Is it possible to replace electrons in a computing circuit with faster photons in order to immediately dramatically increase the speed of calculation? In principle, yes, although this requires significant rework of the computing circuit itself. But the results are unlikely to be indisputably positive
By today's microelectronics standards, quantum optical integrated circuits are not so "micro" (Source: Imec)
"Photonics" is an extremely polysemy term. First, its origins are at least twofold: for example, English-language sources date back to 1954, when American science fiction writer and publisher John J. Bush went back to 1954. Educated at the legendary Massachusetts Institute of Technology and Duke University, John W. Campbell said in a private letter to the German philosopher Gotthard Günther: "By the way, I decided to invent a new science, photonics. Its relationship with optics will be the same as that of electronics with electrical engineering. In the Russian tradition, it is customary to trace the first use of the term "photonics" back to 1967, when Academician A. N. Terenin, known for his work in the field of biomolecules, published the book "Photonics of Dyes and Related Organic Compound Molecules".
In any case, photonics, as a field of natural science that encompasses all systems in which photons carry information, began to develop actively in the 1960s, beginning with the widespread use of lasers and later optical fibers. Formally, the dawn of the term being truly widely used dates back to the 1980s, when it was mainly applied to telecommunications. But researchers weren't going to limit themselves to transmitting data over fiber optics from the start. The maximum speed of light in our universe, where photons move, promises unprecedented performance of logic circuits in the future, in which electromagnetic radiation quanta will act as carriers of information instead of electrons, but it turns out that it is not easy to design and implement such a scheme in practice. But the researchers are not giving up – although we should definitely not expect to directly replace electronic chips with photonics as part of the PCs, servers, and smartphones that surround us.
Where is there no silicon?
Initially, the integrated devices that put the ideas of photonics into practice were mainly related to telecommunications, so the best substances from an optical point of view were used as the material basis for their construction: doped quartz glass, lithium niobate, indium phosphide, etc.: silicon-based electronics were independent, and the end nodes of the reception/transmission of optical signals based on photonics were independent, and there were corresponding converters between them. However, it is only when electronic circuits are integrated that electronic circuits begin to truly conquer the world – and are therefore suitable for mass production with significantly lower unit production costs.
Silicon photonics is attractive because it allows in-line production of chips using methods that have been excellently tested on VLSI (source: Imec)
For this reason, integrated solutions for photonics also seem to be the most promising. Fortunately, (poly)silicon is not a substance that is not suitable for the creation of (near)optical systems: it is transparent to infrared radiation (wavelengths greater than 1 μm) and, in the near-infrared range, has a dimensionless refractive index of about 3.5. For comparison: the various types of optical glasses used in the manufacture of lenses for the use of visible light are refracted with indicators in the range of 1.4 to 2.2. This means that silicon lenses will be able to deflect infrared radiation fluxes very efficiently – even at very small physical sizes (more precisely, optical apertures). If so, theoretically, nothing prevents the creation of pure silicon integrated circuits for the needs of photonics, in which electronic and optical components will be located on a single substrate.
Today, silicon photonics (SiPh) is a fast-growing IT field with a material basis of hybrid quantum optical integrated circuits (PICs; The Russian abbreviation is KOIS; It is not entirely correct to simply call them "photons", because there are electronic components, and "photonics" is somehow discordant). PIC/KOIS is produced using classical lithography methods, using silicon on an insulator (SOI) wafer, ensuring an acceptable cost and mass production of the resulting product. So why do the new CPUs, GPUs, and NPs (neural processors) from generation to generation represent semiconductor electronics VLSI, and not KOIS? There are many reasons for this, and in order to understand them, it makes sense to go back to the basic concept of the electronic bandgap of semiconductors, and more broadly, conductivity is a physical phenomenon.
The scale of the production process required to produce KOIS is relatively large, allowing production to be carried out even without state-of-the-art lithography machines (Source: TU/e)
Recall that even at absolute zero temperatures, metals are distinguished from all other substances by the presence of free electrons (the so-called conductivity electrons) in them. In addition, the presence is very noticeable: about 3-3 units in 10cm22 or more. At temperatures above absolute zero, nonmetals also acquire a certain concentration of conductive electrons, because thermal vibrations near the position of the atoms in the crystal lattice cause (with a certain probability) the separation of the electrons from their atomic shells and their transition to a free state. This probability is determined by the difference in energy between the electrons located in the shell of the atom (in the valence band) and the electrons that gain freedom (stay in the conductive region). In fact, the minimum gap between the "bottom" of the conduction band and the "upper limit" of the valence band is called the band gap width - as we have pointed out more than once in the series of articles, this parameter "does not suit you", and many exotic semiconductors are preferable to silicon. In general, at room temperature, semiconductors have 10 13-10 17 free electrons in a volume of 1cm3, while semimetals have 10 17-10 22 free electrons. Of course, these boundaries are conditional.
Thus, the two stable electron regions in matter – valence and conductivity – can be located symmetrically or asymmetrically. This refers to the representation of these regions on the graph of the dependence of the electron energy on its momentum (by the way, it becomes obvious why there is a "bottom" in the conduction region and a "ceiling" in the valence zone). Of course, it would be more accurate to talk about quasi-momentum (crystal momentum) - the characteristics of quasiparticles (the same moving electrons in the periodic field of the lattice) accepted in solid-state physics, but in this case it is not so important. On the other hand, how symmetry manifests itself in cases where the extrema of the corresponding plot (the minimum value of the band descriptor and the maximum value of the valence zone image) falls on the same (quasi-)momentum value. It is clear that in a real crystal, it is necessary to take into account its three-dimensional structure, which makes the band diagram of the electron spectrum more complex. But in any case, for a semiconductor at room temperature, when the density of free electrons is relatively low, we can assume that only states near the spectral maximum in the valence band are released, and only states near the spectral minimum in the conduction band are occupied. This type of semiconductor is called a straight band.
Quasi-momentum diagram of electron energy for straight-band (left) and non-straight-band semiconductors (Source: Wikimedia Commons)
Thus, in non-direct-band semiconductors, the minimum energy in the conduction band is inconsistent with the maximum energy in the valence band. Silicon, on the other hand, is a non-indirect wavelength semiconductor, which greatly reduces its efficiency as an LED material. We're talking about hybrid integrated photonics, where the optical emitter, waveguide, and classical logic loop are done all at once on a common substrate, right? Thus, in straight-band semiconductors, under certain conditions, the electrons with the minimum conduction band energy are able to lose part of their energy by emitting photons and move to the maximum of the valence band without changing their (quasi) momentum, i.e. in an extremely energy-efficient manner. If the minimum of the first region and the maximum of the second region are separated in the momentum space, then the free electron must also adjust its momentum to become valency. This happens due to the formation of quasiphononic particles – roughly speaking, part of the composite energy (free electrons with a hole where the valence electrons should exist) are ultimately used to heat the sample.
As a result, the efficiency of radiative recombination is very low: the transition of electrons from freedom to valence, emitting photons and phonons at the same time, takes several orders of magnitude longer than direct-band semiconductors with only photons. This, in turn, creates a new problem: since even a real sample of high-purity crystalline silicon is not completely free of defects, it is possible for free electrons and holes in the valence band to recombine on these defects without emitting photons, since this turns out to be more energetically advantageous than "waiting" for the slow formation of phonons for radiative recombination. Fortunately, they learned to cope with this problem at the beginning of the 21st century thanks to the introduction of certain types of additives in silicon, and by far silicon photonics offers the widest range of devices – waveguides, modulators, photodetectors, and combinations of these components assembled into hybrid KOIS. For example, SiPh transceivers can exchange data at speeds of up to 400 Gbit/s or even higher while consuming much less power than microelectronic transceivers, while SiPh lidar can form highly detailed images of large areas of the Earth's surface in a negligible amount of time. In conclusion, silicon photonics is good for everyone, but it is difficult to organize logical elements on this basis. Another question is whether it's worth it?
Here we need to be logical
Any logic circuit needs not only a gate to perform the logic operation correctly, but also a channel for signal propagation. For microelectronics, channels are traditionally made of metals that conduct electricity well, such as aluminum and copper. As far as silicon photonics is concerned, at first glance, the channels should not have any special problems either. As mentioned earlier, silicon itself is characterized by a refractive index of about 3.5 at the 1550nm operating wave (infrared radiation) of many SiPh devices, while the refractive index of silica SiO2 is about 1.6. In fact, it is for this reason that silicon wafers on insulators are chosen as the basis for the production of silicon photonic chips, and not blanks made of pure silicon: semiconductors (silicon) ridges located on a SiO2 substrate, limited by an air gap or the same silica on the sides and top, become excellent waveguides for infrared photons of the desired spectral range: they reflect well from the boundaries between media with significantly different refractive indices (the effect of total internal reflection) and propagate along the crest with acceptable losses.
(a) – structure of billet (silicon on insulator) used in the manufacture of KOIS with a blue stripe with a thickness of 0.3 μm – SiO2 layer; (b) is a waveguide formed on the surface of this blank with a width of 3 μm, in which a beam of photons propagates - the redder the pear-shaped region of the beam section, the higher the flux intensity there (source: NVIDIA)
Everything will be fine, but only the mention of the characteristic wavelength of the working radiation - 1550 nm - should remind even the most attentive reader who knows a little about the marketing scale of today's microelectronic processes that "5 nm" and "3 nm" - by the way, the "1 nm" promised by chipmakers is not far off. Earlier, in a series of articles devoted to lithography production, we explained to what extent the marketing name of the technological process differs from the actual minimum resolution limits in microcircuit manufacturing, but in any case, these are the initial tens of nanometers - not hundreds and thousands. Thus, a waveguide with a solid-state VLSI (VLSI), with its characteristic channel width in microns, looks like an ordinary city street, flanked by buildings and trees built with Lego bricks, rather than real full-size bricks.
And this is just a direct waveguide! On the circuit diagram, the current-carrying bus is usually represented by straight line segments; If it is required for the implementation of a given logic element, it is connected by right angles. In fact, the conductive metal track is also perfectly bent in the position and manner at the command of the design engineer – it will not go astray if the electrons inside the metal bus suddenly turn 90°. The photons inside the waveguide are another matter: in order to be reflected in the desired direction, they must meet the wall at an angle far from the plumb. This means that if the waveguide from one point on the diagram to another cannot be straightened, it does not have to be disconnected at right angles, which is allowed by the current-carrying bus, but for the silicon already mentioned, be careful to bend with a minimum radius of about 250 μm. Yes, in recent years, the characteristic dimensions of cross-sectional waveguides have been estimated to be hundreds of nanometers, and their permissible bending radius is measured in microns, but these scales are still significantly higher than those of pure microelectronics.
One of the basic elements of KOIS, the directional coupler, can be bent: (a) is its general appearance; (b) is the design parameter of the circuit, including an indication of the minimum permissible bend radius (source: Micromachines)
In any case, logic circuits built solely on the basis of optics will expand compared to microelectronic analogues, both due to the scaling of the signal transmission channels (waveguides instead of metal tracks) and the need to rearrange the entire circuit in order to replace the rectangular turns of the channels with elegant fillets. The good news is that, unlike buses used for charge propagation, optical waveguides can freely intersect without causing short circuits and data loss: this alone makes it possible to use up to two layers of structure instead of stacking metal interconnect layers on the silicon substrate of the chip. However, the scale of the retreat along the frontier of manufacturing specifications is still discouraging: hybrid KOIs are much larger than full microelectronic KOIs. However, the use of semiconductors other than pure silicon offers some promise – for example, silicon nitride (SiN)-based optoelectronic circuits make it possible to use shortwave radiation for data transmission down to the visible region of the spectrum. By the way, semiconductor lasers based on straight-band materials such as indium phosphide (InP) are becoming a source of light for KOIS - in this case, pure silicon is simply not suitable, that is, silicon photonics requires an external light source. Therefore, an input coupler is required to provide luminous flux to the SiPh device and further redirect the converted light along the optoelectronic circuit. All of this makes the circuit implementation of silicon photonics very complex.
Synopsys, one of the leading developers of various chip engineering design tools in the U.S. and the world, has earned a name in its glossary "What is Silicon Photonics?" The section states directly: "PICs do not form a complete solution that can be used immediately in isolation from 'traditional' electronic integrated circuits." This means, for example, transceivers – key blocks that exchange information via fiber optics and convert data streams from electronic form to photon form and vice versa – also inseparable from the microelectronic circuitry that controls signal modulators and performs additional signal processing. In turn, an optical modulator is required to encode the data in the laser beam, which then propagates along the optical fiber to a photodetector, which is part of the transceiver on the other side of the communication line.
Ultra-narrowband reconfigurable diode lasers can be used for external "illumination" of KOIS (Source: PhotonDelta)
Theoretically, it is possible to create a "dot and dash" stream in Morse code by simply turning the laser beam on and off, but it is more acceptable to modulate a continuous stream of photons by polarization, intensity, or phase, for reasons that will be explained later, for example, based on a simple Mach-Zehnder interferometer (Mach-Zehnder modulator). This modulator is a waveguide that branches into two channels, then converges again, and applies a voltage to one of the channels, the magnitude and duration of which are set by the microelectronic circuit. If the physical parameters of the two channels of the bifurcated waveguide are the same, then after the two beams converge into a single beam, the light will remain exactly the same as it was before it hit the interferometer. However, since the voltage applied to the waveguide changes its characteristics, especially the refractive index, the light waves propagating through this branch of the interferometer also become different from those propagating along the second branch at the same time. Together, the initial (but halved intensity) and modulated photon fluxes form a beam with variable intensity, in effect, the information is encoded. In general, in order to save energy, a high V value is applied not only to one branch of the Mach-Zehnder interferometer, but to one branch and V/2 to another, therefore, the same effect is obtained. Of course, in order to improve efficiency, such waveguides are made of materials with strong electro-optical effects, such as LiNbO3, GaAs, InP, which makes their integration with silicon VLSI more difficult.
The crystallization of causality
In addition to the microelectronic circuitry that controls the operation of the optical modulator, a typical transceiver contains more circuitry that functions due to electronics, such as error correction, packet processing, signal amplification, and so on. For the most part, although silicon photonics in the form allows joint lithography of electronic and optical components on a single blank wafer, the two circuits are executed on separate chips and are carried out according to distinctly different technical processes, and have been combined into a working solution – the transceiver itself – which is already in the final stages of release. The reason is banal and has already been mentioned more than once: if in the early 2000s, there was still a rather serious discussion about the relatively rapid (in 15 years or so) replacement of electronic logic circuits inside computers with photonic circuits, then by the end of the second decade of the 21st century, the futility of technological development in this direction became obvious. The miniaturization of classical VLSI is happening at such an astonishing rate (thanks to Moore's self-fulfilling prophecy) that photonic circuits are hopelessly behind electronic circuits in terms of the number of conditional transistors (basic logic gates) per unit of occupied area – and this lag is simply not something that can be overcome by another technological leap.
Various optical components of KOIS (not a single scale): external lasers, input couplers, optical modulators, waveguides, directional couplers, optical ring resonators, photodiodes, and photonic crystals (source: Comsol)
The "3nm" manufacturing process allows the placement of a flat chip surface on 1 mm² (we are not even talking about the possibility of VLSI vertical packaging here!). About 300 million transistors. Yes, in order to connect them correctly, forming a logical profile, up to a dozen layers of metal tires are required, but since each subsequent layer is made according to a much larger technical standard than a lower one, it does not cause serious production problems. At the same time, the characteristic dimensions of the typical elements of a SiPh chip, even for the same modulator, cannot be smaller than the light propagating through this element of at least a few wavelengths; For the near-infrared range - microns with multiple micrometers. This is not a rule of thumb, but a law of nature: no advances in the field of technology will reduce the wavelength of infrared photons to a few nanometers. And if you're trying to use radiation with significantly shorter wavelengths instead of infrared and visible light, the problem begins with choosing a material that can adequately perceive hard UV and X-ray irradiation, which is about the same transition from DUV to EUV lithography that we considered earlier.
In total, a single modulator on a hybrid SiPh chip will conditionally occupy 10×10 μm (10−4 mm²). If such a chip is made according to the "3-nm" technical standard, then 30,000 transistors will be placed in the same area. But, roughly speaking, a modulator is simply a simple valve capable of allowing or disabling light from passing through, just as a transistor controls the passage of current through its channel. In other words, when trying to create a pure photonic computing circuit, the designer will sacrifice four to five decimal orders of magnitude per unit area of the number of logic components per unit area of VLSI. Let's not forget that photonic waveguides are longer in length compared to metal current-carrying busbars; And the corners need to be rounded smoothly so that the effect of total internal reflection continues to work inside the waveguide that changes direction – all of which can increase the area occupied by a pure photonic logic circuit by another order of magnitude, or even an order of magnitude of two. By the way, for all its advantages, silicon is not very good as a medium for the propagation of infrared radiation - the losses in the channel (the number of photons at the output divided by the number of photons at the input) can reach 90% even at micron distances.
Photonic crystals are very beautiful for starters (Source: LaserFocusWorld)
The von Neumann architecture of almost all modern computers (with the possible exception of analog computers and quantum computers) imposes another limitation on the development of purely optical computers, which means processing operations in the acquisition-execution cycle. In this logic, the microprocessor receives (selects) an instruction from memory, executes the instruction taking into account the data received from the input device, transmits the result to the memory or output device, and repeats the cycle. On the other hand, photons do not exist at rest, so it is difficult to create a memory subsystem that operates exclusively on these particles and functions like DRAM and NAND chips. However, the direction of photonic crystals has been actively developing recently, and this method of organizing purely optical (non-electronic) calculations may well prove promising.
Photonic crystals are fairly broadly represented in nature: they are particularly responsible for the iridescent sheen of opals or the wings of tropical butterflies. The physical basis of the medium defined as a photonic crystal is a periodic structure with a variable permittivity value (as a reminder, the refractive index value of light depends on it), so that in different regions of this medium regions a region is formed, allowing and prohibiting the passage of photons of certain energies (or wavelengths). In a sense, a photonic crystal for optical quanta is a functional analogue of a semiconductor for electrons, which is suitable for organizing logic gates precisely because it contains regions where charges are allowed and forbidden. Photonic crystals are divided into one-, two-, and three-dimensional according to the way the variable permittivity regions in the medium are ordered. As far as applied photonics is concerned, the most relevant at this stage are two-dimensional (planar) photonic crystals, for example, the production of lithography by lithography on ordinary substrates is no more difficult than comb FinFET structures.
A waveguide with perforated walls allows you to control the characteristics of the beam passing through it - as simple as anything ingenious (source: intechopen)
The easiest way to organize a two-dimensional photonic crystal is to form a wider waveguide instead of a narrow silicon waveguide, and without sidewalls in the usual sense. On both sides of the main line, a stream of light quantum will propagate along the line, perforated using standard lithography methods for selective etching. Thus, along the path of the photon beam, two rows of ordered irregularities appear, both in their size and in the distance between them are much smaller than the wavelength of the operating radiation, so that they do not significantly interfere with the intact internal reflection from the waveguide boundary. More precisely, by choosing the geometrical dimensions of the cavities that are periodically located, as well as the distance between them, it is possible to determine the permissible and forbidden frequencies: the light waves of the former will pass freely through the photonic crystal, while the latter will be extinguished due to destructive interference. In addition, if certain irregularities are now introduced into the periodicity of the cavity, a beam with previously forbidden frequencies will again be able to propagate through such a waveguide, but in a more cunning way. As a result, photonic crystals with periodically varying refractive indices turned out to be analogues of semiconductors, with bands that prohibit the entry of electrons, similar to X-ray Bragg diffraction in some areas of the crystal lattice (we also discussed this phenomenon in detail in the article on EUV lithography).
As mentioned earlier, basic semiconductor devices (diodes, transistors) are well suited for organizing logic circuits precisely because they are nonlinear (in the sense that the level of response is out of proportion to the level of the signal supplied to the circuit). Optical systems, on the other hand, were linear, or rather, before the advent of lasers, they behaved only as linear in ground-based laboratories. The high intensity of laser radiation, the electric field formed by it competes for power with the bonds that hold electrons near the nucleus, causing the optical medium to respond in a non-linear manner to the flow of high-energy subtons passing through them. In particular, the role of high harmonics in the polarization structure of electromagnetic waves is increasing: the optical Kerr effect is beginning to manifest itself, the essence of which is that a strong beam in the medium itself can generate a modulated electric field without the need to apply an external electric field. As a result, photonic crystals through which sufficiently strong laser radiation passes do not require external microelectronic circuits to self-modulate – only the geometrical parameters of the lattice that form the waveguide need to be carefully selected and the inhomogeneities in them arranged in order to obtain the desired result at the output.
Organization of the photonic crystal onto the "logical HE": (a) — specify the node on the graph and the corresponding truth table; (b) is the actual embodiment of the gate in a semiconductor; See explanation in the main text (Source: IntechOpen)
And what exactly this result is depends on what kind of material is used to form the photonic crystal. When interacting with a high power laser radiation of intensity I, the refractive index n starts to depend directly on it according to the formula
n （I） = n0 + n2 I，
where n0 is the classical linear refractive index and n2 is its second-order nonlinear component, which is due to the optical Kerr effect. Thus, depending on the material chosen for the photonic crystal, N2 can take either positive or negative values. In this regard, pure silicon is not the most prominent medium, since its n2 index is both positive and extremely small, so researchers are now actively working on many other materials suitable for photonics, which are often very exotic.
Either way, photonic crystals have made it possible to create fully functional logic gates – "NOT", "AND", "OR", and so on. For example, the "HE" gate requires crystals based on materials with a negative value of n2 – an example is a quantum dot based on lead selenide PbSe. In this case, the structure of the cavity that forms the photonic crystal is disturbed because a nanocavity of such size is introduced and the distance from the waveguide is so far away that in the absence of a modulation pulse, a constant flow of photons will pass freely through the crystal (logic "0" at the input end, logic "1" at the output), and the activation of this pulse will cause the output luminous flux to disappear because of the instantaneous change in the refractive index (at the input, "1"). The output is "0"). It is even possible to form a structure similar to semiconductor random access memory (RAM), but it should be borne in mind that the physical size of a pure photonic chip will in any case be orders of magnitude larger than that of a transistor chip that fully ideological corresponds.
That is why the most desirable field for the development of photonics is still located outside the boundaries of the tasks that von Neumann's computer has already successfully solved. Simulation, quantum computing, special nodes for accelerating AI computing, we talked about in the previous article, but not the banal replacement of transistor circuits with photonic crystals. Fortunately, for the foreseeable future, these tasks will not be trivial for classical computing systems and will obviously be most relevant to humans.