The Signal
Researchers propose a common benchmark for photonic simulation
A preprint proposes comparing photonic devices by the time a classical computer takes to sample their output distributions. Its authors report simulation frontiers for several kinds of light using an NVIDIA H100 GPU.

Researchers have proposed comparing photonic quantum devices by the time a classical computer takes to sample their output, using a common simulation benchmark across different kinds of light. In an arXiv preprint, the team reports results from an open-source GPU library that puts single photons, emitter-generated photons, superposition sources and squeezed light on the same scale.
On one NVIDIA H100 GPU, the authors report a simulation frontier of 38.3 single photons and 31.0 entangled photons from one emitter. Those figures use random interferometers with twice as many modes as photonic inputs and detectors that resolve photon number.
A shared measure
The proposed benchmark asks how long a classical machine takes to draw a sample from a specified device’s output distribution, allowing for floating-point rounding. Its frontier is the device size at which the fastest exact sampler the authors implemented produces one sample per second on a fixed machine.
The rate measures throughput from batches of independent samples, rather than the time between requesting and receiving an individual sample. The authors interpolate between measured integer input sizes to locate the crossing, so the reported frontier can be fractional.
The paper describes the frontier as a lower bound on classical simulation capability, not a threshold establishing that a task is hard. It covers the algorithms the team implemented; another exact method could move the frontier.
Different inputs, different results
For squeezed-light inputs with a mean photon number of 0.5, a timing model fitted to measurements places the frontier at 41.5 photonic inputs under the same mode ratio. That value uses a statistic for a run of 10,000 samples that leaves out the rarest photon numbers; the paper notes that it is not the average cost of a real run of that length.
The authors also compared sources that have the same probability of yielding a detected photon but differ in coherence. At the same brightness, their sampler reached 64.2 inputs when vacuum and a photon were mixed incoherently, as for lossy single photons, and 50.0 for coherent superpositions.
The team says that result shows a difference for its sampler, but does not establish that the difference is intrinsic to coherence. Whether a better algorithm could close the gap remains open.
For single photons, feed-forward from measurements of up to three modes lowered the reported frontier from 38.3 to 38.1. The paper also examines a non-linear site, finding a larger change for the tested case while warning that results across its phase scan are provisional because of the sampler’s branch expansion.
Checking the calculation
The library, called photonic-engines, draws samples step by step using conditional probabilities. Its samplers use a shared GPU kernel, and each was checked against a separate simulator that shares no code with it, the authors report.
The paper defines an arithmetic criterion for exact sampling: the relative error of probabilities computed at each step must stay below 10−3. It also reports that the error bound for a complete sample depends on the number of steps, and that the squeezed-light result is closer to the criterion than the single-photon, emitter and superposition results.
Alongside the measurements, the authors set out a proposed shared protocol specifying the task, input instances, size axis, timing statistic and machine. It calls for reporting the arithmetic used and distinguishing measured results from modelled ones.
Sources
Related stories
- Study finds hybrid quantum grid solver did not beat random sampling
- Researchers Model Quantum Gates in a Recoiled Free Electron
- Aaronson Says He Believes AI Proof of Quantum Conjecture Is Correct
- IonQ and EPB Report 95.8% Fewer Parameters in Grid AI Test
- D-Wave posts 21% quarterly revenue growth as deployments expand