1. Introduction

Electronic device scaling in modern microsystems has led to system limitations mainly driven by challenges in signal communication, but also fundamental challenges such as quantum tunneling [1-2]. Although new emerging options including photonics and plasmonics have recently been investigated to overcome the interconnect and communication bottleneck, the fundamental energy efficiency of optical logic devices is still been limited to fJ level due to weak non-linearity [3]. Rather than focusing at the logic level, reducing the energy cost in data communication via both optoelectronic component- and architecture improvements is a viable option [4-5]. While nanophotonic and plasmonic building blocks show performance and integration density promise, metal-optics is unlikely to become a meaningful implementation for optical communication links due to losses. However, separating the function of communication and light manipulation to passive photonics such as Silicon photonics, and plasmonic, respectively, shows superior intra-chip capabilities in terms latency, energy efficiency and packing density [6]. Such hetero-integration shows up to 300% energy improvement and an order of magnitude footprint reduction over traditional photonic networks [7–9]. However, despite these promises, the point-to-point communication characteristic of photonic heterogeneous integration schemes can be restrictive in realizing on-chip interconnect networks for future many-core chips with hundreds of cores. Conventional photonics, on the other hand, can achieve the required connectivity through a bus structure, relying on wavelength division multiplexing (WDM) to support multiple links on such a shared physical communication medium [10]. However, this is achieved at the cost of large area overheads from multiplexers and microring resonators (MRRs) that aid modulation and detection, as well as substantially higher power consumption attributed to the thermal trimming of the MRRs [7]. Hybrid Photonic-Plasmonic Interconnect (HyPPI) based on our previous demonstration, on the other hand, uses photonic plasmonic hybrid modulators directly integrated onto the photonic transmission waveguide [6]. Although this is beneficial in saving footprint and enhancing point-to-point performance, the on-bus modulator introduces superfluous insertion losses at the level of 1 dB/μm even while inactive, rendering the bus unusable for long-distance communications. An off-bus modulator is therefore essential. Similarly, on the receiver side, an off-bus detector is desirable if coupling to the bus can be actively (de)selected. While this could be achieved with MRRs, energy and latency overheads due to the electrical thermal tuning, footprint, and fabrication repeatability would limit the MRR-based approaches in terms of speed, energy efficiency, as well as the packing density. An alternative option of achieving point-to-point connectivity is to connect each core to all the other cores on the chip directly which is however not practical; for instance, a 64-core processor requires a total of 4,032 unidirectional HyPPI links to achieve full connectivity which is several orders of magnitude footprint consuming than the mesh topology [11-12].

To overcome the limitation of point-to-point connectivity while retaining high performance of all-photonic routing, modulation, and detection, we here show a novel photonic-plasmonic hybrid device termed MODetector (hereinafter referred to as ‘MOD’) that combines both the modulation and the detection functions for inter-chip communication. Different from the regular photonic or plasmonic enhanced (e.g. HyPPI) interconnect, where the modulators and detectors are integrated onto the waveguide bus, this MOD-based HyPPI link allows data to bypass photonic devices without causing much energy penalties [6]. Thus, from a network point of view, we re-allocate some of the routings onto the device level by combining the electro-optic modulation, detection, and optical routing into one single structure. The remainder of the paper is organized as follows: 1) MOD structure introduction. 2) Operation principle discussion. 4) Modulation and detection analysis. 3) Network comparison between classic point-to-point (P2P) optical links and MOD-enabled links.

2. On-chip optical transceiver design

2.1 MOD structure

The goal of MOD is to separate the light modulation and detection from the main bus in order to avoid the unnecessary conversions between electrical and optical domains which leads to extra losses. Since there is no conflict in separating either the modulator from the bus or the detector from the bus, we find positive synergies when both functionalities are combined into a single device as discussed here. Moreover, for network topologies like mesh, ring, and bus, some of the cores require bi-directional communication from both directions of the bus, which requires the MOD design to be symmetric. Based on these requirements, we consider a racetrack ring-based MOD structure that integrates an ‘expanded’ germanium photodetector on the ring via a 2 × 2 hybrid plasmonic 3-waveguide switch to provide modulation functionality in Fig. 1(a) [13]. The 2 × 2 switch consists of a central switching island containing a highly optical index changeable material (indium tin oxide (ITO)) ‘sandwiched’ between two gate oxide layers (SiO₂) to form a metal-oxide-ITO-oxide-semiconductor capacitive heterostructure; whereas the detector has a germanium block on top of the racetrack waveguide with that part of the silicon etched down to 100 nm for better light mode overlap with the high absorption region in Fig. 1(b).

 

Fig. 1 Schematic of the MODetector concept. a) 3D overview of MOD with the ITO hybrid switch on the left and Ge photodetector on the right. b) The cross-section of MOD at A plane. Both a) and b) are color-coded and sharing the same legend on the top-right. All the parameters are optimized for the highest coupling efficiency [17].

Download Full Size | PDF

2.2 Operating principles

In order to obtain the bi-functionality of an optical transceiver (i.e. encoding, detection), a bias voltage is applied to MOD at the switch, and detector depending on the desired function in Fig. 2; configuring the switch in the Bar states encodes a logic ‘1’ onto the downstream bus for an unmodulated light beam arriving at MOD represented in Fig. 2(d). At the same time, a modulated signal arriving at MOD can be captured at the detector when the switch is in the Cross state in Fig. 2(e). In this way, each MOD node in the NoC can act as either as a transmitter or receiver depending on the system’s demands.

 

Fig. 2 The switch analysis at the Cross and the Bar states. a) The top view of the MOD with the same color coding as Fig. 1. b) Fundamental TM mode effective indices change of the 3-waveguide switch at the cross-section (BB’) based on ITO carrier concentrations. c)-f) The FDTD simulations of all four functionalities at different switch and detector state combinations: c) switch OFF, detector OFF; d) switch ON, detector OFF; e) switch OFF, detector ON; f) switch OFF, detector ON. All simulations are based on 1550 nm light source. The ITO refractive indices are calculated based on the Drude model. V_bias = V_dd = 4V. Note, the MODetector is simulated in 3D using Lumerical FDTD software as a complete device.

Download Full Size | PDF

For regular encoding operation, this dual functionality is used nontime-concurrently. Interestingly, a time concurrent operation of MOD creates a copy of the data signal, which may have relevance for cyber security applications, yet, this case is not considered in this work. Placing the switch into the Cross state (0V bias voltage) allows dropping the optical signal from the bus to the racetrack ring, which enables three operation modes, namely modulate ‘0’, detect ‘0’, and detect ‘1’ shown in Figs. 2(c), 2(e), and 2(f). For both detection modes, the detector must always be ON in order to generate photocurrent. On the other hand, for both modulation modes, the detector needs to be OFF in order to avoid any false photocurrent at the modulate ‘0’ state as well as saving the energy. Note, the light will still be absorbed by the detector in this case. For regular operation, independent biasing is required, which eliminates the need for coordination logic-circuitry.

For the Bar state, the refractive index of ITO layer shifts from its dielectric region to the metallic region (e.g. epsilon-near-zero, ENZ) due to the carrier concentration change, and thus yields a new coupling length that prohibits the light transverse through the 3-waveguide switch [13-14]. However, this switch could not enable modulating by itself, since the racetrack ring couples any non-detected photons back to the bus which might affect the output. Note, the symmetric shape of the racetrack ring enables bi-directional communication in the network which improves the link utilization and the footprint efficiency.

3. MOD function analysis

3.1 Light modulation

The coupling efficiency of MOD critically determines the overall performance of the 2 × 2 switch. With a higher switch efficiency, the Ge detector’s responsivity improves. In addition, a higher extinction ratio at the bus port downstream could be obtained as well. The design of MOD deploying the hybrid photon-plasmon mode is driven by two main factors; first, its compact size (9.5 μm) is enabled by enhanced light-matter-interactions (LMI) primarily at the hybrid plasmon center island allowing the device length to be much shorter than conventional photonic switch, which is usually in the range of few hundreds of micrometers [15]. Secondly, it combines both sub-λ confinement and long silicon-on-insulator (SOI) waveguide propagation length, which reduces the insertion loss down to 0.12 dB/μm and 0.08 dB/ μm for the Cross and the Bar states respectively, compared to all-plasmonic designs [16]. The fundamental principle of operating this switch is to use bias voltages (i.e. ITO carrier concentration) to control the coupling length (CL) for the different states as shown in Figs. 2(c) and 2(d). The effective mode indices of the first three fundamental TM supermodes (TM₁, TM₂, TM₃) of the switch are used to determine the actual coupling length and the coupling efficiency based on Eq. (1) and Eq. (2) where light source λ is 1550 nm in this case [17]. We sweep the ITO carrier concentration from 10¹⁹ to 10²¹ cm⁻³ where 10¹⁹ cm⁻³ can be regarded as zero bias voltage point in Fig. 2(b). With a higher voltage applied, the effective index difference between TM₁ and TM₃ decreases resulting in a CL longer than the actual physical length of the switch, which forces the light to stay in the bus. The 4V bias voltage (carrier concentration of 6.8 × 10²⁰ cm⁻³) was experimentally obtained in previous work to surpass ITO’s epsilon-near-zero (ENZ) point which shows unity-strong index modulation. We use + 4V for the BAR state, and zero bias for the Cross state of the switch [14, 16]. Based on the capacitance of the switch, 24 fJ energy is needed to charge the ITO layer from Cross state to Bar state. And over 100 GHz theoretical switching speed is calculated by the RC delay of the device (where the resistance R_c = 500 Ω is assumed) [14].

3.2 Photodetection

The second part of MOD is a photodetector, here a Germanium (Ge) p-i-n based detector delivering two functions; a) it is a regular photodetector, and b) it is a ‘light absorber’ to prevent the residual optical energy from coupling back to the bus waveguide when MOD is used in the modulate ‘0’ mode. In order to achieve high responsivity with CMOS compatible material, we select a metal-semiconductor-metal (MSM) design with Germanium integrated on the SOI waveguide. The choice for Germanium is natural given by its high absorption (i.e. high κ) and compatibility with silicon photonics harnessing the mature fabrication process [18]. Germanium parameters used are based on the high-temperature-growth process resulting in absorption coefficient ~5000 cm⁻¹ [19]. However, based on the TM mode required for the 2 × 2 switch, the responsivity we obtain from the detector is low when the Germanium is simply placed on the waveguide due to the small mode overlap. Thus, we consider a silicon waveguide etch down to be just 100 nm thick in the straight part of the racetrack and then backfill it with the 240 nm Germanium shown in Fig. 1(b).

Two gold contacts are placed on both sides of the waveguide along the light propagation direction each with 400 nm width aligned to the outer side of the waveguide. We noticed that the lossy light absorption (α_Au) increases with the metal contact width (while keeping the doping areas constant in the Germanium region) resulting in a higher response speed, however, the metal loss also reduces the responsivity due to the reduction of incident light onto the Germanium (α_Ge). To overcome this problem, we optimize the absorption ratio (α_Ge/α_Au) by widening the Germanium detection region to 1500 nm. In this way, we can address the trade-off between the operating speed and the responsivity by adjusting the metal contact width in Fig. 3(b). Nevertheless, the 100 nm contact width provides the highest responsivity and a detector speed of about 20 GHz, which is limited by the average distance of the electron-hole pair free path, which is mainly determined by the distance between the metal contacts with a constant doping level. In addition, not only the contact width but also the carrier generation rate profile affects the detector speed since electron-hole pairs generated at different positions inside the cross-section of the Si-Ge waveguide vary with the carrier-to-contact distance for charge collection. For example, carriers generated at the center of the Germanium have a relatively short distance to the contact, as compared to carriers generated near the contact in the insets of Fig. 3(b). Thus, a 400 nm contact width with more free carriers generated from the center (i.e. shorter free path distance) achieves an optimized operating speed of 28 GHz with relatively high responsivity (0.7 A/W) and is selected for our detector design.

 

Fig. 3 Detector performance analysis of MOD. a) The cross section FDTD simulated electrical field of the detector part CC’ in Fig. 2(a) where light enters from the left side and propagates to the right side. b) The trade-offs between the metal contacts width and the speed, responsivity as well as the light leakages after the detection region. The insets are the generation rate at the same CC’ cross section. c) The illuminated and the dark current of the detector at different bias voltages.

Download Full Size | PDF

Note, we take the detector speed as the reciprocal of the response time from 10% to 90% of the saturated photocurrent. While a longer detection region could provide higher responsivity with less light leakage, we aim for a detector length (8.5 µm) similar to that of the 2 × 2 switch to avoid any additional footprint. As mentioned above, the second role of detector region is to prevent the light leakage back to the waveguide bus for modulate and detect ‘0’ that returns to the bus waveguide. Although this detector could not absorb all light for a 400 nm contact width, the 2 × 2 switch also creates 1.2 dB insertion loss per coupling. Thereby, only 2.8% of the light will leak to the bus at the Cross state, together with another 5% light leakage caused by the unperfected switching, yields over 10 dB extinction ratio together with 83% light bypassing the MOD at the Bar state.

Regarding the detector performance, the photocurrent under 0.5V reverse bias is about 0.35 mA under 0.5 mW input light power giving a 0.7 A/W responsivity equivalently while the dark current is in the sub-μA level shown in Fig. 3(c). Indeed improving the responsivity can be done in multiple ways to include 1) reducing the interface between germanium and gold such as cylindrical metal contacts; 2) using a wider rib waveguide, or 3) increasing the absorption of germanium by heating [20–22]. Comparing our optimized detector performance with commercial foundry components, we find a similar matching performance with, for instance, IMECs medium-speed photodetectors process [22]. In addition, if a higher output current is required a trans-impedance amplifier could be used, yet noise will also be amplified, so the SNR may not improve during amplification.

4. Towards ‘off-bus communication’

Improving the link-energy efficiency has become a more effective and reliable way of improving the power efficiency on-chip, than reducing the energy consumption of the logic gates due to the inherent static power dissipation arising from quantum tunneling across thin oxides [5]. In recent optical network on chip (NoC) studies, routing protocols based on circuit switching are used as the ‘express lines’ for high speed, long distance communication [7–11]. In such systems, the setup of the physical light path for circuit switching is typically achieved by spectrally sensitive devices such as MRRs [23-24]. For certain types of the network topologies such as the mesh and the ring, the ability to realize an all-to-all core communication in a power- and latency efficient way is highly desirable, and we find that MOD enables high NoC performance, as discussed next in Fig. 4.

 

Fig. 4 The optical 3 × 3 network topology comparison among: a) the P2P-based mesh network; b) the P2P-based all-to-all network and c) the MOD enabled ring network.

Download Full Size | PDF

Traditional ring-based optical NoCs could achieve all-to-all core connectivity by connecting MRRs to a common optical bus, and thermally tuning them to a selected resonance, such interconnection paradigms require a long setup time up to millisecond level depending on the thermal tuning power and the quality (Q) factor of the ring. The ring’s average energy consumption can be as high as 40% of the entire interconnect’s energy consumption [12, 23].

Considering a 3x3 optical mesh network for example shown in Fig. 4, the all-to-all connection can only be achieved by 1) building a mesh network with 18 links that connect to all the adjacent cores or 2) connecting each core with all the other cores with 72 P2P links in Figs. 4(a) and 4(b). Note, since the P2P link is unidirectional, every two cores require 2 links to communicate with each other (i.e. simultaneous communication demanded). The first mesh option in Fig. 4(a) reduces the number of links needed by detecting and regenerating the signal at each core, while the second all-to-all mesh option in Fig. 4(b) saves the OEO conversions by making all connections available at the same time. This would however not only require larger on-chip footprint but more importantly would require massive routers; the number of switches needed scales with a factor of (N-1)²/2 where N is the number of ports needed for the router [24]. Thus, for our small all-to-all example network, one would require a 9 × 9 router with 32 switches, which is significantly more complex than regular size routers (e.g. 5 × 5 requiring 8 switches) used regularly in. Besides, all OEO conversions in the cores of the P2P-based mesh network and the local routing mechanism in the P2P-based all-to-all network introduce on average 12 dB and 5 dB optical loss per core, respectively, even when high-efficiency hybrid photonic-plasmonic integrated technology are considered [25]. Thus, the P2P interconnect option with both modulation and detection devices on the bus does not allow for reduced network complexity, and hence limits the scaling and is not suitable for designing high-connectivity NoCs.

Given the aforementioned drawbacks, an off-bus device that can be actively bypassed when its functionality is not demanded is could simplify NoC routing protocols, while improving bandwidth as discussed in this paper. For the same 3 × 3 network, only one single link is used that creates a joint bus to connect all 9 cores in Fig. 4(c). This off-bus device significantly simplifies the network design while maintaining every core to every core connectivity. We find that the energy overhead for this single-optical link network is as low as the energy cost by a single 2 × 2 switch per core when the signal needs to be bypassed. Furthermore, this reconfigurable MOD device can be a supplement for the Dynamic Data Driven Applications Systems (DDDAS) enabling configuration based on executing applications and real-time traffic feedback controls [26]. Since MOD is plasmonic and hence spectrally broadband, the joint bus could support multiple wavelengths (i.e. WDM) with even higher connectivity and bi-section bandwidth that communicate simultaneously if the MOD connects to a wavelength selective device (e.g. micro-ring resonator).

5. Conclusion

By integrating a hybrid photonic-plasmonic switch with a Germanium-based photodetector into one single device, we design a dual-function modulator-detector. This integrated device is able to detect optical signals up to 28 GHz and generate on-off keying signals up to 100 GHz. Based on the symmetric design, it enables bi-directional all-to-all communication between multiple communication cores with only one bus waveguide, which significantly reduces the area for inter-chip connections. The performance shows over 10 dB extinction ratio and 0.7 A/W responsivity for the modulator and detector, respectively. This dual-functional device acts an optical transceiver capable of both sending and receiving optical data signals in optical networks and communications and could potentially be used as a reconfigurable optical element in analog photonic-optical compute engines and accelerators.