Photonic Computing: Light as a Substrate for Arithmetic

Modern computing is a process of managing the movement and resistance of electrons. In a standard processor, data travels as electrical pulses through copper traces, meeting logic gates where transistors act as switches. As these circuits have shrunk to the scale of nanometres, two physical constraints have become increasingly difficult to circumvent. The first is heat. When electrons flow through a conductor, they collide with the atomic lattice of the material, converting kinetic energy into thermal energy. This parasitic heating, known as Joule heating, limits how closely transistors can be packed and how fast they can be clocked without the chip melting.

The second constraint is latency born of capacitance. Before an electrical signal can transmit a bit of information across a wire, it must first charge the wire itself. This takes time and energy. As data moves between memory and the processor, or between different cores on a single die, the energy cost of moving the data often exceeds the energy cost of the actual computation. We have reached a point where the speed of electronic arithmetic is no longer limited by how fast a gate can flip, but by how quickly we can shuttle charges across a crowded silicon landscape.

Light offers an alternative substrate for these operations. Photons are bosons; unlike electrons, they do not possess charge and do not interact with one another in a vacuum or through most transparent media. Two beams of light can pass through each other without distortion, and they travel through waveguides with negligible heat generation compared to copper. In theory, a photonic processor could operate at frequencies in the terahertz range, far exceeding the three or four gigahertz limit of modern silicon.

However, the shift from electronics to optics is not a simple substitution of materials. It requires a fundamental change in how numbers are represented and manipulated. While electronics rely on the binary state of a voltage, optics can utilise the continuous properties of a wave, such as its amplitude, phase, or frequency. The challenge lies in performing stable arithmetic with these properties and, crucially, managing the transition between the optical and electronic domains.

The physical limitations of copper and silicon

The continued reliance on silicon-based electronics is a matter of immense industrial momentum and sophisticated manufacturing, yet the material properties of silicon and copper are reaching their architectural end. In a contemporary integrated circuit, the interconnects—the tiny copper wires connecting transistors—account for more than half of the total power consumption. As these wires become thinner, their resistance increases, which necessitates higher voltages to drive signals through them. This creates a feedback loop of increasing thermal output.

Beyond heat, there is the issue of RC delay, the product of resistance and capacitance. A copper wire on a chip acts as a capacitor that must be filled before the voltage at the receiving end reaches a readable threshold. This charging time creates a hard limit on signal frequency. While we have improved transistor switching speeds, the wires connecting them have become the primary bottleneck. Communication between chips on a motherboard or between racks in a data centre suffers from even greater inefficiencies, as the electrical signals must be boosted by power-hungry transceivers to overcome the resistance of longer cables.

Silicon itself is a poor material for light. It is an indirect bandgap semiconductor, meaning it does not emit light efficiently when stimulated electrically. To build a photonic computer, engineers must often integrate disparate materials, such as indium phosphide or gallium arsenide, onto silicon wafers. This heterogeneous integration is difficult and expensive. Furthermore, photons are significantly larger than electrons. While a transistor can be less than ten nanometres wide, the waveguides used to carry light must be large enough to contain the wavelength of that light, typically around 1,550 nanometres for telecommunications infrared. This means optical components are inherently bulkier than their electronic counterparts, limiting the density of an optical chip.

Principles of wave interference and arithmetic

The primary advantage of optics in computing is the ability to perform certain mathematical operations almost instantly through the physics of interference. When two light waves meet, they obey the principle of superposition. If the peaks of two waves align, they interfere constructively, and their amplitudes add together. If the peak of one wave aligns with the trough of another, they interfere destructively, and they subtract.

This is a form of passive arithmetic. In a digital electronic circuit, adding two numbers requires a complex arrangement of dozens of transistors forming an adder circuit. In an optical system, addition can be performed by simply merging two paths of light into a single waveguide. The resulting wave is the physical sum of the inputs. This happens at the speed of light through the medium, with no clock cycles required and virtually no energy consumed by the operation itself.

Optical computing allows for the execution of complex linear algebra at the speed of light, bypassing the thermal barriers of traditional silicon.

Multiplication is handled through attenuation or modulation. By passing a light beam through a material that absorbs or redirects a specific fraction of the light, the intensity of the output beam becomes the product of the input intensity and the transparency of the material. When these two capabilities—addition via interference and multiplication via modulation—are combined, they form the basis for Multiply-Accumulate (MAC) operations. These operations are the fundamental building blocks of matrix-vector multiplication, which is the dominant workload in modern artificial intelligence and signal processing.

Encoding numerical data into light properties

To perform these calculations, numerical data must be translated into a physical property of the light wave. The most straightforward method is intensity modulation, where the brightness of the light represents a value. A dim light might represent a zero, while a bright light represents a larger integer or a floating-point value. This is analogous to how voltage levels are used in electronics, but with the potential for a much higher dynamic range.

Another approach is phase encoding. A light wave is a repeating cycle; by shifting the start of that cycle—moving the peaks and troughs forward or backward in time—engineers can represent data. Phase modulation is often more robust than intensity modulation because the phase of a photon is less likely to be degraded by minor absorption or scattering as it travels through the chip. When two phase-shifted beams are brought together, their relative displacement determines whether they add or subtract, allowing for precise control over the resulting interference pattern.

A third method involves wavelength division multiplexing. Because different colours of light do not interfere with each other, a single waveguide can carry dozens of independent data streams simultaneously, each on a slightly different wavelength. This is equivalent to having dozens of copper wires occupying the exact same physical space without short-circuiting. While this increases the throughput of the system, it also increases the complexity of the hardware, as each wavelength requires its own laser source or a highly precise comb generator to split a single laser into multiple channels.

The architecture of the Mach-Zehnder interferometer

The most common component used to control these properties is the Mach-Zehnder interferometer (MZI). This device is the optical equivalent of a transistor, though it operates on entirely different principles. An MZI consists of an input waveguide that splits into two separate arms of equal length, which then recombine at an output. If the light travels through both arms at the same speed, the waves arrive at the output in phase and combine into a bright signal.

To turn this into a programmable computer component, one of the arms is equipped with a phase shifter. This is usually a small heater or an electrode that changes the refractive index of the material in that arm. By applying a tiny amount of heat or voltage, the speed of the light in that arm is slowed down. This creates a phase delay relative to the other arm. If the delay is exactly half a wavelength, the two beams will be perfectly out of phase when they recombine, resulting in total destructive interference—the light essentially cancels itself out, and the output is zero.

By manipulating the phase of light within a waveguide, a Mach-Zehnder interferometer acts as a perfectly efficient analogue multiplier.

By precisely tuning the phase shifter, the MZI can be set to any state between fully on and fully off. This allows it to perform as a variable attenuator, effectively multiplying the input light by any value between zero and one. When arranged in large, interlocking grids known as photonic mesh lattices, these MZIs can perform massive matrix-vector multiplications in parallel. The data flows through the grid, and the result emerges at the far side with a latency measured in picoseconds.

The unresolved challenge in this architecture is the cost of the phase shifters. While the light itself travels without generating heat, the thermal or electrical elements used to tune the MZIs still require energy. In many current experimental designs, the energy saved by using light for the calculation is partially offset by the energy required to hold the optical components in the correct configuration. Furthermore, because these are analogue devices, they are sensitive to environmental noise and temperature fluctuations, which can introduce errors into the calculation that a digital system would ignore.

Passive computation and the elimination of heat

The primary motivation for shifting from electrons to photons is the circumvention of Joule heating. In a traditional silicon processor, information is moved by pushing electrons through conductive channels. These electrons collide with the atomic lattice of the semiconductor, converting kinetic energy into thermal energy. As transistors have shrunk and clock speeds have increased, this waste heat has become the fundamental bottleneck of electronic design, necessitating complex cooling systems and limiting the density of logic gates.

Photons do not carry an electric charge and do not interact with one another in a linear medium. When light passes through a passive optical component, such as a waveguide or a fixed lens, it does not dissipate energy in the form of heat. In a photonic integrated circuit designed for a specific mathematical task, the computation occurs as the light waves interfere with one another. If the phase shifters and attenuators are set to a fixed state, the actual act of processing data becomes nearly energy-free. The light enters, undergoes interference, and the output represents the solution to a complex equation.

This passive nature is particularly effective for linear operations. In a neural network, the majority of the computational load consists of multiplying large arrays of numbers. In a digital chip, this requires millions of transistors switching states billions of times per second. In a photonic system, this same multiplication is achieved by the physical geometry of the waveguides. The energy cost is shifted entirely to the light source and the detectors, rather than the processing unit itself.

The efficiency of a photonic processor is derived from the fact that interference, not switching, performs the logic.

However, the lack of interaction between photons, which makes them efficient for transport, makes them difficult to use for non-linear logic. Digital computers rely on the ability of one signal to turn another signal on or off. Achieving this with light requires non-linear optical materials that change their refractive index in response to high-intensity light. Currently, these materials require significant power to activate, meaning that while linear algebra is highly efficient in optics, the conditional logic that defines general-purpose computing remains firmly in the domain of electronics.

The energy tax of data conversion

The efficiency gains of optical computing are frequently eroded by the requirement to interface with existing electronic infrastructure. Most data, from sensor inputs to database entries, is stored and moved as electrical signals. To process this data optically, it must undergo two transformations: electrical-to-optical conversion at the input and optical-to-electrical conversion at the output.

This conversion is handled by modulators and photodetectors. A modulator takes an electrical voltage and uses it to vary the intensity or phase of a laser beam. A photodetector performs the inverse, generating a current when struck by light. Both processes involve a fixed energy cost that does not scale down as easily as the computation itself. In many contemporary laboratory benchmarks, the energy consumed by the high-speed analogue-to-digital converters (ADCs) and digital-to-analogue converters (DACs) accounts for more than eighty per cent of the total system power.

For a photonic accelerator to be viable, the number of operations performed in the optical domain must be large enough to amortise this conversion tax. If a system converts data to light just to perform a single addition, it will be orders of magnitude less efficient than a standard CPU. The advantage only appears when the data remains in the optical domain for a long sequence of complex operations, such as the multiple layers of a deep learning model.

There is also the matter of signal integrity. Unlike digital bits, which are restored to a clean high or low voltage at every logic gate, analogue optical signals accumulate noise. Every time a signal passes through a component, a small amount of light is lost or scattered. To prevent the signal from disappearing into the background noise, it must occasionally be amplified or converted back to digital form to be refreshed. This adds further energy costs and complexity to the system architecture.

Scale and the diffraction limit of light

Physical size presents a secondary challenge to the widespread adoption of photonic logic. The size of an electronic transistor is measured in nanometres, with modern production nodes reaching scales where only a few dozen atoms form the width of a channel. This allows billions of transistors to be packed onto a single square centimetre of silicon.

Light is governed by the diffraction limit, which dictates that a wave cannot be easily confined to a space much smaller than its wavelength. Most silicon photonics systems operate in the infrared spectrum, typically around 1,550 nanometres, to take advantage of the transparency of silicon at these frequencies. Consequently, a single optical component, such as a Mach-Zehnder interferometer or a ring resonator, is often tens of micrometres in length. A photonic gate is therefore roughly a thousand times larger than a modern electronic transistor.

This disparity in scale means that a photonic chip cannot currently match the raw density of a high-end GPU. While a digital chip might contain thirty billion transistors, an optical chip of the same size might only contain a few thousand interferometers. This limits the complexity of the models that can be mapped onto a single photonic processor. To compensate for this, designers focus on throughput rather than density. While the photonic chip has fewer components, each component operates at the speed of light with massive bandwidth, potentially processing more data per second than a denser electronic counterpart.

Current applications in linear algebra accelerators

Despite these constraints, photonic computing has found a niche in the acceleration of specific mathematical workloads, most notably in the field of artificial intelligence. Large language models and computer vision systems rely heavily on matrix-matrix multiplication. These operations are essentially massive grids of weighted sums, a structure that maps directly onto the mesh lattices of optical waveguides.

In these accelerators, the weights of a neural network are programmed into the photonic lattice by adjusting the phase shifters. When a vector of data, represented by varying light intensities, is pulsed through the chip, the output represents the product of the data and the weights. This happens at the speed at which the light traverses the silicon, typically a few millimetres.

Several startups and research groups have demonstrated prototypes that perform these multiplications with significantly lower latency than digital hardware. These systems are often used as co-processors. An electronic CPU manages the overall logic and memory retrieval, while the photonic chip handles the heavy mathematical lifting. This hybrid approach allows the system to benefit from optical speed without needing to solve the problem of optical memory or general-purpose logic.

Current demonstrations have shown that these accelerators can be particularly effective for edge computing, where rapid inference is required with a limited power budget. For example, an autonomous vehicle needs to process visual data in real-time to identify obstacles. If the core image processing can be done optically, the system can react faster while drawing less power from the vehicle's battery. However, these systems are currently limited to inference, where the weights are fixed. Training a model, which requires constant updates to the weights, remains difficult in optics because of the speed and power required to reconfigure the phase shifters.

The path toward hybrid optoelectronic systems

The future of the field likely lies not in a purely optical computer, but in a deeper integration of light and electronics on the same die. The industry is currently moving toward a model where optical interconnects replace copper wires for communication between chips, a transition driven by the need to move data between processors in a data centre without the massive heat generated by high-speed electrical cables.

Once optical inputs and outputs are standard on silicon chips, the barrier to including optical processing elements will lower. Engineers are exploring the use of plasmonics, which involves coupling light to the oscillations of electrons at a metal surface. Plasmonics can potentially squeeze optical signals below the diffraction limit, allowing for smaller components that approach the density of electronics. However, these systems currently suffer from high signal loss, and finding materials that can sustain these waves without absorbing too much energy is an active area of research.

Another promising avenue is the development of non-volatile optical materials, such as phase-change alloys similar to those used in rewritable DVDs. These materials can switch between an amorphous and a crystalline state when hit with a pulse of light or heat, changing their refractive index. Crucially, they stay in that state without requiring any power. This could solve the energy problem of holding the lattice configuration, allowing for truly passive optical accelerators that only consume power when the underlying model is changed.

Integrating optical logic directly into the memory hierarchy remains the most significant hurdle for a general-purpose architecture.

At present, the established reality is that optics can perform linear mathematical operations with higher bandwidth and lower latency than electronics. It is also established that the conversion between light and electricity is the dominant energy cost in these systems. It remains contested whether photonic processors can ever achieve the density required to compete with the sheer scale of modern silicon transistors, or if they will remain specialised tools for specific workloads.

A shift toward purely optical computing would require a breakthrough in optical memory and non-linear switching that operates at the single-photon level. Without such a development, the most plausible trajectory is the gradual encroachment of optics into the data centre, first as a means of communication, then as a specialised tool for linear algebra, and eventually as a permanent layer within the architecture of hybrid processors. The success of this transition depends less on the speed of light itself and more on the efficiency with which we can translate our electronic world into its domain.