Silicon based integrated photonic chip technology for datacenter optical interconnection

With the rise of the modern Internet of Things, big data and other industries, the human society’s demand for information increases rapidly, and the demand for broadband network capacity shows a dramatic growth trend. As its core support, optical communication system and data center are facing major challenges. With the characteristics of small volume, light weight and low power consumption, optoelectronic integrated devices are the key devices to break through the bottleneck of optical communication system and data center optical interconnection. In this paper, silicon-based integrated optical switching technology and multiplexing technology applied for datacenter optical interconnection are studied. The implementation methods and technical bottlenecks of silicon-based integrated optical switching chips and mode multiplexing/demultipelxing chips are discussed. Moreover, the future of silicon-based integrated photonic technology is prospected.


Introduction
With the deepening of information technology in today's society, new technologies such as cloud computing, big data mining and block chain emerge along with the development of information technology, changing our lives all the time.In particular, information technology, represented by computers and the Internet and mobile communications, connects everyone.A variety of new Internetbased user experiences, such as high-definition movies, virtual reality (VR), live streaming and video conferencing, have created a huge demand for traffic.
In recent years, the development of cloud computing has resulted in an explosive growth of the traffic in data centers.In this case, the interconnection of data centers will face the following technical challenges, including high bandwidth requirements, high switching capacity requirements, low cost requirements and low energy consumption requirements.Due to the physical characteristics of photon, photon communication technology has the characteristics of large bandwidth, ultra-high speed, strong security and so on.Data center is an important application field of photon technology, and optical connection has quite obvious advantages.The optical interconnection architecture is capable of providing greater transmission bandwidth.The optical switching architecture can achieve higher switching capacity due to the use of light switching.Due to the high transmission quality of optical fiber link, the network overhead of optical fiber link can be lower than that of traditional cable link.Because optical fiber has lower loss, it is easy to achieve longer transmission distance, so the optical link can adopt lower transmission power.Moreover, there is no need of optic-electric-optical conversion process in all-optical network.Optical switching architecture can be constructed using passive or low-energy optical devices.The communication between high-density data interfaces requires a large number of optical fibers, which solves the interconnection problem, but it is bulky and difficult to maintain later.
Silicon optical chip has the characteristics of compact structure, portability and low energy consumption, which can meet the demand of bandwidth and cost of data center.It is an opportunity for the development of silicon photonic integration technology.In this paper, silicon based integrated largescale optical switching technology and high speed optical transmission technology suitable for data center are studied.The switching time of MZI electro-optical switch based on plasma dispersion effect is usually in the order of nanosecond and has a large bandwidth.In the post-Moore era, reducing chip size is a crucial development direction.With the helical waveguide structure, the original device size can be further reduced without reducing the length of the modulation arm.In addition to the size of the device, the operating bandwidth also plays a decisive role.Larger bandwidth, more stable optical ratio means that a wider range of conditions can be met.In 2017, Wang et al. proposed a wideband 2×2 electro-optical switch based on MZI and reverse symmetric DC [3].By replacing the conventional coupler composed of uniform waveguide with reverse symmetric coupler, 50% spectral ratio can be achieved in a wide wavelength range.At 1550nm operating wavelength, when the driving voltage is 6V, the extinction ratio is greater than 15dB, the bandwidth is 85nm, and the insertion loss of all channels is less than 8.5dB.

Silicon-based optical switching technology
Due to the absorption of optical signals by plasma dispersion effect, it has an adverse effect on crosstalk.Compared with silicon-based electro-optical switch, the response speed of silicon-based optical switch based on thermal-optical effect is only in the order of microseconds, but the influence of thermal tuning on loss and crosstalk is relatively small.In recent years, with the continuous optimization of manufacturing process and structure, the performance of silicon based thermos-optic switch has been further improved.

MRR optical switch.
Compared with MZI optical switch, the optical switch based on microring resonance effect has smaller size and lower loss.However, because resonance is more sensitive to temperature and wavelength, the bandwidth of MRR optical switch is smaller than that of MZI optical switch, and the microring structure puts forward higher requirements on the fabrication process.MRR optical switch includes traditional MRR optical switch and DR-MZI optical switch [4][5].The typical MRR type 2×2 optical switch unit is shown in figure 2, which is mainly composed of input-direct waveguide, upload-download waveguide and tunable microring.The signal light is input from the input port.When the microring does not resonate, the signal light is output Through the port.When the microring has harmonic vibration, the light is output from the Drop port after coupling through the resonant cavity, so as to realize wavelength switching.In the actual design and preparation process, MRR optical switch faces two problems to be solved.One is that the spectral characteristics of the microring resonator limit the bandwidth, which is not conducive to the suppression of cross-talk between channels.Second, the preparation of the microring resonator has high technological requirements.The physical size and structural accuracy should effectively meet the resonant conditions, and minimize the influence of external temperature changes and wavelength drift of incident light on the MRR output.In order to increase the working bandwidth, a microring cascade structure can be used to achieve spectrum flattening and continuous tuning.
Due to its own characteristics, MRR optical switch will inevitably cause large crosstalk, while MZI optical switch can regulate crosstalk well, but its size and loss are relatively large.In order to further enhance the silicon-based optical switch performance and reduce the size, Lu proposed a DR-MZI silicon-based optical switch structure [6].The advantage of this scheme combines resonance and interference method, realized the compared to using MZI alone or MRR structural lower crosstalk and loss, effectively reduce the power consumption of the device and size.It provides a new idea and direction for constructing high performance large-scale switch array.

Topology of waveguide optical switch array
The optical switch array in silicon-based photonic chip is usually composed of 1×2 switching unit or 2×2 switching unit connected and combined according to certain logical relation.The large-scale silicon optical switch array is mainly composed of input ports, switching networks and control units.The topology of switching networks, namely optical switch arrays, can be roughly divided into blocking, reconfigurable non-blocking and strictly non-blocking types according to different blocking types.

Blocking architecture.
Blocking network architecture is the first proposed topological structure, typical of which is Butterfly architecture [7], as shown in figure 3. The topology structure has horizontal symmetry.The number of switch units required by N×N switch array is 1 2 log 2 , the number of switch units in each path is log 2 , and the number of crossover nodes in each path is  − log 2  − 1.Compared with reconfigurable non-blocking networks and strictly non-blocking networks, its advantage is that the number of switch units is small, and can meet most of the switching needs.A smaller number of switch units is beneficial to reduce loss and packaging difficulty, control the cost, and enhance practicability.The challenge is that the network can not meet the special needs of each path independent of each other.

Reconfigurable non-blocking architecture.
A typical reconfigurable non-blocking switch is the Benes architecture [8].As shown in figure 4, the total number of switch units required by the N-port array is  log 2  −  2 , and the number of switches in each path is 2 log 2  − 1.The whole network architecture presents a center symmetric structure.The flexible configuration of the Benes network enables non-blocking transmission of optical signals by resetting the cell state.In the non-blocking network framework, Benes has the lowest number of switching units and has the advantage of lower losses compared to strictly non-blocking networks.The disadvantage of Benes network is that each path is not completely independent.Since the same switch can control two transmission optical signals, the crosstalk is relatively high.Therefore, it is necessary to further optimize the unit switch to reduce the crosstalk performance to improve the overall performance of the array.The schematic of strict non-blocking architecture.The total number of switch units in the network of port Crossbar is  2 , and the number of switches in each path is 2 − 1, with no light cross.The main characteristics of the Crossbar network architecture are its strict non-blocking characteristics and zero waveguide crossing.However, each path of the Crossbar structure passes through different number of switch units, that is, the length of the optical signal path is different.The shortest path passes through only one switch unit, while the longest path needs to pass through 2 − 1 switch units.Therefore, different paths lead to a large difference in loss, thus affecting the loss uniformity among the array switch paths.Although the total number of switch units in the PILoss network is  2 , the number of switches on each path is N, and the number of crossing nodes on each path is  − 1.Compared with Crossbar networks, PILoss architecture has better path loss uniformity.Loss is path independent, but the network structure is relatively complex.
The total number of switches in the N-port S&S network architecture is 2 2 − 2, the number of switches in each path is 2 log 2 , and the number of fork nodes in each path is ( − 1) 2 .Each path of the S&S network has a completely independent optical link, and there is no overlap between paths except the cross waveguide.Therefore, crosstalk between paths of the S&S network is extremely low.However, with the increase of the number of ports, the complexity of the network increases rapidly, and the number of crossing nodes of each path increases rapidly with  2 , and the loss and crosstalk will deteriorate accordingly, which is difficult to achieve.From the perspective of practical application, reconfigurable non-blocking network can meet most application requirements.With the increase of the number of ports, the number of switches is relatively small, typical Benes network architecture has many applications in large-scale array optical switches.
In 2019, Suzuki et al of AIST in Japan reported a 32×32 strictly non-blocking optical switch array based on silicon thermo optical effect [9], with a switching time of about 10us.The unit switch adopts DC coupler and realizes fiber -fiber loss of 10dB through LC converter.In the same year, the team reported an optical switch array based on Si/SiN double-layer waveguides.In this scheme, 32×32 3D optical switching array is realized by using the free ports in PILoss topological architecture, and the polarization dependence of silicon waveguide is overcome by SiN/Si double-layer waveguide and crossover structure.

Wavelength division multiplexing (WDM) technology
WDM technology is to realize parallel transmission of multi-channel data on a single fiber/waveguide by using multiple different wavelengths of light, which greatly expands the communication capacity of optical interconnection.Its key function is how to combine or separate the multiplex data carried by different wavelengths, the corresponding key device is WDM device.The basic principle of WDM devices is to use beam interference, which can be divided into two categories: double beam interference and multi-beam interference.Compared with dual-beam interferometer (such as Mach-Zehnder interferometer), multi-beam interferometer can achieve narrower bandwidth filtering and easily achieve multi-channel dense wavelength multiplexing.The most common multi-beam interference wavelength division multiplexing devices include AWG, etched diffraction grating (EDG), MRR, etc.Among them, AWG and EDG are quite similar in structure and principle, which can realize parallel multi-channel.MRR can realize serial multi-channel through cascading structure.

AWG
In the process of conventional silicon nanowires AWG, with the decrease of wavelength channel interval, the size of AWG device increases significantly, and the corresponding device performance deteriorates significantly.Therefore, it is a challenge to realize AWG devices with intensive silicon nanowires and excellent performance.In order to solve this problem, it is worth trying to introduce the design of cascading comb filter.A comb filter is utilized to input a set of channels with Δλch interval signal (λ1, λ2, λ3, λ4, ... λN) divided into odd groups (λ1, λ3, λ5...) and even groups (λ2, λ4, λ6...).Then two channel interval of 2 Δλch WDM devices are used to separate the odd group of channels, even group of channels respectively.This method can significantly reduce the difficulty of implementing WDM devices [10].Comb filters have been used in intensive WDM modules based on traditional large section SiO2 optical waveguides, but there are few reports about silicon nanowire comb filters and their integration with AWG monographs.

MRR filter
MRR is a kind of classical integrated photonic device which can realize multiple functions.It has two typical structures: add-drop type and all-pass type.It is widely used in optical filtering and optical modulation.For the application of optical filtering, it is usually hoped to realize the optical filter with square response spectral line, so as to make it have greater tolerance and avoid the adverse effect of resonant wavelength drift caused by environmental interference.In order to realize the square filtered spectral line, the multi-loop cascade structure can be used, and the coupling coefficient of each coupler is designed carefully.In Ref [11], multimode interference (MMI) couplers are used.However, the cost is to introduce additional loss, increase the length of the resonant cavity (Lcav=2πR+2LMMI, resulting in a smaller FSR), and the coupling coefficient of the MMI coupler is a fixed value, which cannot be further adjusted according to the need.To overcome this problem, the Ref [12] introduced a bending directional coupler to replace the MMI coupler.In this structure, the widths of two coupling waveguides (W1, W2) should be optimized to meet the bit matching condition.In this case, even if a large coupling waveguide spacing (such as 150 nm) is selected, a sufficiently large coupling coefficient can be obtained by increasing the length of the coupling region.

Modular division multiplexing (MDM) technology
MDM can further expand the communication capacity by using different modes of light to transmit information in the same optical waveguide.MDM technology adds new dimensions of reuse, enabling optical communications to achieve data transmission speeds and bandwidths that electrical interconnection cannot.Mode multiplexer/demultiplexer is the key device to realize mode multiplexing on chip.The principles of mode multipelxer /demultiplexer can be divided into multimode interference, phase matching, and pattern evolution.
The MMI structure was first proposed by Leuthold in 1998 [13].The device comprises a cascaded MMI plus a phase shifter structure.In 2012, Uematsu transplanted the scheme to On the SOI platform of 220 nm thick top layer silicon [14], as shown in Fig. 6.However, this structure is too complicated.And phase shifter process is sensitive, the realization is very difficult.More importantly, mode multiplexers based on multimode interferometry are difficult to multiple more modes.In contrast, multiplexers based on phase matching and pattern evolution are scalable.
The structure of pattern multiplexer based on phase matching is very simple.By selecting the appropriate waveguide width to meet the phase matching conditions, high mode conversion efficiency can be obtained.Usually this structure has a compact size which facilitates large size, which is benefit to integrate with large scale.Figure 7 shows the schematic of 2-mode mode multipelxer which is composed of asymmetric coupler [15].But this kind of mode multiplexer needs accurate phase matching and the mode propagation constant is closely related to the waveguide width.This means that the mode conversion efficiency is sensitive to the change of waveguide width.In order to improve the process tolerance, a single pull based nondirectional coupler scheme for cones has been proposed [15].The experimental results show that this structure can increase the process tolerance to more than 20 nm, while the traditional asymmetric directional coupler is only a few nm.In addition, in order to be compatible with WDM technology, harmonic vibration microring is introduced into the asymmetric coupler [16], as shown in figure 8.This structure is similar in implementing MDM.It also has the function of wavelength selection.In addition to microring resonators, grating structures can also be combined with asymmetric couplers.Current mode multiplexing also has the functions of bandpass filtering and reverse coupling [17][18].The adiabatic coupler based on mode evolution avoids the strict phase matching conditions in the whole coupling region.It is necessary to match phase at some point in the coupling process, which greatly increases the process tolerance [19].Mode excitation in adiabatic couplers is one-to-one, and unwanted modes are not excited.This means that only the coupling length needs to be long enough to achieve perfect mode conversion efficiency.So, mode multipelxer based on adiabatic coupling has received extensive attention.
In addition to adiabatic couplers, asymmetric Y branches are also representative structures based on the principle of pattern evolution [20].The structure has the advantage of high scalability.Figure 9 shows a 4-way based on a cascade asymmetric Y branch Channel pattern multiplexer.The disadvantage of this structure is that very precise machining accuracy is required to obtain low loss Y fraction Branch.The weakness of devices based on the mode evolution principle is that they are usually very long (>100 μm).So one is called adiabatic A shortcut method has been proposed to optimize the structure of adiabatic devices [21][22].Reduce the size of the device, and still maintain the advantages of large bandwidth and large process tolerance.This approach opens up a path to designing small size and high reliability photon integration.

Discussion
Compared with traditional optical switches, silicon optical switches and arrays are relatively late to start, but because of their advantages such as small size, low power consumption and CMOS process compatibility, they have gained wide attention in research and application.One of the main problems faced by silicon optical switch is that crosstalk affects the array scale.The other is that the transmission loss is relatively large.So it is necessary to enhance the process tolerance, improve the performance of the switching unit, and further optimize the topology structure to achieve effective expansion of the array performance and scale.In addition, the power consumption of the driving circuit is also one of the urgent problems in the development of silicon optical switch.With the increase of the number of switch ports, the scale of the external drive circuit also increases geometrically, which means that the reliability of large-scale silicon optical switch array chip size and heat dissipation should be considered.Based on the electrode design, the power consumption of switching units can be reduced by push-pull drive, and a new topology structure can be constructed to optimize the optical transmission path.Three-dimensional cross-connecting technology can be adopted to reduce the number of switching units and reduce the overall power consumption.In the future, the way of combining low-dimensional materials or phase change materials with silicon waveguides can be explored to break through the bottleneck in the development of the current integrated optical path on the premise of ensuring the compatibility of CMOS technology, and realize the optical switch chip that can meet the needs of information exchange in the post-Moore era.
The on-chip multiplexing system based on silicon photonic devices has many advantages, but there are still several major problems that need to be solved.Silicon-based photonic devices can be divided into thick silicon optical devices and thin silicon optical devices according to the thickness of silicon core layer.The thickness of the thick silicon core layer is in the order of several pm, which can achieve an approximate square waveguide cross section and has the characteristics of polarization insensitivity.However, the large size of such devices makes it difficult to integrate optical devices on a large scale.Business flow platform mainly adopts thin silicon wafer.At 1550 nm band, the refractive indexes of silicon core layer and silica cladding refractive are 3.47 and 1.44.The waveguide has high refractive index difference, which can generate the strong restriction, light field to decrease the size of the device.But it also brought strong waveguide dispersion effect, the response of silicon-based photonic devices limited bandwidth.It is difficult to make large bandwidth silicon optical devices.In addition, the asymmetric structure of silicon wafer causes a large refractive index difference between the quasi-Tbase mode and the quasi-TM-base mold, which results in polarization sensitivity.Polarization sensitivity leads to the coupling loss between optical chip and optical fiber mode and affects the receiving signalto-noise ratio of detector, which is difficult to be integrated into traditional optical network units.

Conclusion
In this paper, silicon-based integrated optical switching technology and multiplexing technology applied for datacenter optical interconnection are investigated.First, the principle of basic optical switching unit is discussed, including MZI optical switch and MRR optical switch.On this basis, blocking architecture, Reconfigurable non-blocking architecture and Strict non-blocking architecture are studied.Next, onchip WDM multiplexing/demultiplexing devices are discussed, including AWG and MRR filter.Then, different kinds of on-chip MDM multiplexing/demultiplexing devices are studied.Finally, the future of silicon-based integrated photonic technology is prospected.

2. 1 .
Silicon-based optical switching unit 2.1.1.MZI optical switch.MZI structure is a common basic structure in integrated optoelectronic devices.It has the advantages of simple structure, large working bandwidth and good thermal stability, and is widely used in modulator and optical switch devices.The optical switch structure based on MZI is shown in figure1.The switch unit consists of two 2×2 splitters and two modulation arms.The common optical splitters are usually directional coupling type (DC) or multi-mode interferometer (MMI) structures, and the modulation arms can be divided into thermos-optical type and electro-optical type[1][2], which are based on the plasma dispersion effect and thermos-optical effect, respectively.

Figure 1 .
Figure 1.The structure of MZI optical switch.The switching time of MZI electro-optical switch based on plasma dispersion effect is usually in the order of nanosecond and has a large bandwidth.In the post-Moore era, reducing chip size is a crucial development direction.With the helical waveguide structure, the original device size can be further reduced without reducing the length of the modulation arm.In addition to the size of the device, the operating bandwidth also plays a decisive role.Larger bandwidth, more stable optical ratio means that a wider range of conditions can be met.In 2017, Wang et al. proposed a wideband 2×2 electro-optical switch based on MZI and reverse symmetric DC[3].By replacing the conventional coupler composed of uniform waveguide with reverse symmetric coupler, 50% spectral ratio can be achieved in a wide wavelength range.At 1550nm operating wavelength, when the driving voltage is 6V, the extinction ratio is greater than 15dB, the bandwidth is 85nm, and the insertion loss of all channels is less than 8.5dB.Due to the absorption of optical signals by plasma dispersion effect, it has an adverse effect on crosstalk.Compared with silicon-based electro-optical switch, the response speed of silicon-based optical switch based on thermal-optical effect is only in the order of microseconds, but the influence of thermal tuning on loss and crosstalk is relatively small.In recent years, with the continuous optimization of manufacturing process and structure, the performance of silicon based thermos-optic switch has been further improved.

Figure 2 .
Figure 2. The structure of MRR optical switch.In the actual design and preparation process, MRR optical switch faces two problems to be solved.One is that the spectral characteristics of the microring resonator limit the bandwidth, which is not conducive to the suppression of cross-talk between channels.Second, the preparation of the microring resonator has high technological requirements.The physical size and structural accuracy should effectively meet the resonant conditions, and minimize the influence of external temperature changes and wavelength drift of incident light on the MRR output.In order to increase the working bandwidth, a microring cascade structure can be used to achieve spectrum flattening and continuous tuning.Due to its own characteristics, MRR optical switch will inevitably cause large crosstalk, while MZI optical switch can regulate crosstalk well, but its size and loss are relatively large.In order to further enhance the silicon-based optical switch performance and reduce the size, Lu proposed a DR-MZI silicon-based optical switch structure[6].The advantage of this scheme combines resonance and interference method, realized the compared to using MZI alone or MRR structural lower crosstalk and loss, effectively reduce the power consumption of the device and size.It provides a new idea and direction for constructing high performance large-scale switch array.

Figure 3 .
Figure 3.The schematic of Butterfly architecture.

Figure 5 .
Figure 5.The schematic of strict non-blocking architecture.The total number of switch units in the network of port Crossbar is 2 , and the number of switches in each path is 2 − 1, with no light cross.The main characteristics of the Crossbar network architecture are its strict non-blocking characteristics and zero waveguide crossing.However, each path of the Crossbar structure passes through different number of switch units, that is, the length of the optical signal path is different.The shortest path passes through only one switch unit, while the longest path needs to pass through 2 − 1 switch units.Therefore, different paths lead to a large difference in loss, thus affecting the loss uniformity among the array switch paths.Although the total number of switch units in the PILoss network is  2 , the number of switches on each path is N, and the number of crossing nodes on each path is  − 1.Compared with Crossbar networks, PILoss architecture has better path loss uniformity.Loss is path independent, but the network structure is relatively complex.The total number of switches in the N-port S&S network architecture is 2 2 − 2, the number of switches in each path is 2 log 2 , and the number of fork nodes in each path is ( − 1) 2 .Each path of the S&S network has a completely independent optical link, and there is no overlap between paths except the cross waveguide.Therefore, crosstalk between paths of the S&S network is extremely low.However, with the increase of the number of ports, the complexity of the network increases rapidly, and the number of crossing nodes of each path increases rapidly with  2 , and the loss and crosstalk will deteriorate accordingly, which is difficult to achieve.From the perspective of practical application, reconfigurable non-blocking network can meet most application requirements.With the increase of the number of ports, the number of switches is relatively small, typical Benes network architecture has many applications in large-scale array optical switches.In 2019, Suzuki et al of AIST in Japan reported a 32×32 strictly non-blocking optical switch array based on silicon thermo optical effect[9], with a switching time of about 10us.The unit switch adopts DC coupler and realizes fiber -fiber loss of 10dB through LC converter.In the same year, the team reported an optical switch array based on Si/SiN double-layer waveguides.In this scheme, 32×32 3D optical switching array is realized by using the free ports in PILoss topological architecture, and the polarization dependence of silicon waveguide is overcome by SiN/Si double-layer waveguide and crossover structure.

Figure 7 .
Figure 7. Mode multiplexer of based on non-symmetrical directional coupler.

Figure 9 .
Figure 9. Mode multiplexer based on asymmetric Y Branch.