This paper presents the first chip-scale demonstration of an intra-chip free-space optical interconnect (FSOI) we recently proposed. This interconnect system provides point-to-point free-space optical links between any two communication nodes, and hence constructs an all-to-all intra-chip communication fabric, which can be extended for inter-chip communications as well. Unlike electrical and other waveguide-based optical interconnects, FSOI exhibits low latency, high energy efficiency, and large bandwidth density, and hence can significantly improve the performance of future many-core chips. In this paper, we evaluate the performance of the proposed FSOI interconnect, and compare it to a waveguide-based optical interconnect with wavelength division multiplexing (WDM). It shows that the FSOI system can achieve significantly lower loss and higher energy efficiency than the WDM system, even with optimistic assumptions for the latter. A 1×1-cm2 chip prototype is fabricated on a germanium substrate with integrated photodetectors. Commercial 850-nm GaAs vertical-cavity-surface-emitting-lasers (VCSELs) and fabricated fused silica microlenses are 3-D integrated on top of the substrate. At 1.4-cm distance, the measured optical transmission loss is 5 dB, the crosstalk is less than −20 dB, and the electrical-to-electrical bandwidth is 3.3 GHz. The latter is mainly limited by the 5-GHz VCSEL.
© 2012 OSA
The performance of microprocessors continues to improve with technology scaling, especially through the increase of the number of cores. Communications within these chips, e.g. between processor cores and at the memory/processor interface, will demand larger bandwidth density, smaller latency and better signal integrity. To meet these demands, conventional electrical interconnects need better materials to minimize transmission loss, and increased circuit complexity (e.g. equalization) to achieve larger bandwidth, both of which increase energy consumption. Therefore, a fundamental change is required for the inter- and intra-chip interconnects. Optical interconnect exhibits inherent advantages in loss, delay and bandwidth compared to its electrical counterpart, and can potentially lead to significant performance gains and energy savings [1, 2]. For inter-chip communications, optical interconnects with point-to-point topologies have already been developed, typically using on-board waveguides and directly modulated lasers [3, 4].
For intra-chip communications, however, optical interconnect schemes previously proposed create new challenges: packet-switching optical interconnects require either all-optical switching, which is still difficult for silicon, or repeated electrooptic and optoelectronic conversions, which largely defeats optical interconnect’s advantages in latency and energy efficiency. Circuit-switching optical interconnect needs wavelength division multiplexing (WDM) to satisfy the bandwidth density requirement. WDM, however, requires precise optical filters (e.g. microrings) with accurate wavelength control and minimal transmission loss , which are difficult to fabricate in large-scale chips and consume significantly more power if thermal tuning is needed. In addition, all of these waveguide-based systems use an power-hungry external laser as the optical power supply, which is difficult to integrate and of high cost, especially in the WDM case.
As an alternative, free-space optics can be used to overcome some of the technical challenges of waveguide-based systems. Free-space optics has been successfully applied in board-to-board [6, 7] and inter-chip applications [8–11]. These earlier proposals use arrays of discrete vertical-cavity surface emitting lasers (VCSELs) and photodetectors (PDs), and interconnect them using free-space optics. In , diffractive mirrors and fused silica slabs were used to guide light beams. Due to the small diffraction angle of the mirrors, a large number of reflections are required within the silica slab, resulting in a long optical path even for a short distance. In [6, 8], lenses with large aperture size and focal length are shared among multiple VCSELs and photodetector arrays to increase bandwidth density. However, these macro-scale lenses are too thick to be fabricated using standard microfabrication processes, and the long throw distance results in a large link latency. Improved designs using microlenses [10, 11] has been proposed to overcome these limitations.
Recently, we proposed an intra-chip optical interconnect for future multi-core processors based on free-space optics and 3-D integrated photonic devices [12, 13]. The main objective is to construct an all-to-all communication fabric with high bandwidth density, low latency, and good energy efficiency without routing or switching. As shown in Figs. 1(a) and 1(b), this free-space optical interconnect (FSOI) system consists of a photonics layer and a free-space optical guiding medium constructed using micromirrors and microlenses, which are stacked on top of the CMOS electronics layer by 3-D integration. The light beam generated by an electrically modulated VCSEL is focused by a microlens at the backside of the GaAs substrate, similar to . Guided across the chip by the mirrors on the chip and package, it is focused by another microlens onto a PD, where it is converted into an electrical signal and then processed by the CMOS receiver. In this system, optical links are constructed directly between communicating nodes in a totally distributed fashion, without a centralized arbitration system. Therefore, it has significant signaling and networking advantages. First, because FSOI avoids packet switching and the associated intermediate routing, buffering, and arbitration delays in electrical networks or packet-switching optical networks, it can achieve extremely low latency. Second, FSOI exhibits very low propagation loss, minimal dispersion, and no bandwidth degradation with transmission distance. Third, FSOI saves a significant amount of energy by a) eliminating packet-switching related energy consumption, b) powering VCSELs down in low duty-cycle operation, and c) avoiding thermal tuning of sensitive E/O modulators in WDM systems. Fourth, FSOI’s good signal integrity simplifies CMOS transceiver electronics, e.g. implemented as a single laser driver in the transmitter and a single amplifier in the receiver. Finally, this system can also be used for inter-chip communications to form a unified optical interconnect fabric throughout the entire computer system. In this paper, we report the design, fabrication, integration and measurement results of the first chip-scale prototype to demonstrate this intra-chip FSOI system.
2. Performance evaluation
In , we calculated the energy consumption, bandwidth and latency of the proposed FSOI system. The analysis demonstrated that the FSOI system can support 10-Gb/s data rate with 0.5-pJ/bit energy efficiency for up to 3.24-cm transmission distance, which is diagonally crossing a typical 2.3x2.3-cm2 microprocessor chip. The link latency is only 115 ps, limited by the free-space propagation of the light beam. The system achieves less than 10–12 BER with a total bandwidth density of 6.25 Tbps/cm2 and total aggregate bandwidth of 10 Tbps when scaled up to 36 nodes.
In this paper, we expand the analysis to compare a WDM-based optical interconnect similar to the one in  with the proposed intra-chip FSOI system to illustrate the advantages of our approach. The operation wavelength of the WDM systems is chosen to be 1550 nm, while the FSOI system still uses 980 nm. To emphasize the effects of photonic devices and optics design, the transceiver circuits are excluded from the calculation. To simplify the calculation, the PDs in both systems are assumed to have 100% quantum efficiency and adequate bandwidth to support 10-Gbps data rate, which gives the WDM system an unfair advantage. Table 1 lists the optical and photonic device parameters for both FSOI and WDM systems.
In the FSOI case, the VCSELs are based on a design demonstrated in , except with In0.2Ga0.8As quantum wells and GaAs barriers. The optical transmission loss is calculated by adding an insertion loss of 0.1-dB for a microlens and 0.08-dB for a mirror. Gaussian beam clipping loss at the lenses is determined by beam divergence and lens aperture size. As the number of nodes increases, the size of the lenses are shrunk to satisfy that up to 50% of chip area is covered by microlenses, and hence optical clipping loss at the lenses increase from 0 at 4 nodes up to 47% at 64 nodes.
In the WDM system, the light source is an external WDM laser, which is assumed to have 20% efficiency, several times larger than the state-of-the-arts. The silicon microring modulators are based on the design in , operating in the carrier depletion mode. The microrings have 10-μm radius with zero power transmission at the through port when the resonance is tuned to 1550 nm. As a good compromise between microring insertion loss and modulation power consumption, Δλ is chosen to be 0.383 nm, leading to a 0.0437-dB insertion loss (ILmod) when the microring is off-tuned to make the waveguide transparent to optical data flux. A 0.77-nm wavelength channel spacing and a total of 13 channels in each node are selected to utilize the entire 10-nm free-spectral-range (FSR). Each channel experiences an additional 0.0662-dB loss per node (ILmod2) due to the contribution from microrings for the other 12 adjacent channels. Hence, the total modulator induced optical insertion loss per node increases to 0.11 dB. The overall link optical loss is calculated by adding 0.3-dB/cm waveguide loss , which is very optimistic, 0.05-dB bend loss, and 0.11-dB modulator insertion loss per node.
The chip has N communication nodes equally spaced in a grid. The chip size is 2.3×2.3 cm2, typical for microprocessors . The longest path is calculated by multiplying the distance between two adjacent nodes, , where L=2.3 cm is the chip dimension, with the number of nodes traveled, N – 1. In addition, there are also bends on the waveguide. Adding these two loss terms, the waveguide related optical loss for the longest path can be expressed as:Fig. 2(a). The optical loss in the WDM system increases much faster than FSOI with the number of nodes. This can be mainly attributed to the linear increase of the modulator insertion loss with N.
Next we compare the energy efficiency of a single link in both systems. Assuming that the required output current from the PDs in both systems is 150 μA, the laser/modulator optical power is calculated based on the optical loss, followed by their energy consumption. For example, the VCSEL and external WDM laser (per wavelength) need to provide 0.32 and 0.84-mW maximum optical power at 36 nodes, respectively. The average electrical power consumed for a single WDM link is 6.1 mW, corresponding to 0.61-pJ/bit at 10-Gbps data rate, including 0.183 pJ/bit to switch on/off the modulator. A VCSEL in the FSOI system consumes 0.69-mW average power, corresponding to a 0.069-pJ/bit energy efficiency. The energy efficiencies of both systems are shown in Fig. 2(b). The FSOI system performs better with N scaling, thanks to its lower optical loss.
Note that the calculations are highly optimistic for the WDM system. First, it does not include the power consumption for thermal tuning, required to accurately control the wavelengths of microrings. Second, the state-of-the-art WDM laser sources have energy efficiency of a few percentage, much worse than our assumption. Finally and more importantly, silicon microring modulators exhibits insertion loss more than 0.5 dB [16, 19], resulting in over 60-dB optical loss in a single 36-node link.
3. Design of FSOI chip prototype
As shown in Fig. 3, the chip-scale prototype of the proposed intra-chip FSOI system is constructed as a 3-D chip stack. It is designed based on the following specifications and constraints: the chip has an area of 1×1 cm2, limited by the mask writer, mask size and autostepper lithography tools. The longest optical path is 1.4 cm, diagonally crossing the chip. The Ge substrate is used to build PDs and serve as a carrier for the VCSELs and microlenses. The microlenses are fabricated on a 525-μm thick fused silica substrate. The VCSELs used in the prototype is a commercial VCSEL array (Finisar V850-2092-001S, 1x4 array) with a pitch size of 250 μm, and provides 2-mW optical power at 850-nm with a 5-GHz modulation bandwidth. The pitch size between microlenses is chosen as 250-μm, matching the pitch size of the VCSEL array. To facilitate wirebonding VCSELs and PDs to the Ge carrier, a silica spacer is inserted between the microlenses and the VCSEL/PDs. A prism is used instead of micromirrors for testing convenience.
3.1. Optics design
The optics design involves several device parameters: VCSEL aperture size/divergence angle, microlens aperture size and focal length, and device spacing, as shown in Fig. 4. The 850-nm VCSEL has an aperture size of 8 μm, and the full-width half-maximum (FWHM) far-field divergence angle is measured as 20° in free space at the operation bias point. Since there is a 200-μm height difference between the VCSEL chip and PD, the microlens aperture and focal length need to be adjusted correspondingly. The VCSELs and PDs are located 200-μm and 400-μm away from the back of the fused silica layer, or 565-μm and 765-μm from the lenses in air. The corresponding focal lengths are slightly less than these values because VCSELs and PDs need to be placed further away from the focal planes.
As discussed later in Sec. 4, microlenses are fabricated based on the photoresist melt-and-reflow technique. The desired microlens aperture size is defined by lithography, and its focal length is achieved by choosing the specific photoresist thickness. Assuming that the VCSEL is placed near the focal point of the VCSEL lens to capture more than 98% of the light, the minimal aperture size should be 200 μm. The corresponding photoresist thickness to achieve a 200-μm aperture and a 560-μm focal length lens is 10 μm. Similarly, the same photoresist thickness results in 725-μm focal length for a 220-μm aperture microlens on the PD side.
Using the designed focal length and aperture sizes, the beam waist size and its distance from the VCSEL microlens is calculated with respect to Δz, the relative position of the VCSEL to the focal point of its microlens, based on Gaussian beam propagation. As shown in Fig. 5, the beam waist distance after the VCSEL lens is larger than 7 mm, i.e., half of the longest link distance, when the VCSEL is placed between 10 μm and 40 μm away from the lens. Within this range, the beam waist size changes between 110 μm and 160 μm. Cascading it with the PD lens, the normalized optical power transmitted through the lenses for the 1.4-cm link distance is plotted with respect to Δz in Fig. 6(a), when the PD is assumed to capture all incoming light. Optical transmission peaks at Δz between 10 and 20 μm, matching that of the beam waist distance in Fig. 5.
Considering the power loss due to the finite PD size, the detected optical power is calculated as shown in Fig. 6(b). The detected power does not change for PDs beyond 45-μm PD diameter. Adding tolerance for possible integration errors, the PD size is chosen as 62×62 μm2. Next, we evaluate the optical loss with respect to its distance for different Δz. As shown in Fig. 7(a), smaller Δz leads to better transmission at longer distances primarily because of the smaller divergence angle. When Δz=8 μm, the total transmission loss from the VCSEL to the PD is approximately 2 dB for link distance up to 1.4 cm, and increases to 9.3 dB at 3.5 cm.
The impact of misalignment of the VCSEL and PD with their microlenses is also examined. As shown in Fig. 7(b), when the VCSEL is misaligned to the central axis of its microlens, the optical transmission does not change for in-plane misalignments up to 6 μm at 1-cm distance. At 1.4-cm and 2-cm distance, the transmission decreases by 3 dB for a misalignment of 4.5 μm and 2.5 μm, respectively. To keep the transmission loss below 1 dB at 1.4 cm, the misalignment needs to be less than 3 μm, which is within the accuracy of typical flip-chip bonding processes.
In order to evaluate the whole link performance, the values of the PD responsivity, VCSEL slope efficiency, and the device bandwidth are needed. Based on our prior work , an MSM Ge PD with a hydrogenated amorphous Si (a-Si:H) layer is designed and simulated in DAVINCI. The a-Si layer increases the Schottky barrier height of holes, and passivates the Ge surface states. It hence mitigates the large dark current and low frequency gain of the MSM PD. A 62×62-μm2 device with a 20-nm thick a-Si:H layer, 1.25-μm contact width, 2-μm contact spacing, and a low stress 106-nm thick Si3N4 anti-reflection coating achieves 0.37-A/W responsivity and 0.224-μA dark current at 7-V bias (Fig. 8(a)). As shown in Fig. 8(b), the simulated bandwidth is 12.1 GHz with a 80 fF device capacitance, corresponding to a 40-GHz extrinsic bandwidth when terminated with 50 Ω. The commercial VCSEL has a slope efficiency of 0.4 W/A and 2-mW optical power at 850-nm.
In a single-bit FSOI link, the electrical-to-electrical link gain is expressed as the product of the VCSEL slope efficiency (0.4 W/A), PD responsivity (0.37 A/W) and the optical transmission (−2.1 dB):12].
4. Device fabrication, characterization, and integration
4.1. Ge PD and carrier
The chip prototype is designed to support 7.5-Gbps data rate per link at 850 nm. Ge is chosen over other semiconductor materials thanks to its CMOS integration compatibility, which is crucial in large-scale electronic-photonic integrations, and its good optical and electrical properties, allowing PDs to exhibit good sensitivity and large bandwidth. A metal-semiconductor-metal (MSM) structure is chosen over p-i-n ones to reduce parasitic capacitance per area, to allow less stringent microlens-to-PD alignment for efficient light coupling, and to simplify the contact definition in the fabrication to a single step, while achieving reasonably good bandwidth and responsivity .
The fabrication procedure begins with the passivation of the Ge wafer by consecutive HCl and HF dip cycles, followed by 20-nm thick hydrogenated amorphous Si (a-Si:H) deposition, serving as the surface passivation and Schottky barrier enhancement layers in order to provide low dark current and large bandwidth. Unlike , increasing the hydrogen content in this layer reduces the defect centers at the interface between a-Si and Ge, and increases the material bandgap slightly, therefore further reducing the dark current. Then, 240-nm thick low-stress Si3N4 is chemical-vapor-deposited (CVD) on top of the a-Si layer, serving as an isolation between the electrical lines and substrate, as well as an anti-reflection coating in the active region. In the first lithography step, the alignment markers are etched into the carrier chip. A 1.45-μm deep groove for the VCSEL is then patterned in the second lithography and etched into the Ge substrate by plasma enhanced reactive-ion-etching (RIE). The third lithography step defines the active PD area, where the metal contacts are connected to the substrate. The next step is the removal of 140 nm of the 240-nm thick Si3N4 leaving a 98-nm thick Si3N4 film in the active region. In the final lithography step, the contact widths of 1.25 μm and spacings of 2 μm are patterned. The 98-nm thick Si3N4 layer at the metal contact regions is etched down to the a-Si:H layer. Subsequently, 11-nm thick Ti and 230-nm thick Au are evaporated and then lift-off is performed. The remaining 98-nm thick Si3N4 in between contacts in the active PD region serves as an anti-reflection coating, and improves the responsivity. The final PD structure is shown in Fig. 9(a). The chip photo of Ge carrier with integrated PDs is shown on the left side of Fig. 11.
Each MSM Ge PD occupies 62×62-μm2, and exhibits 83-fF capacitance, 0.315-A/W responsivity and 7-μA dark current (Fig. 9(b)), all improved from . The measured dark current is larger than the simulation primarily due to accidental plasma damage at the a-Si:H layer during etching of the Si3N4 layer, increasing trap surface states at the a-Si/Si3N4 interface. This problem can be mitigated using diluted HF etching to remove Si3N4 close to the surface. The measured responsivity is less than the simulated results due to the 12% reflections from the surface caused by an 8-nm error in the anti-reflection coating thickness, and 4% scattering loss caused by the surface roughness. As shown in Figs. 9(c) and 9(d), the PD bandwidth is 9.3 GHz at a 7-V bias, mainly limited by the transit time of carriers, and wirebonds from the chip to the PCB trace.
We choose fused silica as the microlens material, which offers very low optical transmission loss at 850-nm wavelength, is compatible with CMOS processes, and can be easily 3-D integrated with silicon substrates. The lenses are built by melting and reflowing of the photoresist into a spherical shape and then dry etching the pattern into silica wafer. Fabrication begins with the plasma etching of the wafer-to-wafer alignment marks into the fused silica wafer. In the second lithography step, 9.8-μm thick cylindrical photoresist patterns with 200 and 220-μm diameters are defined, and subsequently melted and reflown on a temperature controlled hot plate similar to . The height of the photoresist curvature is 17 μm for 220-μm aperture lens, corresponding to a focal point of 710 μm in air (Fig. 10(a)). After the lens curvature are defined, the photoresist is transferred into the fused silica wafer without significantly deviating from a spherical wave shape. The sample is finally cleaned with O2 plasma to remove the photoresist remnants. The final lens thickness is measured as 15.3 μm. The fabricated microlenses for VCSELs and PDs have 200 and 220-μm aperture size with focal point 580 and 730 μm in air, respectively (Fig. 10(b)), exhibiting a 1-dB optical scattering transmission loss per lens. Note that the total electrical-to-electrical loss of an FSOI link calculated in Sec. 3, 20.6 dB, is much higher, and hence the 1-dB scattering loss is tolerable for our prototype. The total fused silica chip area, shown on the right side of Fig. 11, is 0.84×0.84 cm2.
4.3. Chip prototype integration
The VCSEL and fabricated microlens chips need to be integrated with the Ge carrier with high horizontal accuracy and minimal tilt and rotational errors. We use non-conductive epoxy to bond the chips together and use wirebonding to electrically connect them. The tilt and rotational errors between each chip are checked under the optical interferometer and minimized before curing the epoxy. The alignment tolerance is limited to a few microns due to the 0.5-μm optical stage resolution and a maximum +/− 5-μm axial placement uncertainty of the VCSEL chip. The detailed integration steps are explained as following:
First, the 200-μm high VCSEL chip is mounted on the designated 1.45-μm deep grooves on the Ge substrate via UV-curing non-conducting epoxy. Since the grooves are etched using SF6/O2 dry-etching and coated with gold via evaporation, the surface is very smooth and flat, and does not have etch depth differences with respect to the surface of Ge carrier. Therefore, the VCSEL chip placed inside the groove has a very low tilt, which is measured as 0.15 μm and 0.27 μm along the 250-μm and 1-mm VCSEL chip width and length via an optical interferometer, respectively. After the flatness of the VCSEL chip is verified and non-conductive epoxy applied to the two sides of the chip is UV-cured, silver conducting epoxy is placed at the bottom edges of the VCSEL chip to provide a low resistance ground contact. The VCSELs and PDs are 0.75 cm apart, and VCSELs are wirebonded to the 1-mm long 50-Ω transmission lines on the Ge substrate. Each PD also has a 1-mm long feed line and a pad at the end of the line for testing. Then, a 380-μm thick fused silica spacer layer is placed on the Ge substrate and glued by UV-curing to provide a large enough gap between bondwires on the chip and the microlens layer. Finally, the fused silica chip is aligned to the Ge carrier using a high precision optical stage, and glued using non-conductive epoxy after controlling the flatness of the fused silica layer with respect to the Ge carrier via optical interferometer. The measured tilt of the microlens chip after the integration is less than 4 μm from one edge to the other. To improve the in-plane alignment of the VCSEL and microlens, the VCSELs in the array chip are turned on during the alignment process to ensure that each optical beam is centered in the middle of the microlenses. Therefore, even if the VCSEL chip is misaligned with the Ge carrier, it will be aligned well with the microlenses. This approach causes some misalignment between the microlenses and the PDs; however, the large PD area compensates for that.
5. Measurement results
In preparation for testing, the pads for the VCSELs and PDs are wirebonded to a printed circuit board (PCB) with RF connectors. This chip prototype assembly is mounted on an optical test bench with a prism which functions like two face-to-face mirrors with a 45 degree angle. Because the distance between the VCSEL chip and PDs is fixed at 0.75 cm, the prism is moved up and down to change the transmission distance.
The electrical-to-electrical current gain of the optical link is measured as −24.4 dB and −26.6 dB at 1-cm and 1.4-cm distances using a network analyzer. These results are used to extract the optical transmission with respect to distance based on VCSEL slope efficiency and PD responsivity. Similarly, the crosstalk between the adjacent links is also measured and extracted with respect to distance. As shown in Fig. 12(a), the FSOI link achieves 4-dB optical transmission loss at 0.8-mW VCSEL optical power and 1-cm distance. The loss increases to 5.3 dB and 8 dB at 1.4 and 2-cm distance. The crosstalk between adjacent links is −23 dB at 1 cm and −16 dB at 2 cm. Crosstalk up to −20 dB does not impact the estimated BER of the FSOI system, because BER is mainly limited by shot noise of the PD. The 3-dB discrepancy between the measured and calculated optical loss values is primarily due to 1) larger scattering loss per microlens, and 2) 1.2-dB loss from the in-plane misalignment of the VCSEL and the microlens, which is estimated as +/− 3 μm from the measurement results. The transmission loss can be further alleviated by (a) using a VCSEL with smaller divergence angle, (b) a lens with larger aperture size and focal length to reduce the optical clipping losses, (c) smoother lens surface with anti-reflection coating to reduce reflection and scattering losses, and (d) better alignment accuracy using wafer-level or flip-chip bonding techniques.
The measured small-signal 3-dB bandwidth of the FSOI link is 3.3 GHz with a 5.1-mW total power consumption (Fig. 12(b)). Note that the frequency response and hence the 3-dB bandwidth does not change with transmission distance. The bandwidth does not show any discrepancy when the PD is biased at 7 V, because it is primarily limited by the VCSEL bandwidth and the bondwires from the chip to the PCB.
To illustrate the VCSEL bandwidth limitation, the small-signal response of the optical link is simulated based on the experimental and simulation results of the photonics devices. The simulated bandwidth of the MSM PD is 12.1 GHz at 5-V bias, and the VCSEL has a modeled small-signal bandwidth of 5 GHz at 3 mA forward bias current and 1.7-V bias. The overall bandwidth is estimated as 4.9 GHz limited by the VCSEL bandwidth (Fig. 13(a)). To demonstrate the potential bandwidth of the FSOI link, a faster VCSEL from an earlier experiment and the same PD are mounted on two PCBs and placed facing each other. The measured frequency response of this setup is shown in Fig. 13(b). The −3-dB bandwidth is now 6.55 GHz, which can support a data rate of 9 Gbps.
In order to further improve the chip prototype performance, the sensitivity of the PDs needs to be increased. Although MSM Ge PDs with the described barrier enhancement and surface passivation techniques reduce dark current by more than three orders of magnitude, these dark current values are still high enough to affect the signal-to-noise ratio and bit-error-rate for the FSOI system adversely. For example, a PD with 7-μA dark current and the same optical power levels in the performance evaluation section, the bit error rate cannot be lower than 10–9 with a 0.5-pJ/bit energy efficiency at 10-Gb/s data rate. Simply burning more power at the VCSEL to increase the signal-to-noise ratio reduces energy efficiency. Shrinking PD area, e.g., by more than 10 times can reduce the dark current correspondingly, hence leading to a BER less than 10–12. However, smaller PD area puts constraints on the microlens and optics design, requiring larger focal length and aperture size to achieve smaller spot size at the PD, and therefore, limiting the bandwidth density of the interconnect system. This can be overcome by designing one large lens to connect a link with multiple bits (VCSEL-PD pairs), similar to . Alternatively, changing the PD device to a p-i-n structure can significantly reduce the dark current at the expense of process complexity.
An intra-chip free-space optical interconnect (FSOI) system is presented. The FSOI system achieves low loss, low latency, large bandwidth and large energy efficiency for future many-core chips. A performance evaluation shows that energy efficiency of the FSOI system is significantly better than the waveguide-based WDM optical interconnects for a large number of nodes. The first 3-D integrated chip-scale prototype with a 1×1-cm2 area is designed based on real-world optical and photonic device parameters. MSM PDs are fabricated on a Ge carrier chip, and GaAs VCSELs and fused silica microlenses are 3-D integrated on the Ge carrier. The prototype achieves 4-dB optical loss, −23-dB crosstalk, and 3.3-GHz small signal bandwidth at a 1-cm transmission distance. The loss increases slightly to 5 dB with −21-dB crosstalk when the distance increases to 1.4 cm.
This work was partially supported by NSF grants CCF0829915 and DMR1124601, and the DOE Office of Inertial Confinement Fusion under Cooperative Agreement No. DE-FC52-08NA28302, the University of Rochester, and the New York State Energy Research and Development Authority. The support of DOE does not constitute an endorsement by DOE of the views expressed in this article. The authors would like to thank Cornell Nanofabrication Facility for their support.
References and links
1. J. W. Goodman, F. J. Leonberger, S. -Y. Kung, and R. A. Athale, “Optical interconnections for VLSI systems,” Proc. IEEE 72(7), 850–866 (1984).
2. D. A. B. Miller, “Optical interconnects to silicon,” IEEE J. Sel. Top. Quantum Electron. 6(6), 1312–1317 (2000).
3. L. Schares, J. A. Kash, F. E. Doany, C. L. Schow, C. Schuster, D. M. Kuchta, P. K. Pepeljugoski, J. M. Trewhella, C. W. Baks, R. A. John, L. Shan, Y. H. Kwark, R. A. Budd, P. Chiniwalla, F. R. Libsch, J. Rosner, C. K. Tsang, C. S. Patel, J. D. Schaub, R. Dangel, F. Horst, B. J. Offrein, D. Kucharski, D. Guckenberger, S. Hegde, H. Nyikal, C. -K. Lin, A. Tandon, G. R. Trott, M. Nystrom, D. P. Bour, M. R. T. Tan, D. W. DolfiL. ScharesJ. A. KashF. E. DoanyC. L. SchowC. SchusterD. M. KuchtaP. K. PepeljugoskiJ. M. TrewhellaC. W. BaksR. A. JohnL. ShanY. H. KwarkR. A. BuddP. ChiniwallaF. R. LibschJ. RosnerC. K. TsangC. S. PatelJ. D. SchaubR. DangelF. HorstB. J. OffreinD. KucharskiD. GuckenbergerS. HegdeH. NyikalC. -K. LinA. TandonG. R. TrottM. NystromD. P. BourM. R. T. TanD. W. DolfiIBM T.J. Watson Research Center, “Terabus: terabit/second-class card-level optical interconnect technologies,” IEEE J. Sel. Top. Quantum Electron. 12(5), 1032–1044 (2006).
4. I. Young, E. Mohammed, J. Liao, A. Kern, S. Palermo, B. Block, M. Reshotko, and P. Chang, “Optical I/O technology for tera-scale computing,” IEEE Int. Solid-State Circuits Conf. 468–469 (2009).
5. R. G. Beausoleil, J. Ahn, N. Binkert, A. Davis, D. Fattal, M. Fiorentino, N. P. Jouppi, M. McLaren, C. M. Santori, R. S. Schreiber, S. M. Spillane, D. Vantrease, Q. XuR. G. BeausoleilJ. AhnN. BinkertA. DavisD. FattalM. FiorentinoN. P. JouppiM. McLarenC. M. SantoriR. S. SchreiberS. M. SpillaneD. VantreaseQ. XuHP Labs, Palo Alto, CA, “A nanophotonic interconnect for high-performance many-core computation,” 16th IEEE Symp. High Performance Interconnects, HOTI ’08 182–189 (2008).
6. D. V. Plant, M. B. Venditti, E. Laprise, J. Faucher, K. Razavi, M. Chteauneuf, A. G. Kirk, and J. S. Ahearn, “256-channel bidirectional optical interconnect using VCSELs and photodiodes on CMOS,” IEEE J. Ligthwave Technol. 19(8), 1093–1103 (2001).
7. J. Jahns, “Planar packaging of free-space optical interconnections”, Proc. IEEE 82(11), 769–779 (1994).
8. M. W. Haney, M. P. Christensen, P. Milojkovic, G. J. Fokken, M. Vickberg, B. K. Gilbert, J. Rieve, J. Ekman, P. Chandramani, F. KiamilevM. W. HaneyM. P. ChristensenP. MilojkovicG. J. FokkenM. VickbergB. K. GilbertJ. RieveJ. EkmanP. ChandramaniF. KiamilevGeorge Mason Univ., Fairfax, VA, “Description and evaluation of the FAST-Net smart pixel-based optical interconnection prototype,” Proc. IEEE 88(6), 819–828 (2000).
9. H. Thienpont, C. Debaes, V. Baukens, H. Ottevaere, P. Vynck, P. Tuteleers, G. Verschaffelt, B. Volckaerts, A. Hermanne, M. HanneyH. ThienpontC. DebaesV. BaukensH. OttevaereP. VynckP. TuteleersG. VerschaffeltB. VolckaertsA. HermanneM. HanneyVrije Univ., Brussels, “Plastic microoptical interconnection modules for parallel free-space inter- and intra-MCM data communication,” Proc. IEEE 88(6), 769–779 (2000).
10. C. Debaes, M. Vervaeke, V. Baukens, H. Ottevaere, P. Vynck, P. Tuteleers, B. Volckaerts, W. Meeus, M. Brunfaut, J. Van Campenhout, A. Hermanne, and H. Thienpont, “Low-cost microoptical modules for MCM level optical interconnections,” IEEE J. Sel. Top. Quantum Electron. 9(2), 518–530 (2003).
11. M. J. McFadden, M. Iqbal, T. Dillon, R. Nair, T. Gu, D. W. Prather, and M. W. Haney, “Multiscale free-space optics interconnects for intrachip global communication: motivation, analysis, and experimental validation,” Appl. Opt. 45(25), 6358–6366 (2006). [PubMed]
12. B. Ciftcioglu, R. Berman, J. Zhang, Z. Darling, S. Wang, J. Hu, J. Xue, A. Garg, M. Jain, I. Savidis, D. Moore, M. Huang, E. G. Friedman, G. Wicks, and Hui Wu, “A 3-D integrated intra-chip free-space optical interconnect for many-core chips,” IEEE Photon. Technol. Lett. 23(3), 164–166 (2011).
13. J. Xue, A. Garg, B. Ciftcioglu, Jianyun Hu, S. Wang, I. Savidis, M. Jain, R. Berman, P. Liu, M. Huang, H. Wu, E. Friedman, G. Wicks, and D. Moore, “An intra-chip free-space optical interconnect,” 37th Int. Symp. Computer Architecture ISCA 94–105 (2010).
14. D. Louderback, O. Sjolund, E. R. Hegblom, S. Nakagawa, J. Ko, and L. A. Coldren, “Modulation and free-space link characteristics of monolithically integrated vertical-cavity lasers and photodetectors with microlenses,” IEEE J. Sel. Top. Quantum Electron. 5(2), 155–165 (1999).
15. Y. -C. Chang and L. A. Coldren, “Optimization of VCSEL structure for high-speed operation,” IEEE 21st Int. Semiconductor Laser Conf., ISLC 159–160 (2008).
16. P. Dong, S. Liao, D. Feng, H. Liang, D. Zheng, R. Shafiiha, C. -C. Kung, W. Qian, G. Li, X. Zheng, A. V. Krishnamoorthy, and M. Asghari, “Low Vpp, ultralow-energy, compact, high-speed silicon electro-optic modulator,” Opt. Express 17(25), 22484–22490 (2009).
17. J. Cardenas, C. B. Poitras, J. T. Robinson, K. Preston, L. Chen, and M. Lipson, “Low loss etchless silicon photonic waveguides,” Opt. Express 17(6), 4752–4757 (2009). [PubMed]
18. “International Technology Roadmap of Semiconductors,” www.itrs.net. (2009).
19. Q. Xu, S. Manipatruni, B. Schmidt, J. Shakya, and M. Lipson, “12.5 Gbit/s carrier-injection-based silicon microring silicon modulators,” Opt. Express 15(2), 430–436 (2007). [PubMed]
20. B. Ciftcioglu, J. Zhang, R. Sobolewski, and H. Wu, “An 850-nm normal-incidence germanium metal-semiconductor-metal photodetector with 13-GHz bandwidth and 8-μA dark current,” IEEE Photon. Technol. Lett. 22(24), 1851–1853 (2010).
21. F. T. O’Neill and J. T. Sheridan, “Photoresist reflow method of microlens production part I: background and experiments,” Optik 113(9), 391–404 (2002).