Silicon Photonics For A Post-Moore Era
What solutions are in the pipeline to support higher performance devices in a post-Moore era? Koji Yamada, head of the Silicon Photonics Group at the Electronics and Photonics Research Institute - a division of Japan's National Institute for Advanced Industrial Science and Technology (AIST), explores the opportunities provided by silicon photonics and discusses the next steps for this exciting PIC platform.
End of Moore's Law in data transmission systems
Moore's Law expresses an empirical trend in device integration in electronic circuits, such as micro-processor chips. The device integration status is defined as the number of elemental devices -- such as transistors -- integrated on a chip, which doubles every two years. However, such exponential growth defined by Moore's law is now coming to an end, because -- ultimately -- miniaturised devices see the de Broglie wave of electrons (more details below). In other words, we are now entering the post-Moore Era.
At the same time, we are seeing explosive growth in information systems, which is one of the major applications of electronic circuits, and improvements in their performance must continue. To cope with such an information explosion, various post-Moore electronic circuit technologies are now being developed under guidelines referred to as "˜More Moore', "˜More than Moore', and "˜Beyond CMOS'.
In "˜More Moore', further geometrical integration is achieved by introducing novel materials and three-dimensional (3D) integration methods. "˜More than Moore' is a system-on-chip (SoC) approach, where non-digital devices are implemented. "˜Beyond CMOS' refers to devices based on novel principles, such as electron spins, and novel architectures suitable for these novel devices.
Figure 1: Trend of switching capacity in a typical large-scale datacenter.
Moore's Law in data transmission, which is an important aspect of information systems, is also facing its end. For example, as shown above, switching capacity in a typical large-scale data centre is growing at a rate of 100 to 1000 times every 10 years1, which is overwhelming Moore's Law in electronic circuits. Since electronic circuits are used for switching systems at present, such explosive growth will become unsustainable in the near future. In the global network system, we are also facing the end of Moore's Law.
For more than 30 years, data transmission capacity per fibre has been increasing with a growth rate of 80% a year through various paradigm shifts in technology, such as from TDM to WDM and multi-level modulations2. However, it has now reached the nonlinear Shannon limit.
Figure 2: Trend of data transmission capacity per fibre.
In order to cope with such explosive growth in data transmission, we also need post-Moore technologies for photonics as well as for electronics - because data transmission systems consist of electronic and photonic elements. For the development of post-Moore photonics technology, the same approaches in electronics can be applied, and silicon photonics provides a platform for post-Moore photonics.
Silicon photonics as a post-Moore photonic technology
More Moore: further geometrical integration (two-dimensional)
Thanks to the very strong optical confinement ability of silicon photonics technology, photonic circuits can be miniaturised considerably, and data transmission capacity per unit chip area can be increased. An example is an integrated WDM receiver chip consisting of multi-channel wavelength filters, photodiode (PD) array, and electrodes for signal output. By using silicon photonics, an arrayed-waveguide-grating (AWG) wavelength filter, which is a standard multi-channel wavelength filter, can be miniaturised to 1-mm square. Moreover, germanium PDs and through-silicon via (TSV) electronic wiring can be scattered over the whole the chip.
Figure 3: Estimated chip area of integrated WDM receiver sub-assembly.
Therefore, as shown above, the required area of a 100-ch WDM receiver chip will be reduced to 1 cm2, which is 1/100 of the area of chips based on conventional technology. Assuming PD operation at 25 Gbps, total capacity will reach 2.5 Tbps/cm2. In parallel transmission systems, where no wavelength filter is required, total bandwidth can be increased up to 30 Tbps/cm2 3.
Spatial division multiplexing (SDM) can also be categorised as two-dimensional geometrical integration. By using silicon photonics technology to make ultra-small optical coupling structures within the area of a fibre core, highly integrated SDM modules consisting of an SDM fibre interface and optical processing circuits can be constructed4.
More than Moore: photonics-electronics SoC
Integration with electronics is the most impactful post-Moore technology for photonic circuits. In particular, integration with modulation drivers, transimpedance amplifiers for PDs, and various control circuits on a silicon photonics chip can significantly reduce the size of photonic-electronic integrated modules. Moreover, since electronic circuits can be placed very close to photonic devices, high-frequency performance can be significantly improved5.
Figure 4:ã€€Photonics-Electronics convergence. (a) monolithic and (b) TSV-based hybrid integration.
Since silicon photonics is based on silicon electronics technology and the silicon platform is reliable and robust, electronic circuits can be integrated by both monolithic and hybrid approaches. Concepts of such photonics-electronics convergence can be seen above. The monolithic approach is very attractive from the viewpoint of ultra-high-volume production, and some short-range data transmission modules have already been commercialised6.
However, we must consider that device performance might be degraded because of narrow margins in the fabrication process for photonics-electronics convergence. For example, in monolithic photonics-electronics convergence, the dark current of germanium PDs is likely to increase. Moreover, typical CMOS electronics technology cannot provide the high-speed electronics required for high-bit-rate optical data transmission.
For improving device performance both in photonic and electronic circuits, the hybrid approach is attractive. Since both types of circuits can be fabricated by their respective optimised fabrication process, high-performance photonics-electronics convergence can be achieved. Hybrid integration is performed by using various wafer-bonding/die-bonding techniques with TSV and micro-solder bump technologies7. Since hybrid integration technology is a kind of 3D integration technology, it can also contribute to geometrical integration in the More Moore approach.
Novel-principle devices and architectures: replacement of electronics with photonics
Since current optical fibre transmission systems show excellent performance, we currently have very small margins for novel-principle devices. However, if we consider data transmission systems as a whole, replacing electronic circuits with photonic ones corresponds to an approach for implementing novel-principle devices and architectures. This approach can be seen in the dynamic optical path network (DOPN) using large-scale optical switch matrixes8,9. Conventional switching systems are made of electronic devices, where numerous small packets are independently routed to a number of unspecified terminals. Thus, power consumption for the switching increases proportionally to the data traffic.
Figure 5: Dynamic optical path network.
This conventional packet switching scheme, however, is not suitable for the transmission of bulky video data to specified terminals, which has been a dominant factor in the recent traffic increase. In the DOPN, as shown above, dedicated optical transmission channels are temporarily constructed by an optical switching matrix. Since packet switching is not utilised in the DOPN, the power consumption for bulky data transmission can be greatly reduced. Recently, a DPON field trial using a silicon-photonics-based optical switch matrix has been carried out8.
Evolution of silicon photonics
As mentioned, silicon photonics technology provides immediate solutions for data transmission systems in the post-Moore era. However, to deal with the expected increase in traffic in the future, silicon photonics itself must evolve much further. This is because there are serious obstacles to further performance improvement and miniaturisation.
One hurdle is the severe fabrication tolerance in silicon photonics, which is a fatal problem. Typical fabrication errors in thickness and width of Si waveguide cores are nanometres in scale. Since a silicon waveguide is very small, such errors can result in large crosstalk and strong polarisation dependence, which significantly degrade performance10.
Nonlinear effects in silicon waveguides are also serious problems in practical applications. Silicon has a large nonlinear coefficient, which is about 100 times larger than that of silica. Moreover, the optical field size in a silicon waveguide is extremely small. Thus, nonlinear effects, such as four-wave mixing and two-photon absorption, are greatly enhanced. The resulting large channel crosstalk and poor power tolerances significantly degrade device performance.
Poor carrier mobility in silicon is also a serious obstacle to achieving ultra-high speed device operation of over 40 Gbps.
Figure 6: End of miniaturization scaling.
But what about the end of Moore's Law in silicon photonics? As discussed earlier, Moore's Law in electronics is ending because device size has reached the wavelength of the de Broglie wave. Referring to Figure 6(a) -- for example -- as an electronic quantum well becomes thinner, the number of energy levels decreases. Eventually, the number is reduced to a single energy level, and this is the end of miniaturisation. The critical thickness is about 10 nm. Any thinner and electron confinement becomes very weak and the device will not work well.
The same analogy is applicable to optical waveguides, as shown in Figure 6(b). In other words, a single-mode waveguide, whose core size is comparable to the optical wavelength, is the end of miniaturisation. Any smaller and photon confinement becomes very weak and the device will not work well. The critical core size is about half of a micrometre for infrared transmission in the 1.55-Âµm telecommunications band. Since the single mode waveguide has been the most essential element in silicon photonics, Moore's Law has already ended.
Through these discussions, we cannot help but conclude that silicon photonics requires post-Moore technology, and -- we believe -- three approaches in the development of post-Moore electronics can be applied here again for the development of post-Moore silicon photonics technology.
More More: further geometrical integration (3D)
For 3D photonic integration, backend photonics is a promising technology. Backend photonics is a performance-assisting technology using additional waveguide systems made of various materials. The requirements for the additional waveguide systems are 1) a low-loss and compact photonics system comparable to the silicon photonics system; 2) functionalities comparable to those in the silicon photonics system; 3) low-loss interlayer coupling between the additional waveguides and silicon waveguides; and 4) the ability to fabricate the additional waveguide system without damaging the silicon photonic system underneath it. In other words, the additional waveguide systems should be constructed by using the backend fabrication technology for silicon semiconductors, which is the reason we refer to the additional waveguide systems as backend photonics.
Figure 7: Integration density and fabrication tolerance in AWG wavelength filters. (200 GHz x 16 ch, neighbouring channel crosstalk < -20 dB)
Silicon-nitride-based materials, such as SiOx, SiON and SiN, are promising for backend photonics. The most attractive feature of these materials is that their refractive index can be largely tuned11. These materials cover a wide refractive index range, and optical waveguides using these moderate-index materials would significantly relax fabrication tolerance. The above chart shows the estimated integration density and fabrication tolerance as a function of refractive index contrast of a waveguide. Using moderate-index materials, fabrication tolerance is significantly relaxed while keeping still high integration density.
Table 1: Nonlinear and thermo-optic coefficients of Si and backend photonics materials
Table 1 shows nonlinear coefficients and thermo-optic coefficients of such materials. These coefficients are less than one-tenth of silicon's. Recently, low-loss interlayer coupling structures have been developed using inverted tapers12,13. By using PECVD technology, the deposition temperature can be reduced to less than 350 degC, which would not damage the silicon/germanium devices underneath the additional waveguide system.
This silicon-nitride-based backend photonics technology has been used to develop low-crosstalk high-resolution AWG wavelength filters and integrated them with silicon-based modulators and germanium-based PDs on a silicon photonics platform14,15.
More than Moore: heterogeneous photonic SoC
The additional waveguide system can integrate various additional photonic functionalities on the silicon photonics platform, because it can increase the degree of freedom in device design. For example, by using silicon-nitride-based waveguides constructed on silicon waveguide systems; polarisation manipulation16,17 and fibre-mode MUX/DEMUX18, have already been demonstrated. Because of material incompatibilities, these functionalities were constructed outside the silicon photonic chip by using bulky conventional optical components.
Novel-principle devices: restarting Moore's Law
As mentioned above, the core size of a silicon waveguide cannot be reduced to less than a half of the wavelength, and Moore's Law has already ended in silicon photonics. However, we know that radio waves, which have metres of wavelength, can be confined in a centimeter-diameter cable. Radio waves are confined as a TEM-like wave supported by surface electron density waves in metals. Here, recall that the miniaturisation limit of electron density waves is determined by the de Broglie wavelength, which is around 10 nm. Thus, we can restart Moore's Law in photonics. Roughly speaking, a photonic system based on such a metallic confinement is referred to as plasmonics.
Figure 8: Structure and optical mode field of plasmon waveguide.
The above schematic shows a photonic waveguide based on plasmonics on an Si photonic platform. The optical field is confined in a 100-nm-wide gap between two metallic plates. By filling this gap with electro-optic (EO) polymers and applying voltage to these metallic plates, we can change the refractive index of the EO polymer. Thus, we can construct a very compact optical phase shifter, which is a fundamental element of optical modulators.
Recently, such a plasmonic-polymer waveguide system has been used to develop Mach-Zhender interferometer optical modulators19. Since the gap is very narrow, the electric field in the gap is very strong and the modulation efficiency becomes extremely high. For example, the voltage-length product -- which is a measure of the efficiency of phase shifters -- reaches 0.006 V.cm, which is less than 1/100 of that of a typical silicon modulator. Thus, a device a few tens of micrometers long can function as a modulator with practical modulation depth, and such a small device can operate at very high frequencies of over 100 GHz. Such plasmonic devices would be a significant breakthrough in the post-Moore photonics technology.
Moore's Law in data transmission systems is nearing its end, and we need post Moore-photonics technology. In the development of post-Moore photonics technology, we can apply the same approaches that have been established in electronics, and silicon photonics can provide immediate solutions for post-Moore photonics technology. However, we must note here that silicon photonics itself requires post-Moore technology because of the quantum limit in miniaturisation and poor material characteristics. Here again, the same approaches established in electronics are applicable to post-Moore silicon photonics with the help of backend photonics and plasmonics.