1. Introduction

High-speed interconnects are now highly demanded in data centers and high-performance computing and optical technologies have been proposed and widely studied to overcome the low-speed and high electromagnetic interference bottlenecks of electrical interconnects [1–3]. For chip-scale and inter-chip interconnects, silicon photonics technology has been investigated and integrated optical interconnect transceivers have recently been demonstrated [4, 5]. On the other hand, for card-to-card interconnects, conventionally electrical cables have been utilized. However, electrical cables have several fundamental limitations for high-speed operation, including limited bandwidth, heat dissipation, electric power consumption, high transmission loss and latency, as well as electromagnetic interference [3].

The use of short-range parallel optical links to realize high-speed card-to-card interconnect has been widely studied, including both multi-mode fiber ribbons and polymer waveguides based architectures [6, 7]. However, both kinds of schemes are inherently point-to-point and non-reconfigurable, so they are not suitable for dynamically interconnected electronic cards that require flexibility for efficient operation. Several reconfigurable card-to-card optical interconnect architectures using free-space signal propagation have been proposed and investigated [8, 9]. In these architectures, the modulated optical beam directly propagates in the free-space to the final destination card and since no pre-determined waveguide is utilized, the optical signal can be switched along different directions in the air with a link-selection block based on either liquid crystal on silicon or Opto-VLSI processors. However, both link selecting techniques are based on signal diffractions, which are comparatively complicated and have low efficiency for large tuning angles due to the use of high-order diffraction signals. To overcome these limitations, in previous studies we have proposed and experimentally demonstrated a 3 × 10 Gb/s reconfigurable optical interconnect architecture based on MEMS steering mirrors, which have simple link reconfiguration mechanism through just reflection and high link selection efficiency through proper coating [10]. While this scheme is simple and cost-effective, it cannot provide the broadcasting capability, which is typically required for control purposes and parallel computing.

In this paper, we show that by modifying the previously proposed optical interconnect architecture, additional broadcasting function can be easily implemented, making the optical interconnect architecture more practical [11, 12]. Experimental results show that 10 Gb/s data can be successfully broadcast to all channels with a worst-case receiver sensitivity better than −12.20 dBm. For multicasting, 10 Gb/s reconfigurable point-to-point link and multicasting to multiple arbitrary-selected receivers are achieved simultaneously and compared with the broadcast scenario, demonstrating a power penalty in receiver sensitivity of ~1.3 dB.

2. Proposed optical interconnect architecture with broadcast capability

The architecture of proposed free-space based reconfigurable card-to-card optical interconnect with broadcast capability is shown in Fig. 1, which is similar to that reported in [10]. A dedicated optical interconnect module is proposed to be integrated onto each electronic card (typically a PCB) and inside the module a VCSEL array is used in conjunction with a collimating lens array to generate digitally-modulated collimated Gaussian optical beams. A MEMS-based transmitter mirror array is employed to adaptively steer the optical beams to various destinations, thus providing interconnect reconfigurablity and operation flexibility. After propagating in free-space, at the receiver side another receiver MEMS mirror array is used to appropriately steer the modulated optical signals to focus them onto the corresponding PD elements. To realize broadcasting, an additional large-size MEMS steering mirror (1.5 mm diameter here and the size is larger than the size of micro-lens array) is employed at the receiver side, which guides the signal to the center of the receiver micro-lens array. Due to the Gaussian beam divergence, the beam footprint becomes larger after free-space propagation and all receiver elements are illuminated, hence, the data modulating a VCSEL beam can be broadcast to all channels.

 

Fig. 1 Architecture of the proposed reconfigurable card-to-card optical interconnect with both broadcast and multicast capabilities.

Download Full Size | PDF

The proposed reconfigurable optical interconnect architecture with broadcast capability can also be used to realize multicasting, where selected receiver mirrors are appropriately steered to guide the signal to corresponding PD elements. While a VCSEL element is operating in the multicast mode, i.e., supplying data to specific receiver elements, reconfigurable point-to-point interconnects with the other receiver elements can still be realized simultaneously with minimal crosstalk, since steering angles for the PD elements operating in multicast modes are different than those of the PD elements operating in point-to-point interconnection modes. In addition, the data can be selectively multicast to receivers both in adjacent or separated. For the non-adjacent case, the signal is directed to the middle of receiving MEMS mirrors requesting multicasting and the crosstalk is still minimal even if a point-to-point link in between is in operation, because the various optical interconnects operate at their optimal steering angles.

3. Experiments and discussions

Experiments were carried out to demonstrate the proposed reconfigurable optical interconnect architecture with broadcast capability using the setup shown in Fig. 2. A PCB-based optical interconnect module was designed, fabricated and assembled (as shown in the inset of Fig. 2 [10]). A 1 × 4 VCSEL array (~17° divergent angle, 850 nm, 250 μm pitch, and 1 mm overall dimension), the corresponding driver circuits (4 QFN packaged drivers), a 1 × 4 PD array (60 μm active diameter, 250 μm pitch, ~0.6 A/W responsivity at 850 nm, and 1 mm overall dimension), and 4 trans-impedance amplifier (TIA) chips were assembled onto a single small-size PCB. A micro-lens array with ~236 μm clear aperture and 250 μm pitch was then aligned and mounted on top of the VCSEL and the PD arrays to collimate the radiated VCSEL beams and to focus the received optical beams onto the active windows of corresponding PD elements. Each of the micro-lens arrays was attached to a 3-axis translational stage, and the distance between the VCSEL/PD plane and the lenses was manually changed (the distance was equal to the focal length). Separate MEMS mirror chips were used to switch the optical beams to various ports or cards due to the device limitation. The MEMS mirror chips were also attached to 3-axis translational stages and they were steered by changing the applied voltage. In addition, as mentioned in the previous section, another larger MEMS mirror with ~1.5 mm diameter was employed for the broadcast function.

 

Fig. 2 Experimental setup (not to scale) for demonstrating the proposed reconfigurable optical interconnect architecture with both broadcast and multicast capabilities.

Download Full Size | PDF

To demonstrate the broadcast capability of the proposed reconfigurable optical interconnect architecture, only one VCSEL was turned on and modulated. The modulated VCSEL beam was transmitted towards the large-size broadcast MEMS mirror for broadcasting. The beam reflected off the broadcasting mirror was then steered to the center of the receiving micro-lens array and after focusing, the optical signal for each receiver channel was detected. With VCSEL 1 being used as the broadcasting source (10 Gb/s on-off-keying modulation and 3 mW transmission power), the measured bit-error-rate (BER) performances of all four PD elements with respect to the horizontal distance between the transmitter and receiver modules are shown in Fig. 3(a). It should be noted that in the experiments there was not lateral displacement between the transmitter and receiver modules, while in practical applications, the lateral displacement is required due to the non-blockage consideration, as shown in Fig. 1. If lateral displacement takes place, the BER performance of proposed system will be worse due to the longer signal propagation distance and the lower power collected by the receiver. It is clear from Fig. 3(a) that receivers 1 and 4 (or 2 and 3) have similar BER performances. This is mainly because both receivers have the same distance to the Gaussian beam center, resulting in almost the same optical signal power being received by both receivers. In addition, receivers 2 and 3 always performed better than receivers 1 and 4, and this can be attributed to the fact that receivers 2 and 3 are closer to the Gaussian beam center. When VCSEL 2 was turned on for broadcasting, the BER performances of the four receivers are shown in Fig. 3(b). Compared with the results shown in Fig. 3(a), it can be seen that the performances of all four channels follow the same trend, although the BER performances in Fig. 3(b) are slightly worse than those displayed in Fig. 3(a). This is mainly due to that VCSELs 1 and 2 have different divergence angles, resulting in different beam footprints after propagating in free-space.

 

Fig. 3 BER versus horizontal distance. (a) VCSEL 1 served as transmitter; and (b) VCSEL 2 served as transmitter (reprinted from [12]).

Download Full Size | PDF

The receiver sensitivities of the broadcast function were also measured, and the BER results are shown in Fig. 4. Here, VCSEL 1 served as the broadcasting source and the horizontal distance between the transmitter and receiver modules was fixed at 30 cm. It is clear from Fig. 4 that all channels have similar receiver sensitivity (~-12.20 dBm). Compared with previous results of reconfigurable point-to-point data transmission [10], better receiver sensitivity is achieved for the broadcast function since there is no inter-channel crosstalk. When the horizontal distance between the transmitter and receiver was reduced to 20 cm, similar receiver sensitivities were achieved. In addition, it should be noted that the receiver sensitivity is defined at the BER of 10⁻⁹. While such a BER is not sufficient for typical card-to-card optical interconnects applications, the BER performance of our proposed architecture is not limited to 10⁻⁹ and better performance can be realized, as evident from Fig. 3 and Fig. 4. Furthermore, if forward-error-correction codes (FEC) is employed, the BER performance can further be improved at the cost of some overhead (7% or 20%) can be achieved [13].

 

Fig. 4 Receiver sensitivity for the broadcast function. VCSEL 1 served as the source and the bit rate was 10 Gb/s (reprinted from [12]).

Download Full Size | PDF

In addition to the broadcast function demonstration where the signal was distributed to all receiver channels, arbitrary multicasting was also experimentally investigated. Here VCSEL 1 served as the multicast signal source (10 Gb/s and 2 mW power) and the modulated signal was multicast, for proof of concept, to PD elements 1 and 2. VCSEL 4 (10 Gb/s and 2 mW power) was also turned on for demonstrating the capability of the proposed architecture to simultaneously realize reconfigurable point-to-point interconnects (PD element 4 was used for signal detection). For the multicast function, MEMS mirrors 1 and 2 at the receiver side were appropriately steered to guide the signal to the corresponding PD elements. The measured BER performances of these PD elements with respect to the horizontal distance between the transmitter and receiver modules are shown in Fig. 5(a). It is clear that channel 4 has better BER performance than the multicast channels. This is mainly because that the multicast signal needs to cover two receivers, leading to a smaller received power.

 

Fig. 5 BER versus horizontal distance. (a) VCSEL 1 multicast signal to receivers 1 and 2; and (b) VCSEL 1 multicast signal to receivers 2 and 4 (reprinted from [11]).

Download Full Size | PDF

To demonstrate the reconfigurability and flexibility of the proposed architecture, a second scenario was considered, where the modulated optical beam of VCSEL 1 was multicast to receivers 2 and 4 and VCSEL 2 was point-to-point interconnected to receiver 1. The BER performances in this scenario are shown in Fig. 5(b). It can be seen that similar to the results shown in Fig. 5(a), the point-to-point interconnect channel still has better BER performance than the multicast channels.

The receiver sensitivities in the selective multicasting scenario were also measured and the results are shown in Fig. 6. Here, VCSEL 1 served as the multicast signal source towards receivers 1 and 2 and VCSEL 4 was used for the point-to-point interconnects. The horizontal distance between the transmitter and receiver modules was fixed at 20 cm. It is clear that channel 4 exhibits the best sensitivity while the sensitivity for channel 2 is the worst. This is mainly because receiver 4 is further away from the other channels, resulting in smaller inter-channel crosstalk. In addition, compared with the results shown in Fig. 4, the receiver sensitivity is ~1.3 dB. This power penalty can be attributed mainly to inter-channel crosstalk.

 

Fig. 6 Receiver sensitivity in the multicast scenario. VCSEL 1 multicast signal to receivers 1 and 2 and bit rate was 10 Gb/s (reprinted from [11]).

Download Full Size | PDF

Finally, it should be noted that in the experiments, the maximum horizontal distance between the transmitter and receiver modules measured was limited to 30 cm. This is mainly due to the use of a 250 µm pitch micro-lens for signal collection at the receiver side. For a longer free-space signal propagation distance, the signal power that can be collected drops significantly because of the Gaussian beam divergence. Although this free-space distance may be not enough for some practical applications, as discussed in [10], by using larger focusing lenses at the receiver side, much longer interconnection range can be realized.

4. Conclusion

In this paper, a novel free-space based reconfigurable card-to-card optical interconnect architecture with broadcast capability has been proposed and experimentally demonstrated. The broadcast function has been achieved by adding one dedicated larger-size MEMS mirror at the receiver side and by using the Gaussian beam divergence to illuminate all receivers. 10 Gb/s data has been successfully broadcast to all channels with a worst-case receiver sensitivity better than −12.20 dBm. In addition, it has been shown that multicasting can be achieved with the same system architecture. 10 Gb/s reconfigurable point-to-point and multicast interconnection to selected receivers have been experimentally realized simultaneously. Compared with broadcast scenarios, the power penalty in receiver sensitivity has been shown to be ~1.3 dB for simultaneous point-to-point and multicast interconnects.

In should be noted that for free-space based reconfigurable card-to-card optical interconnects, dust accumulation, mechanical vibrations, as well as atmospheric turbulence affect the system performance and the robustness. The turbulence has been shown to result in ~1 dB power penalty [14] and other impacts require further study. In addition, the initial installation of optical interconnect modules requires high-accurate alignment. For the architecture proposed in this paper, the micro-lens arrays can be assembled with the VCSEL/PD arrays with a spacer and the major challenge is aligning the MEMS mirrors. The impact of misalignment and the required installation alignment accuracy also require further consideration.