Silicon photonics is one of the most promising areas in electronics, which promises a significant reduction in energy consumption and an increase in bandwidth. This technology allows electro-optical chips to be built on a single silicon chip, allowing individual chips to communicate through optical rather than electrical signals. It took IBM about 12 years to create the first working hybrid chip. The increased performance of systems with such chips makes it possible to create much more powerful supercomputers than those currently in operation.
Thus, the use of light pulses instead of electrical pulses allows for the rapid transfer of very large amounts of information both within one chip and between different parts of an electronic computing system. Previously, the corporation managed to create a photonic transceiver that provided the function of multiplexing channels according to the wavelength of light. Now the company was able to place chips made using silicon photonics technology directly on the processor module. 
Bert Offrein, head of the photonics group at IBM Research - Zurich, together with colleagues from Europe, the USA and Japan, proposes to consider chips made using silicon photonics technology on a par with conventional silicon processors. The technology for manufacturing such chips is also offered as a hybrid one. The team demonstrated the efficient operation of the hybrid chip, which suggests a possible breakthrough in silicon photonics technology. Current designs typically involve the use of an optical transceiver at the edge of the board. But this is not a solution, since the transceiver is located far enough from the processor, and system performance is significantly reduced.
Blue lines are optical fibers that transmit information in the form of light pulses. The orange-yellow structures are copper conductors through which high-speed electrical signals pass. The developers managed to integrate both types of conductors on one chip.
The development of hybrid chips makes it possible to achieve a multiple increase in the performance of the entire system where such chips are used. The development team was able to develop a method for connecting polymer and silicon light guides, despite the fact that the sizes of such structures are very different.
Computer systems with hybrid chips of this type will be used to work with huge amounts of data, which will allow for analytical calculations, processing data in a matter of seconds. Cognitive computing supersystems can help take technology and science to a new level. But specialists still have a lot of work to do before all this becomes possible.
Today, optical connections are used primarily at the device-to-device level or in optical networks. Their main components and operating principles are discussed in one of the previous ones. However, there are three other categories of interconnects—board-to-board, chip-to-chip, and in-circuit communications—for which the main difficulty in implementing optical interconnects lies in the need to combine optical and electronic functions on a common semiconductor substrate. This problem may be solved by silicon photonics, which uses silicon-based materials to generate, transmit, control and detect light.
Reasons
Interest in the development of optical communication channels at the board level was caused by the creation of blade servers. The obvious target for optical technology here is the backplane. It typically supports high-speed point-to-point or multipoint connections with typical lengths of up to 1 m. Key benefits of optical patch panels include low crosstalk and high bandwidth. However, many of today's optical patch panels are more like patch panels. They demonstrated a range of optical technologies, including polymer silicon fibers, ribbon fibers integrated with vertical cavity surface emitting lasers (VCSELs), planar fiber chains, and photodiodes. But none of them, with the exception of some niche applications, have replaced copper connections.
It is difficult to predict whether the clock speed race in the processor industry will stop, because extrapolating from Moore's law, we can expect the appearance of chips with clock frequencies of about 10 GHz by the end of 2010. However, even at existing frequencies it is becoming increasingly difficult to provide the necessary bandwidth in printed circuit boards or modules based on copper busbars. Loss on copper FR-4 (Flame Resistance 4) PCBs has been shown to increase rapidly at frequencies above 1 GHz, with signal-to-noise ratios deteriorating and timing errors occurring. In addition, crosstalk limits wiring density. High-speed optical channels up to 10 cm long between microcircuits have a number of advantages compared to copper ones. They have lower losses with greater bandwidth, and are not subject to electromagnetic crosstalk. In the last 20 years, optical technologies have been proposed to overcome the limitations of copper wiring, but the relatively high cost and use of exotic materials have made them unsuitable for large-scale production.
Designing electrical connections within integrated circuits operating at multi-gigahertz frequencies is also becoming increasingly complex. In such a situation, optical channels with a typical length of less than 1 cm become potentially attractive. The following reasons contribute to this:
- reduction of delay times compared to the use of copper conductors;
- large bandwidth that does not limit growth clock frequencies transistors;
- reduced power consumption;
- insensitivity to electromagnetic interference.
However, today, efforts to integrate optics and electronics are not only in their early stages, but are also quite expensive compared to traditional copper-based technologies.
Intel is conducting very intensive research in this area, whose approach to solving the problem is based on silicon photonics. The main building blocks of the proposed integrated platform here are a tunable External Cavity Laser (ECL), a silicon modulator, a silicon-germanium photodetector, and low-cost interconnect technology.
Silicon light sources
Although silicon-based lasers are not yet achievable, work on such light sources emitting in the visible and infrared ranges is underway widely around the world. Silicon sources are one of the organic parts for monolithic integration, since they allow the fabrication of both optical elements and control electronics on a single substrate. When using silicon light guides, the radiation should be in the infrared range with a wavelength of more than 1.1 microns, since it is in this window that losses are minimal.
Currently, most research is being conducted in the direction of using the effect of electroluminescence - radiation obtained as a result of electrical pumping. Until reliable and efficient silicon emitters are obtained, the possibility of hybrid integration, i.e., the use of non-silicon light sources connected to silicon light guides, is being considered.
The difficulty in manufacturing silicon light sources is caused by the presence of a band gap with indirect transitions. This leads to the fact that the probability of nonradiative transitions (in particular, Auger recombination) becomes higher than those with light emission.
To obtain infrared radiation, appropriate impurities, such as erbium, must be introduced into silicon. Silicon light guides doped with erbium emit in the infrared range if they are additionally doped with oxygen to form optically active ions in the lattice. However this type devices has a significant drawback: although the radiation intensity is relatively high at 100° K, at room temperatures it drops sharply.
The next way to increase the efficiency of light output in silicon is to reduce the number of nonradiative transitions during electron-hole recombination. This is achieved by reducing the diffusion of carriers to nonradiative recombination centers in the lattice, which increases the probability of light-emitting transitions. One method of such limitation, compatible with VLSI technology, is based on the use of nanocrystals. Other means involve the use of quantum wells in GeSi or crystal lattice defects.
Impurities other than erbium can be included to produce radiation at other wavelengths. For example, terbium provides radiation with wavelengths of 0.98 and 0.54 microns. However, the lifetime and reliability of such devices are too low for practical use.
Another limitation for all types of silicon light sources with direct current is the low direct modulation speed - about 1 MHz. This means that they require external modulators to create high-speed channels.
Device architecture
Work on creating silicon light sources continues, but they are still far from complete. And until a reliable and efficient silicon light source appears, integrated photonics systems will require traditional materials of groups III-V of the periodic table.
Following Intel, we give an example of how an external cavity laser and a silicon light guide with a Bragg grating can be used as a filter for group III-V light generated by the crystal in order to obtain the desired wavelength for optical communications. The strong thermo-optical effect in silicon can be used to tune the generated wave.
The Bragg grating was made by etching a number of 1.2 × 2.3 × 3.4 μm grooves onto a silicon-on-insulator (SOI) wafer. Then, after appropriate processing, the details of which we omit, the Bragg grating was placed in the light guide. The ELC was built by connecting a light guide containing a Bragg grating to an amplifier chip. The resonator was formed between a Bragg grating, serving as a mirror on one side, and an amplifier chip with a 90% reflective coating, forming a mirror on the opposite side. The light guide with the Bragg grating was connected to the amplifier chip at an angle of 8°, which, together with the non-reflective coating, reduced the effective reflectivity of the face to 10-5. The generated beam exited from the side of the laser diode on which a 90% reflective coating was applied and entered the cone of a single-mode optical fiber with a lens (Fig. 1). The lens served to increase the coupling between the optical fiber and the laser. To better understand the operating principle of a laser with an external cavity using a Bragg grating, we present its diagram using more traditional components (Fig. 2).
Silicon modulators
So, above we described a tunable laser based on a complex semiconductor diode of groups III-V and a silicon Bragg grating. However, the laser output produces a continuous wave, which does not carry information. To transmit data over optical communication channels, an optical modulator is required. Such devices with modulation frequencies above 1 GHz were typically made from either ferroelectric lithium niobate crystals (LiNbO3) or complex semiconductors with multiple quantum wells, which exploit the localized Stark effect (splitting of the spectral lines of an atom under the influence of an external electric field) or the electroabsorption effect . The modulation frequency in these devices reaches 40 GHz.
Market demand for low-cost solutions has stimulated the development of silicon-based modulators. In addition, silicon photonics makes it possible to obtain monolithic integrated optical elements based on CMOS technology.
Silicon-based optical modulators have been proposed and demonstrated by many research centers. We present here an experimental version of a device based on a Mach-Zehnder interferometer (MZI). Thanks to the original development of a phase-shifting circuit based on a MOS capacitor built into the MZI passive silicon waveguide, a modulation frequency of 2.5 GHz can be achieved for a wavelength of 1.55 µm.
A schematic representation of the MCI is shown in Fig. 3. The incoming light is split into two equal parts and directed into the two arms of the interferometer. Each of them can contain an active section, which, using an applied voltage, slightly changes the speed of light propagation in the arm. Due to this, a phase shift of the beams is obtained at the output, which, due to interference, leads to intensity fluctuations in the resulting beam.
Silicon photodetectors
The final active component that must be integrated into an all-silicon optical platform is the photodetector. Silicon photodetectors are already widely used for applications using the visible range of light (0.4-0.7 μm), e.g. digital cameras and scanners, due to its high efficiency for these wavelengths. However, most semiconductor lasers used in communications operate in the near-infrared region, typically 850, 1310 and 1550 nm, a range in which silicon is transparent and a poor detector. The most common way to increase the output current of silicon photodetectors is to add germanium, which reduces the bandgap and increases the wavelength of the detected light.
In Fig. Figure 4 shows a cross-section of a photodetector based on SiGe light guides developed by Intel. It is made on the same SOI platform as the previously discussed modulator. The SiGe layer is located on top of the silicon bead of the light guide.
The first version of the detector used 18 quantum wells based on Si0.5Ge0.5 as a light-absorbing material. Sensitivity for some devices reached 0.1 A/V at a light wavelength of 1316 nm. The developers believe that with some improvements the sensitivity can be increased to 0.5 A/V. The bandwidth was below 500 MHz due to a significant shift of the valence band, which prevented hole transport. However, it is believed that this drawback can be corrected by changing the composition of the film. Simulations show that throughput can reach 10 Gbps.
Research in the field of silicon-based planar optics has been ongoing in many laboratories around the world for several decades, but industrial samples have not yet been obtained. However, recently there has been significant progress in understanding current problems and possible ways their decisions.
Quantum wells
A quantum well is a potential well that limits the movement of particles. Getting into it, particles that previously moved freely in three-dimensional space can only move in a flat region, essentially two-dimensional. The effect of motion limitation appears when the size of the quantum well becomes comparable to the de Broglie wavelength of the carriers (usually electrons or holes). Let us consider at a qualitative level how a quantum well is created.
As is known, in accordance with band theory, the energy spectrum of a semiconductor consists of three bands (from bottom to top): valence band, band gap band, and conduction band. If a thin layer of a narrow bandgap semiconductor is placed between two layers of wide bandgap semiconductors, then the conduction band electrons of the middle thin layer, which have an energy lower than the energy level of the wide band gaps of the adjacent semiconductors, will not be able to penetrate the potential barrier formed by them. Thus, two heterojunctions restrict the movement of electrons on both sides, i.e., the electrons are locked in one direction. We can say that the movement of electron gas in a quantum well becomes two-dimensional.
The past 2007 was very successful for the development of many Intel technologies, including in the field of silicon photonics. MIT Technology Review magazine compared Intel's latest breakthrough achievements in this area to a triple win at the races - this is how observers of the leading publication assessed a series of official announcements by the corporation.
According to Justin Rattner, chief technology officer and head of Intel's Corporate Technology Group, “We have empirically demonstrated that manufacturing technologies compatible with CMOS silicon design enable the creation of semiconductor optical devices. Proving this fact was a huge achievement, but no less significant steps are needed for the further development of this technological direction. We now need to learn how to integrate silicon photonics devices into standard computer components; We still don’t know how to do this. But at the same time, we continue to actively work with the divisions involved in the development of various types of products to offer manufacturers models for using semiconductor photonics in Intel solutions.”
Researchers at Intel have developed the world's first semiconductor chip capable of producing high-quality continuous laser beams. Eight lasers are integrated into one silicon chip.
Silicon photonics as a means to eliminate bottlenecks on the road to the era of tera computing
Silicon photonics is the most important component Corporate Technology Group's long-term development strategy aimed at accelerating the transition to tera-computing. The point is that as we develop multi-core processors With enormous computing power, engineers face new challenges. For example, the demand for communication speed between memory and processor will soon exceed the physical limitations imposed by copper conductors, and the transmission speed of electrical signals will become slower than the speed of the processor. Already, the performance of powerful computing systems is often limited by the speed of data exchange between the processor and memory. Today's data transmission technologies are designed for much lower bandwidth compared to photonics, and as the distance over which data is transmitted increases, the transmission speed becomes even slower.
“It is necessary to bring the data transfer speed between the components of the computing platform in line with the speed of the processors. This is indeed a very important task. We see silicon photonics as a solution to this problem and are pursuing a research program that positions us at the forefront of this field,” said Kevin Kahn, Distinguished Research Engineer at Intel Corporation.
Tests of a prototype optical memory module have shown that light, rather than electricity, can be used to access server memory.
A team led by Intel's leading optics researcher, Drew Alduino, is developing an optical communication system between the processor and memory for Intel platforms. A test platform has already been created based on fully buffered FB-DIMM memory, on which it loads and runs Microsoft Windows. The current prototype is proof of the ability to connect memory to the processor using optical communication lines without compromising system performance.
Creating a commercial version of such a solution has enormous benefits for users. Optical communication systems will eliminate the bottleneck between memory bandwidth and processor speed and improve the overall performance of the computing platform.
From research to implementation
The Photonics Technology Lab, led by Intel Distinguished Research Engineer Mario Paniccia, has proven that all optical communications components—laser, modulator, and demodulator—can be manufactured from semiconductors using existing manufacturing technologies. PTL has already demonstrated critical silicon photonics components operating at record-breaking performance, including modulators and demodulators delivering data rates of up to 40 Gbps.
To implement semiconductor photonics technology, six main components are required:
- laser emitting photons;
- a modulator for converting a stream of photons into a stream of information for transmission between elements of the computing platform;
- waveguides, which act as “transmission lines” to deliver photons to their destinations, and multiplexers to combine or separate light signals;
- a case, especially necessary for creating assembly technologies and low-cost solutions that can be used in mass production of PCs;
- a demodulator for receiving streams of photons carrying information and converting them back into a stream of electrons available for processing by a computer;
- electronic circuits to control these components.
%%%
The issue of implementing all these optical communication components using semiconductor technologies is widely recognized as a major research problem, the solution of which will lead to a huge technical breakthrough. PTL has already set a number of world records by developing high-performance devices, modulators, amplifiers and demodulators that provide data rates of up to 40 Gbps. Over the next five years, Intel will look to integrate these components into actual products.
One of the key components of silicon photonics is a modulator that provides transmission speeds of up to 40 Gbit/s.
In the field of semiconductor photonics, Intel has already entered the home stretch. Research in the field of integration of optical elements has already moved from the stage of scientific or technological development to the stage of creating commercial products. The research team is now focused on identifying the capabilities and specifications for designing innovative products based on this revolutionary technology. Ultimately, Intel teams create prototypes and work closely with product development teams to accelerate adoption new technology.
In addition to its own activities, Intel is funding some of the most promising research in this area outside of CTG - in particular, it is collaborating with the University of California at Santa Barbara, which is developing a hybrid semiconductor laser. Talented graduates from various universities from other countries also undergo internships at the PTL laboratory.
Intel's leading optics researcher, Richard Jones, said: “We are facing two major challenges in the current hybrid semiconductor laser project. First, we must move hybrid laser pilot production from the University of California to the Intel plant. Secondly, we have to combine a hybrid laser, a high-speed semiconductor modulator and a multiplexer to prove that we can create a single optical transmitter based on production technology CMOS compatible."
The introduction of silicon photonics technologies will involve the development of new manufacturing processes for producing lasers at high volume scale. Intel's successes in the field of photonics will allow it to significantly outperform potential competitors. PTL Laboratory has already registered about 150 patents. The most prestigious publications, such as Nature, noted the unprecedented achievements of Intel specialists. Additionally, Intel was awarded the 2007 EE Times ACE Award for Most Promising New Technology.
Chasing photons
Unlike existing well-established transistor production processes that have been proven for decades, the technology for creating elements for semiconductor photonics is completely new. There are certain problems on the way to its implementation: optimizing devices, increasing design reliability, developing test methodology, ensuring energy efficiency, and developing subminiature devices.
Test bench for 40 Gigabit silicon laser modulator
One of the most important problems is optimization, because the PTL laboratory develops optical devices for mass computing. While there are no other similar products, standards or other reference points, engineers developing a new technological process themselves search for solutions that best meet the needs of computer applications.
Currently, a group of researchers from the PTL laboratory, relatively small by photoelectronics standards, is gradually switching to the commercialization of semiconductor photonics solutions and expects that mass adoption of this incredible technology could begin as early as 2010. A group of optics specialists from the Digital Enterprise Group (DEG) under under the leadership of Victor Krutul, it is developing applications that will provide the basis for the development of new technology. "We believe that by mastering optical communications, Intel products will continue to comply with Moore's Law," says Krutal.
When transferring information between components of the same computing platform and between different systems not electrons, but photons will be used, the next computer revolution will take place. Leading electronics manufacturers around the world have already joined this race, seeking to gain a competitive advantage. The significance of the new technology can be compared to the invention of integrated circuits. Intel is leading the way in this research and in the development of semiconductor photonics-based components.
65 nanometers is the next goal of the Zelenograd plant Angstrem-T, which will cost 300-350 million euros. The company has already submitted an application for a preferential loan for the modernization of production technologies to Vnesheconombank (VEB), Vedomosti reported this week with reference to the chairman of the board of directors of the plant, Leonid Reiman. Now Angstrem-T is preparing to launch a production line for microcircuits with a 90nm topology. Payments on the previous VEB loan, for which it was purchased, will begin in mid-2017.
Beijing crashes Wall Street
Key American indices marked the first days of the New Year with a record drop; billionaire George Soros has already warned that the world is facing a repeat of the 2008 crisis.
The first Russian consumer processor Baikal-T1, priced at $60, is being launched into mass production
The Baikal Electronics company promises to launch into industrial production the Russian Baikal-T1 processor costing about $60 at the beginning of 2016. The devices will be in demand if the government creates this demand, market participants say.
MTS and Ericsson will jointly develop and implement 5G in Russia
Mobile TeleSystems PJSC and Ericsson have entered into cooperation agreements in the development and implementation of 5G technology in Russia. In pilot projects, including during the 2018 World Cup, MTS intends to test the developments of the Swedish vendor. At the beginning of next year, the operator will begin a dialogue with the Ministry of Telecom and Mass Communications on the formation of technical requirements for the fifth generation of mobile communications.
Sergey Chemezov: Rostec is already one of the ten largest engineering corporations in the world
The head of Rostec, Sergei Chemezov, in an interview with RBC, answered pressing questions: about the Platon system, the problems and prospects of AVTOVAZ, the interests of the State Corporation in the pharmaceutical business, spoke about international cooperation in the context of sanctions pressure, import substitution, reorganization, development strategy and new opportunities in difficult times.
Rostec is “fencing itself” and encroaching on the laurels of Samsung and General Electric
The Supervisory Board of Rostec approved the “Development Strategy until 2025”. The main objectives are to increase the share of high-tech civilian products and catch up with General Electric and Samsung in key financial indicators.
It is possible that someday, using silicon photonics, the entire huge data center can be turned into a single hyperscalable computer, and if we take into account the successes achieved by that time in the field of artificial intelligence, it is not difficult to imagine something like the Ocean on Solaris, described by Stanislav Lem. In the meantime, current servers and data centers resemble PCs in their condition before the advent of SATA and USB: inside there are awkward ribbon cables, outside there are serial and parallel ports for a mouse, keyboard and speakers. But already in 2025, the picture will become different: everything will be unified and connected via optical fiber, which will provide a qualitatively different approach to a number of tasks, in particular, to scaling and high-performance computing. And all this will be possible thanks to advances in silicon photonics.
Silicon photonics is the synergy of two groups of technologies - electronics and optics, which makes it possible to fundamentally change the data transmission system at distances from millimeters to thousands of kilometers. In terms of significance, the result of the introduction of silicon photonics is compared with the invention of semiconductors, because its implementation allows for many years to come to maintain the effect of Moore's law, which forms the basis for the development of information and communication technologies.
For those who are interested in the fundamental principles of this direction, we can recommend the popular science book “Silicon Photonics: Fueling the Next Information Revolution” (Daryl Inniss, Roy Rubenstein “Silicon Photonics: Fueling the Next Information Revolution”), published in 2017. More serious introductions to silicon photonics are the book “Silicon Photonics III: Systems and Applications” by a group of authors and “Silicon Photonics: An Introduction” (Graham T. Reed, Andrew P. Knights). There are also some useful materials on this topic on the Mellanox website.
How it works
If we limit ourselves to practical applications to computing, then, as in the case of electronics, optics and solid state physics can be left aside. To understand at a system level, the most superficial information about the subject is sufficient. It would seem that everything is obvious: the sequence of electrical signals is converted by the transmitter T into a sequence of optical signals. It travels along the cable to the receiver R, which returns them to electrical form. Several types of lasers can be used as light sources, and single- or multimodal cables can be used for transmission.
But we should not forget about the scientific and engineering complexity of the problems that arise when implementing the principles of silicon photonics. It can be judged by the fact that the first experimental work in this direction dates back to the mid-80s of the twentieth century, attempts at commercial development were made in the early 2000s, and the first commercial results were obtained only after 2016. Forty years... Despite the fact that the practical use of fiber-optic communications began in the mid-sixties, and experimental work - much earlier.
The crux of the problem with silicon-based materials is their inability to operate at the same frequencies used in fiber optics, and the use of alternative materials is practically impossible for economic reasons. Enormous investments have been made in existing semiconductor manufacturing technologies. To implement the principles of silicon photonics, they need to be adapted to existing technologies. A solution may be to include miniature receivers and transmitters in the microcircuits and lay the corresponding waveguides between them. This is a most difficult engineering and technical task, which, as of 2017, has been solved.
Intel managed to do this before others - the corporation has already offered its products to the market. We should expect announcements from IBM soon, followed by Mellanox, Broadcom, Ciena, Juniper and a number of other major companies. At the same time, startups that have achieved success are being bought up. The process has begun, but not quickly. The difficulties are caused by the fact that creating new products requires significant funds and time, which gives advantages to the largest vendors.
Four levels of communication
Silicon photonics technologies already make it possible to create 100 Gbit Ethernet, and in the foreseeable future 400 Gbit and 1 Tbit. Such data exchange speeds open up opportunities for convergence modern architectures into qualitatively new ones - at the RSA (Rack-Scale Architecture) rack level and at the ESSA (Extended-scale system architecture) data center level. The limit of the first is limited to the so-called pod (one or more racks), the second covers the entire data center. The components of these infrastructures communicate remotely via the PCIe bus (PCIe-bus interconnects at a distance).
Using silicon photonics, a hierarchical communication system is created, divided into 4 levels:
Level 1 "Chip": The implementation of silicon photonics technologies inside a chip is interesting for several reasons:
- There are significantly more chips than racks, therefore, the need for receivers and transmitters is great, and these technologies will develop rapidly.
- Off-chip communication speeds will increase significantly, so system design principles may change significantly.
- In the long term, one can imagine that optical communications can be used between chip components, for example, for exchange between cores. But at such short distances, copper will retain its position for a long time.
