There has been growing excitement around optical computing, including efforts by major tech companies. But what is it about light that could make optical computers outperform conventional electronics? This post dives into the fundamental physics differences between photons and electrons that researchers are trying to harness.
Why Optical Computing?
You likely own at least one optical computing device already – your digital camera! It uses a silicon image sensor chip to convert light into electrical signals. So we know optics can process visual information. But today’s computers rely on electrons moving through transistors etched on silicon chips. Why consider photons instead?
Several factors motivate the development of optical computers:
- Electronics face challenges scaling to smaller transistors and higher clock speeds. This limits how much more powerful electronic chips can get (see Moore’s Law).
- Some applications like neural networks and scientific computing are incredibly computationally intensive, so alternative hardware is appealing.
- Components used for optical communications have improved enormously. Photonic integrated circuits now integrate many optical components in a compact chip.
- Optics have some profoundly different physics from electronics that could confer advantages. Exploiting these differences is the key to beating electronics with optics.
The 11 Unique Features of Optical Physics
Researchers have identified 11 properties of light that set it apart from electronics for computing:
1. Considerable bandwidth
The bandwidth of optics refers to the range of frequencies of light that can be generated, transmitted, and detected. This bandwidth spans over 300 THz – from 400 to 700 nm wavelength. By comparison, the bandwidth of electronic circuits is typically just a few GHz. This 100,000x larger bandwidth is a source of major advantage for optics. It enables parallelism by assigning different frequency channels to different computational tasks. And it allows optical systems to be switched extremely fast – in tens of femtoseconds rather than the picosecond timescale of even the fastest electronics.
2. Vast spatial parallelism
Because the wavelength of visible light is so small (~500 nm), a massive number of separate optical beams can be packed into a tiny space. Optical systems leveraging spatial parallelism can far exceed the number of components integrated on electronics chips. In free space optics, the number of resolvable voxel locations in a volume just 5cm x 5cm x 5cm can exceed 10^15. Compare this to even advanced electronic chips which may have around 10^11 transistors in the same area footprint. This vast spatial parallelism allows optics to process huge datasets in parallel.
3. Nearly dissipationless dynamics
One of the most remarkable properties of photons is that they can propagate through free space with virtually no energy loss. Computation can happen in optics simply through interference and propagation phenomena, without needing to dissipatively drive currents and voltages like in electronics. This dissipationless operation, if harnessed properly, can dramatically lower the energy costs of computation. It also avoids heat generation challenges that limit the density and speed of electronics.
4. Low-loss transmission
In addition to free space propagation, optical fibres allow the transmission of light with five orders of magnitude lower loss compared to electrical cables. Signals can be sent long distances before needing regeneration. This allows high bandwidth optical interconnects to overcome the limitations of electrical interconnects for transmitting signals within and between datacenters. For optical computing systems, low-loss data transmission between components can reduce energy costs and latency.
5. Beams can cross without interference
Unlike electrical wires which cannot occupy the same physical space, optical beams can pass through each other with negligible crosstalk or interference. This allows compact, three-dimensional optical systems to be constructed. It also simplifies realizing systems with massive fan-in and fan-out. However, careful optical engineering is still needed to minimize unwanted diffraction, scattering, and reflection in practical systems.
6. Reconfigurable beam steering
The direction of optical beams can be steered using devices like spatial light modulators or acousto-optic deflectors. These devices allow dynamic reconfiguration of free space optical systems, with switching speeds on the order of microseconds to milliseconds. This enables reconfigurable optical interconnects within a processor, which is not possible with fixed electrical wiring. However, electrical switching networks can also reroute signals dynamically, so optics provides a different functionality rather than an outright advantage here.
7. Different fan-in and fan-out
Optics enables different methods for combining signals (fan-in) and distributing signals (fan-out) compared to electronics. This leads to different tradeoffs. Optics can readily support fan-in and fan-out of over 1000 channels, whereas electrical fan-in/out is typically limited to below 10 channels. This very high optical fan-in/out enables efficient distribution and summation of massively parallel data.
8. One-way propagation
Light naturally flows in one direction set by the source, although reflections can occur at interfaces. By contrast, electrical signals can easily flow backwards from outputs to inputs. This reverse propagation can cause issues in certain computing architectures. Optics’ natural one-way propagation avoids these problems and can simplify system design.
9. Wave physics at room temperature
Observing quantum wave phenomena of individual photons is easy with visible wavelength optics but extremely difficult with electrons at room temperature. This accessibility of quantum wave effects for optics could aid certain algorithms, like quantum simulation. But classical wave computing is also possible for some applications without needing true single photon quantum behavior.
10. Quantum nature visible
Storing optical data with just a few photons per bit is possible at room temperature without those quantum signals becoming drowned out by noise, unlike in microwave electronics. Operating optical computers at very low photon numbers and leveraging single photon detection could enable the lowest possible energy operation by reducing quantum noise.
11. Realizations of optimization principles
General computational principles like minimizing time or dissipation can be harnessed in unique ways with optics. An example is light propagating via the fastest path through a principle of least time. Both analog optics and digital electronics can leverage thermodynamic optimization principles, but the realization in optics is often distinct from electronics, providing different capabilities.
But Beware of Pitfalls!
Despite having advantages on paper, optical computers face major practical challenges:
One of the biggest pitfalls is the interface between optics and electronics. These conversions between optical signals and electronic signals, and between analog and digital domains, are often major bottlenecks limiting speed and efficiency. Unless optical computers are designed to avoid or minimize these interfaces, the improvements from optics can be squandered.
Lack of optical transistor
Another major challenge is the lack of an optical transistor that can compete with state-of-the-art electronic transistors. Electronic transistors with switching energies below 10-17 J have been demonstrated. But no optical transistor reaching this performance has been shown that could enable general optical digital logic. This makes it very difficult for optics to compete with electronics for general-purpose processing.
Massive scale required
To gain a substantive advantage, optical computers likely need an incredibly massive scale in terms of the number of parallel channels. For example, an optical matrix-vector multiplier may need dimensionality exceeding 10,000 to beat electronic chips on throughput. Developing schemes that can be manufactured and operated at this daunting scale is highly non-trivial.
Losses and noise
Unwanted losses, noise, and manufacturing variations remain problematic in optical systems. Photon loss mechanisms like material absorption and scattering accumulate, reducing signals. Precision optical components are also prone to noise from vibrations and thermal fluctuations. Fabrication variations easily throw off finely calibrated interference phenomena that optics depend on.
Algorithm co-design needed
Many algorithms running on today’s computers have been optimized for electronic hardware. To maximize their benefit, optical computers likely require co-designing novel algorithms suited to optics’ unique features. Simply running existing algorithms on optical processors is unlikely to be optimal. This poses a challenge since creating new competitive algorithms is difficult.
Given these pitfalls, most optical computing research targets specialized analogue computing tasks rather than general digital logic. Matrix multiplications for neural networks are one key focus. The high complexity of these parallel matrix computations compared to data transfer costs is appealing.
Promising Ideas to Beat Electronics
Despite the challenges, researchers have proposed many promising concepts for harnessing optics:
Optical matrix-vector multipliers
A common primitive operation across applications like neural networks and graph analytics is the multiplication of a matrix by a vector. It turns out this computation can be performed by simply propagating light through certain linear optical systems. The matrix values are encoded in properties of the optical system, while the input vector is encoded in an optical beam. This optical matrix-vector multiplication happens almost dissipationlessly due to the nature of interference. Large free-space optical systems demonstrating over 500,000 simultaneous matrix-vector multiplications per pass have been realized. This massive parallelism is promising for acceleration.
Optical neural networks
Neural networks are one of the leading application domains being targeted for optical computing acceleration. Optics appears robust to the noise and quantization issues arising in analog computing, making it well-suited for neural network inference. Optical convolutions and matrix multipliers can be used to implement key operations like convolutional and dense layers. Systems using lenslet arrays for free-space optical convolutions, and integrated photonic circuits for matrix multiplication have been demonstrated. Optical neural networks leveraging analog operation, high bandwidth, and parallelism could potentially be very fast and efficient.
Special-purpose Ising machines aim to solve challenging combinatorial optimization problems like the traveling salesman problem. Some proposed Ising machines use networks of optical parametric oscillators. These will oscillate in the mode with lowest loss set by the problem parameters. The physics of minimizing loss can then solve hard optimization problems. This approach exploits natural optical computational principles to tackle problems believed to be intractable on conventional computers when scaled up.
Massive neural network models already require optical interconnects using fibre optics and free-space links to avoid communication bottlenecks. This motivates replacing electronics entirely with optical processing to avoid repeated optical-electrical conversions. With both inputs and outputs staying optical, optics may show benefits even for sub-components of highly parallel systems. Optical interconnects customized for neural network architectures like transformer models have been proposed and simulated.
Most algorithms today are optimized for digital electronic hardware, which differs greatly from analog optical systems. Developing novel optical algorithms suited to leverage properties like optimization principles, fan-out, and reconfigurability could better unlock benefits. Optical computers likely need a synergy of tailored algorithms, architectures, and devices to fully realize advantages over electronics. Algorithm innovation remains relatively unexplored but could have a big impact.
The Outlook for Optical Computing
There are good physics reasons optical computing could beat electronics in certain domains if engineered properly. Key criteria for success include:
- Targeting applications suited to optics – for example, with inputs already in optical form.
- Minimizing energy and speed costs of optics-electronics conversions.
- Combining optical advantages like dissipationless operation with huge parallelism.
- Developing optical equivalents of electronics building blocks like nonlinearity and cascadable gain.
- Dramatically scaling up the number of parallel optical processing channels.
- Improving optical device losses, precision, and noise characteristics.
This is by no means easy, but optics has physical differences that can potentially be translated into orders-of-magnitude improvements in speed and energy efficiency.
The optics community still has much work ahead to transform these physics principles into practical and revolutionary optical computers. But proving it is possible would be a major scientific achievement with big impacts. Exciting optics-for-computing breakthroughs likely lie ahead in this decade and beyond!
This post explained the key physics of optics that researchers are trying to harness for computing. Light has distinctive properties like huge bandwidth, massive parallelism, nearly dissipationless operation, and accessibility of quantum effects. These form the basis for proposals to beat the limits of electronics and usher in a new generation of ultra-fast and efficient optical computers!