When photonic AI is discussed in the press, a familiar claim appears: photonic chips can perform operations with far less energy than electronic chips, sometimes by orders of magnitude. The claim is often based on a calculation that considers only the optical computation itself — the energy dissipated in the waveguides or resonators where interference performs the math. That calculation can be correct and still misleading, because it omits the energy consumed by everything else in the system.
The real metric for photonic AI is system-level energy per operation: the total energy consumed by the accelerator to perform a useful computation, divided by the number of operations, including all overhead. This article explains why the distinction matters, what belongs in the energy budget, how different architectures compare, and how to evaluate energy claims critically.
What “energy per operation” means
Energy per operation is a deceptively simple metric. In electronics, it is often quoted as the energy per multiply-accumulate operation, or MAC. A MAC is the basic operation in neural network inference: multiply two numbers and add the result to an accumulator. A modern GPU or tensor accelerator might perform trillions of MACs per second, with an energy per MAC that depends on the process node, the data movement, and the workload.
In photonics, the analogous metric is the energy per MAC performed optically. The optical MAC itself can be very efficient, because interference does not dissipate energy in the way that charging a wire does. But the optical MAC is not the whole computation. The input must be encoded onto light, the weights must be set, the light must be generated, the output must be detected, and the result must be stored or transferred. Each of those steps consumes energy.
A fair comparison requires defining the system boundary. Does the energy budget include the laser? The modulator drivers? The detector amplifiers? The digital-to-analog and analog-to-digital converters? The memory access? The control electronics? The cooling? Different papers draw the boundary in different places, which is why reported energy figures can vary by orders of magnitude and why direct comparisons are often meaningless without careful reading.
The components of the energy budget
A system-level energy budget for a photonic AI accelerator includes several components.
Light generation: The laser or lasers that provide the optical carrier. The wall-plug efficiency of the laser — the fraction of electrical power converted to optical power — matters, as does whether the laser is on continuously or pulsed, and whether one laser serves many channels or each channel needs its own.
Input encoding: Converting the input data from electronic to optical form, typically with modulators. Modulators consume energy to drive, and their efficiency depends on the modulation format, the speed, and the voltage required.
Weight programming: Setting the weights of the photonic matrix. In a Mach-Zehnder mesh, this means driving phase shifters. In a microring design, it means tuning resonators. In a phase-change design, it means switching the material. If weights are set once and held, the static power may dominate. If they are updated frequently, the dynamic power matters.
Optical computation: The propagation and interference of light through the photonic core. In an ideal system, this dissipates little energy. In a real system, there are propagation losses, coupling losses, and insertion losses that must be compensated by higher laser power.
Detection: Converting the optical output back to an electronic signal with photodetectors and amplifiers. Detection energy depends on the required signal-to-noise ratio, the bandwidth, and the amplifier design.
Analog-to-digital and digital-to-analog conversion: If the system interfaces with digital electronics, converters are needed. Converters can be energy-hungry, especially at high speed and high resolution. Some photonic designs avoid converters by staying in the analog domain, but they then face precision and noise challenges.
Memory and data movement: Moving data between memory and the accelerator, and between layers. In many AI workloads, data movement dominates energy consumption in electronics, and photonics does not automatically solve this. If the photonic accelerator still needs to fetch weights and activations from electronic memory at every step, the memory energy may dominate.
Control and calibration: Maintaining the operating point of the photonic devices. Microring resonators, for example, are sensitive to temperature and require active stabilization. Phase shifters may require feedback loops. These control systems consume power, and they are often omitted from device-level energy calculations.
Cooling: If the system requires temperature control or cryogenic operation, the cooling energy must be included. For room-temperature photonic accelerators, cooling is the datacenter cooling that applies to all hardware. For cryogenic systems, it is a major additional cost that can dwarf the computation energy.
Each of these components can be significant. A system that is efficient in one component but inefficient in another may not be efficient overall.
Why device-level numbers mislead
Device-level energy figures are useful for comparing devices, but they are not a proxy for system efficiency. A polariton switch that switches with a very low control energy, for example, may still be part of a system that consumes high energy if the pump laser must be on continuously, if the device requires precise temperature control, or if the signal must be amplified at each stage to compensate for loss.
A common pitfall is to compare a device-level photonic energy figure with a system-level electronic figure. For example, comparing the energy of a single photonic MAC — counting only the optical energy — with the energy of a GPU MAC — counting the full system including memory and control — will make photonics look far better than it is in practice. A fair comparison must use the same system boundary for both.
Another pitfall is to assume that energy per operation is independent of workload. The energy per MAC depends on the matrix size, the sparsity, the precision, the batch size, and the data reuse. A photonic accelerator that is efficient for large, dense matrices may be inefficient for small or sparse ones, and vice versa. The relevant comparison is for the workload the accelerator will actually run, not for an idealized benchmark.
A third pitfall is to ignore the energy of training. Most energy discussions focus on inference, because inference is repeated many times and dominates total energy in deployed systems. Training, however, is also energy-intensive, and photonic accelerators that are designed for inference may not be suitable for training, which requires higher precision, backward passes, and frequent weight updates. A system that accelerates only inference should be evaluated on inference energy, not on training energy, and the distinction should be made explicit.
Latency, throughput, and energy: the tradeoff triangle
Energy per operation is not the only metric. It trades off against latency and throughput, and a system that optimizes one may sacrifice another.
Latency is the time to complete a single inference. It matters for real-time applications, such as autonomous driving or interactive AI, where a result must be available within a deadline. Photonic propagation is fast, but total latency includes encoding, detection, and any electronic control. A system with low energy per operation but high latency may be unsuitable for latency-sensitive workloads.
Throughput is the number of inferences per second. It matters for datacenter workloads where many requests must be served concurrently. A photonic accelerator can achieve high throughput through parallelism — many wavelength channels, many spatial modes — but throughput is limited by the input-output bandwidth and by the speed of the slowest stage.
Energy, latency, and throughput must be considered together. A common way to visualize the tradeoff is energy-delay product, which combines energy and latency into a single figure. Even that is incomplete, because it does not capture throughput or accuracy. The best evaluation is a set of measurements for a specific workload: energy per inference, latency per inference, throughput, and accuracy, all measured on the same system.
Photonic vs. electronic: where photonics can win
Photonics is most likely to win on energy per operation where data movement dominates electronic energy. In electronics, moving data between memory and compute can consume more energy than the computation itself. If a photonic accelerator can perform many operations on the same data without moving it — for example, by using wavelength multiplexing to process many channels in parallel, or by keeping weights in place in a photonic matrix — it can avoid some of that data movement energy.
Photonics is also promising where the workload is dominated by large, dense matrix multiplications with high data reuse, because the photonic core can perform many MACs in parallel with low incremental energy. It is less promising where the workload is dominated by small matrices, by sparse operations, by control flow, or by operations that are not easily mapped to photonics, such as nonlinear functions that require complex digital logic.
Excitonic devices, if they succeed, could improve the energy story by providing a low-energy nonlinearity. In a photonic accelerator without an optical nonlinearity, the signal must be converted to electronics at every layer, which adds energy for detection, conversion, and re-encoding. An efficient optical nonlinearity could eliminate or reduce that conversion energy, making the system-level energy significantly lower. The benefit depends on how much energy the nonlinear device itself consumes, including its pump and control, and on how many layers it allows the signal to stay optical.
How to evaluate an energy claim
When you encounter a claim about energy per operation for photonic AI, ask a set of questions.
What is the system boundary? Does the figure include the laser, the modulators, the detectors, the converters, the control, and the memory? If not, what is omitted, and how large is the omitted energy likely to be?
What is the workload? What matrix size, what precision, what batch size, what model? Is the workload representative of a real application, or is it an idealized benchmark chosen to favor photonics?
Is the figure measured or projected? A measured figure on a prototype is more credible than a projection that assumes ideal components, zero loss, or perfect yield. Projections can be useful for exploring potential, but they should be labeled as projections and their assumptions should be stated.
How does it compare to the incumbent? Is the comparison to a GPU or tensor accelerator measured on the same workload, with the same system boundary, and with the same accuracy? Or is it to a theoretical peak that no real system achieves?
What is the accuracy? A system that achieves low energy by sacrificing accuracy is not directly comparable to a system that maintains accuracy. The accuracy-energy tradeoff should be reported.
What is the operating condition? Is the system operating at room temperature, with realistic thermal variations, or under carefully controlled laboratory conditions? Does it require active stabilization, and is that energy included?
A claim that answers these questions clearly is worth taking seriously. A claim that does not should be treated as a hypothesis, not a result.
The path to credible energy measurements
The field needs standardized benchmarks for photonic AI energy. Standardized workloads, standardized measurement methodologies, and standardized reporting would make comparisons meaningful and would help the field progress. Efforts in this direction are emerging, including benchmarks for AI accelerators generally and for photonic accelerators specifically.
Until such standards are widely adopted, the best practice is transparency: report the system boundary, the workload, the measurement conditions, and the assumptions. Report energy per inference, not just energy per MAC, because energy per inference captures the full system. Report latency and accuracy alongside energy, because a low-energy system that is slow or inaccurate may not be useful.
For excitonic devices, the need for transparency is especially acute, because the devices are novel and the system-level implications are not yet well understood. A paper that reports a low switching energy for a polariton device should also report the pump energy, the operating temperature, the lifetime, and the signal quality, so that readers can assess whether the device could plausibly contribute to a low-energy system.
Conclusion
Energy per operation is the right metric for photonic AI, but only when it is defined at the system level and measured for a real workload. Device-level figures, however impressive, do not establish that a photonic accelerator is more energy-efficient than an electronic one. The full budget — light generation, encoding, weighting, computation, detection, conversion, memory, control, and cooling — must be counted.
Photonic AI has a plausible path to energy efficiency for certain workloads, particularly large dense matrix multiplications with high data reuse. Excitonic devices could strengthen that path by providing a low-energy optical nonlinearity. Whether they do so in practice will depend on the system-level energy of the nonlinear device, including all overhead, and on whether the system can be manufactured and operated reliably at scale.
For readers evaluating photonic AI claims, the discipline is simple: ask what is included, ask what workload, ask whether it is measured or projected, and ask how it compares to the incumbent on the same basis. The answers to those questions matter more than any single energy figure.
Sources and further reading
This article synthesizes review literature on photonic AI accelerators, on AI hardware energy, and on benchmarking. For photonic computing energy and architectures, see reviews in Nature Photonics, Nature Reviews Electrical Engineering, and journals covering integrated photonics. For AI hardware energy and data movement costs, see reviews and perspective articles on AI accelerators in journals such as Nature Electronics, Proceedings of the IEEE, and Nature Reviews Electrical Engineering. For polariton device energy and its system implications, see reviews of exciton-polariton devices in Nature Photonics and Nature Reviews Physics. No single energy-per-operation figure is asserted here, because values depend strongly on system boundary, workload, precision, and operating conditions; readers should consult the reviews above and primary experimental papers for measured figures in specific contexts, and should verify the system boundary for any figure they cite.