Wavelength-selective optical switches help scale the datacenter

Oct. 20, 2021
Datacenters will be transformed by the use of silicon photonic integrated circuits; here, an 8 × 4 multiwavelength-selective ring-resonator-based switch matrix is used for interconnecting electronic packet switches in scalable datacenters.

Wavelength-selective optical switches, when used with electronic switches in datacenters, allow scaling of datacenters with fewer physical connections. These switches can reconfigure bandwidth on demand and have lower latency than single-wavelength switches. Wavelength-selective switches have wavelength-division-multiplexed (WDM) signals as inputs. These switches can route none, one, or many of the wavelengths from an input port to an output port. Single-wavelength switches can route only one wavelength from an input to an output. Figure 1a shows the application of a wavelength-selective switch in a leaf-spine datacenter architecture and Figure 1b shows the application of the switch in Hyper X (another type of datacenter switching architecture).1

The contribution of this work is as follows:

  •  Multiple wavelength-selective switches have been designed and fabricated in a 220 nm silicon photonics foundry as a part of the AIM Photonics program.
  • The optical loss of the switch was reduced by 70% in the new 4 × 4 switch as compared with an 8 × 4 switch.
  • Error-free data transmission was shown with a 40 Gbit/s nonreturn-to-zero (NRZ) signal and multiple input channels.
  • Error-free transmission was shown through different paths of the switch with a 111 Gbit/s pulse-amplitude-modulation 4-level (PAM4) signal.
  • A wavelength-locking scheme was shown with on-chip wavelength monitors.
  • A multicasting and wavelength selective switch was proposed with half the number of devices as compared with the original design.

Wavelength-selective switch architecture

Figure 2a shows the architecture of the switch. The switch has N inputs, N outputs, and M wavelengths per input. There are L microring-resonator filters at each cross point. The switching blocks are connected with optical waveguides and the microring resonators are thermally tuned. The waveguide crossings use multimode interference to reduce crosstalk between the crossing waveguides. The microring resonators should tune by more than 20 nm for 8 WDM channels at 200 GHz spacing, should have low off-resonant pass-through loss, and should have high tuning efficiency.

Due to the high index contrast of silicon photonic waveguides, microring resonators and devices are smaller in size as compared with other fabrication platforms. The size of the switch is pad-limited. These electrical pads are 100 × 100 μm in size as compared with 10 × 10 μm, the size of a single microring resonator.

Time is divided into time slots. At the start of every time slot, the arbitration algorithm assigns connections and wavelengths. There are buffers at each transmitter and each transmitter can modulate any data on any wavelength channel. The switch is rearrangeably nonblocking, but operates equivalently to a nonblocking switch as it is reconfigured at the start of every data transmission slot. For L=1 in Figure 2a, the switch operates similarly to an arrayed waveguide grating router (AWGR). For L>1, we require contention resolution as wavelengths are shared between different ports. A centralized arbitration scheme is used for contention resolution in our architecture.

Figure 2b shows a 8 × 4 switch and Figure 2c a 4 × 4 switch. A foundry-provided process design kit (PDK) was used to design the former switch, while the latter switch was designed with devices from literature. The optical losses were improved by 70% from the 8 × 4 switch to the 4 × 4 switch.4 Ring resonators were used with both thermal and electro-optic tuning in the 8 × 4 switch. These ring resonators have doped carriers on the path of light, which increased the optical loss. With thermal tuning, the rings tuned with tuning efficiency of 0.84 nm/mW with a negligible optical-loss penalty as compared with electro-optic tuning. One can tune by a full free spectral range (FSR) with thermal tuning with negligible loss. Our rings tuned by 11 nm out of 26 nm FSR for our rings in the 8 × 4 switch, 17 nm out of 20 nm FSR in the 4 × 4 switch, and we could tune by more than one FSR in the newest generation devices. With electro-optic tuning, a loss of 3.6 dB was observed for every nanometer tuned.

The ring resonator design was changed in the 4 × 4 switch: ring resonators with Bezier curves were used to reduce scattering loss between the optical waveguide and the ring curve, and p-doped heaters were used that do not touch the ring waveguide. Both these factors reduced the off-resonance loss of the ring resonator, which reduced the path loss of the switch.

High-speed data transmission experiments

In addition, high-speed, error-free data transmission was experimentally demonstrated through different switch paths for both switches. Figure 3a shows bit-error-rate curves for a high-speed transmission experiment.2 In this experiment, input light was simultaneously injected through multiple inputs of the switch with a 40 Gbit/s NRZ signal. Light was split from a single tunable laser, modulated with a high-speed modulator, and passed through an optical splitter; fiber loops of different lengths were then used for different paths to decorrelate the signals from each other. This emulates multiple transmitters at the same wavelength. In this measurement, the signal degradation due to incoherent crosstalk was measured. Incoherent crosstalk is caused due to optical signals at the same wavelength and from different transmitters at different input ports of the switch. It was shown that incoherent crosstalk has negligible power penalty for the switch. Figure 3b shows the bit-error-rate curve for a 111 Gbit/s PAM4 signal. Error-free transmission was shown through different paths of the switch with HD-FEC limit.4

In 2017, we designed the biggest wavelength-selective crossbar switch in silicon photonics.5 The switch had 8 input ports and 8 output ports and had on-chip wavelength monitors to lock the switch rings. Microring resonators are sensitive to temperature variation and require feedback algorithms to lock them to the channel of interest. We demonstrated a wavelength-locking scheme that requires a one-time calibration; the wavelength locking can be done with on-chip monitors. This technique can be used for any photonic switch.

A multicasting and wavelength-selective switch was also shown with half the number of devices and area. This switch is made up of two switches—a multiwavelength-selective crossbar switch and a broadband switch. The switch uses partial-power-drop microring resonators for multicasting. These devices can drop a small fraction of the power from the input port to the output port. The switch has half the power consumption due to almost half the number of devices as compared with the previous switch architecture.6

We proposed and built 8 × 8, 8 × 4, and 4 × 4 switches in a silicon photonics foundry. We performed high speed bit-error-rate measurements for multiple switch paths. Our 4 × 4 switch had a lower optical loss as compared with 8 × 4 and 8 × 8 switches due to thermal tuning in the microring resonators as compared with electro-optic and thermal tuning used in 8 × 8 and 8 × 4 switches. We also experimentally demonstrated a wavelength-locking scheme to lock microring resonators at different stage temperatures.

Avenues for further development

These switches have a large number of electrical pads and the area of the switch is pad-limited. We need to reduce the pad size to increase the port counts of these switches. With optical interposer technology, copper pillars, or through-silicon vias, one can reduce the pad size and build bigger switches. The optical loss of the switch paths increases as the ring resonator off-resonance loss increases. We hope that, with better fabrication techniques and better ring designs, the off-resonance loss of the ring resonators can be reduced.


1. A. A. M. Saleh et al., Proc. OECC, 13 (2016).

2. A. S. P. Khope et al., Opt. Express, 27, 4, 5203–5216 (2019).

3. C. L. Manganelli et al., IEEE Photon. J., 9, 1, 1–10 (2017).

4. A. S. P. Khope et al., Opt. Lett., 45, 19, 5340–5343 (2020).

5. A. S. P. Khope et al., Opt. Lett., 42, 23, 4934–4937 (2017).

6. A. S. P. Khope et al., Opt. Lett., 46, 2, 448–451 (2021).     

About the Author

Akhilesh S. P. Khope | Software Engineer, Microsoft

Akhilesh S. P. Khope is a software engineer at Microsoft (Redmond, WA) and was previously with the Electrical and Computer Engineering Department at the University of California, Santa Barbara.

Sponsored Recommendations

Request a free Micro 3D Printed sample part

April 11, 2024
The best way to understand the part quality we can achieve is by seeing it first-hand. Request a free 3D printed high-precision sample part.

How to Tune Servo Systems: The Basics

April 10, 2024
Learn how to tune a servo system using frequency-based tools to meet system specifications by watching our webinar!

Motion Scan and Data Collection Methods for Electro-Optic System Testing

April 10, 2024
Learn how different scanning patterns and approaches can be used in measuring an electro-optic sensor performance, by reading our whitepaper here!

How Precision Motion Systems are Shaping the Future of Semiconductor Manufacturing

March 28, 2024
This article highlights the pivotal role precision motion systems play in supporting the latest semiconductor manufacturing trends.

Voice your opinion!

To join the conversation, and become an exclusive member of Laser Focus World, create an account today!