design, the switching voltage is calculated as 13.8 V, and the resonance frequency of the actuator including all moving part is calculated as 58.4 kHz.²⁴ The detailed calculation for the switching voltage and resonance frequency can be found in Ref. 25.

Fig. 4

Fabricated silicon photonic MEMS switches. (a) Perspective SEM image of the fabricated silicon photonic switch array. (b) Optical microscope image of the fabricated device. (c) Optical microscope image of the switch unit cell. In each unit cell, there are 2 pairs of the directional couplers, 4 folded springs, 44 pairs of comb fingers, and 1 waveguide crossing. (d) SEM image of the switch unit cell. (e) SEM image of the comb drive and spring. (f) SEM image of the directional coupler.

Fig. 5

(a) Measured transmission response of the switch unit cell with various applied voltages. (b) Measured spectral response of the switch unit cell when the unit cell is at on-state.

4.

Measurement

Figure 5(a) shows the measured transmission characteristics of the switch unit cell with various applied voltages at 1550 nm wavelength. As the switch is designed for TE mode, all the measurement is done with the TE polarization. At 0 V, optical transmission to the drop port is negligible (cell at off-state) and all the optical power goes to the through port. At 9.9 V, the optical transmission to the drop port reaches maximum (cell at on-state), whereas the optical power transmitted to the through port is minimum. The on-off extinction ratios are 50.8 dB for the drop port and 27.9 dB for the through port. Figure 5(b) shows the spectral response of the switch unit cell when the unit cell is at on-state. The switching is optimized at near 1550 nm wavelength. The 20-dB-bandwidth of the through port is measured to be 28.7 nm.

Figure 6(a) shows the distributions of the on-state voltage of the switch unit cells. The average and the standard deviation of the switching voltages are 9.45 and 0.74 V, respectively. The response time of the switch was measured as $50\mu \mathrm{s}$, which is longer than the value inferred from the resonant frequency of the actuator (i.e., 58.4 kHz). This is because the high resistivity of the wafer substrate caused a large RC delay in the actuator. However, as demonstrated in our previous work,²⁴ using a wafer with a low-resistive substrate can reduce the switching time below $10\mu \mathrm{s}$ with the same actuator design. We measured various optical paths in the $32\times 32$ switch to find the path dependence insertion loss of the switch [Fig. 6(b)]. The loss per unit cell is 0.10 dB, extracted from the slope of the linear fit line of the data. The $y$ intercept (= 0.51 dB) includes the optical losses of the two directional couplers at on-state and the optical losses of the routing waveguides to/from the grating couplers. Therefore, the fixed switching loss is $<0.51\mathrm{dB}$. The maximum on-chip insertion loss for the $32\times 32$ switch is only 7.7 dB. There are optical test structures for measuring the propagation loss of the waveguides with cut-back method. The propagation loss of the 450-nm wide fully etched waveguide (same dimension as the waveguide used in the directional couplers) and the 450-nm wide shallow etched waveguide is measured as 1.7 and $0.8\mathrm{dB}/\mathrm{cm}$, respectively. Insertion loss of the grating coupler is measured as 3.8 dB at 1550 nm wavelength.

Fig. 6

(a) Distributions of the on-state voltage of the switch unit cells. (b) Measured path-dependent insertion of the switch.

5.

Conclusion

In summary, we have demonstrated $32\times 32$ silicon photonic MEMS switch on a single-layer silicon-on-insulator substrate with gap-adjustable directional couplers. The switch chip is successfully fabricated at a commercial CMOS foundry on 200-mm wafers. The switch array area is $5.9\mathrm{mm}\times 5.9\mathrm{mm}$. The use of a foundry-based silicon photonic process allows the switch to be intimately integrated with other silicon photonic integrated circuits. The continuously tunable gap allows us to compensate for the variation of the waveguide width in the directional couplers, though this does require a customized bias for each switch cell to achieve the best extinction ratio. One way to control the bias precisely is to use the through port signal for feedback control. Alternatively, the need for precise bias control can be alleviated by replacing directional couplers with adiabatic couplers, as in our other reported switches.^15,21 The operating voltages of the switch cell are quite uniform ($\text{standard deviation}=0.74\mathrm{V}$), and the maximum on-chip insertion loss of the switch is only 7.7 dB. We believe that this work shows a promising path for mass production of large-scale silicon photonic switches at commercial foundries.