The researchers measured the simulability frontier, defined as the device size at which their fastest exact sampler produces one sample per second on a fixed machine, and found it lies at 38.3 single photons on an NVIDIA H100 GPU. This measurement was conducted using photonic-engines, an open-source GPU library validated against a completely separate simulator, and is a lower bound on classical capability, not a hardness threshold.

Photonic-Engines Library Enables Exact Sampling of Quantum States

Photonic-engines, a newly released open-source GPU library, allows researchers to perform exact simulations of quantum states up to 38.3 single photons and 31.0 entangled photons on a fixed machine, establishing a measurable boundary for classical computational capability. The library supports simulations of single photons, entangled photon trains, vacuum-single-photon superpositions, and squeezed light, each validated against a separate simulator sharing no code. The development of photonic-engines addresses a long-standing inconsistency in estimates of when photonic devices surpass classical simulation, where previous claims for single-photon boson sampling ranged from 50 to 98 photons.

This approach acknowledges that the cost of approximate sampling is dependent on the desired accuracy, making exactness the only universally applicable criterion. For squeezed light, the frontier lies at 41.5 squeezed photonic inputs of mean photon number 0.5. Sources that emit a photon with probability one half reach 64.2 when vacuum and photon are mixed incoherently, but only 50.0 when they are in coherent superposition. Feed-forward from mid-circuit measurement of up to three modes lowers the single-photon frontier to 38.1.

Simulability Frontier Defined for Classical Simulation Benchmarking

This threshold, established through rigorous testing with the open-source library photonic-engines, indicates the point at which classical simulation takes longer than one second per sample on a fixed machine. The team’s approach deliberately focuses on exact sampling, recognizing that the cost of approximate methods is inherently tied to the desired level of accuracy. They note that while increased computing power can only logarithmically extend the frontier, a more efficient algorithm can fundamentally alter its growth rate.

Photon Limit Established on NVIDIA H100 GPU

This figure was determined using Haar-random interferometers and photon-number-resolving detectors, with the number of modes doubling the number of photonic inputs. The researchers found that the type of photonic input significantly impacts the simulation limit. Entangled photons from a single emitter lowered the frontier to 31.0 photons due to the increased computational cost of simulating their entanglement.

Squeezed Light Simulation Reaches 41.5 Photonic Inputs

Simulating photonic quantum computations with 41.5 squeezed photonic inputs is now achievable on a single NVIDIA H100 GPU, extending the practical limits of classical simulation techniques, according to new research. This benchmark, established using a timing model fitted to measurements, surpasses the 38.3 photon frontier previously observed with single photons and the 31.0 photon limit for entangled photons originating from a single emitter.

Coherence Impacts Simulation Complexity Beyond 64 Photons

The simulability frontiers of photonic quantum computing are now constrained to 38.3 single photons, according to new measurements of a “simulability frontier” established by researchers. The study demonstrates that the coherence of the photonic source dramatically impacts the ease of classical simulation. Sources emitting a photon with a 50% probability, representing a mix of vacuum and a single photon, can be simulated up to 64.2 inputs.

However, when these sources are prepared in a coherent superposition, the simulation limit drops to just 50.0 inputs, highlighting the increased complexity introduced by quantum coherence. The researchers validated each sampler against a simulator with no shared code, ensuring the rigor of their measurements and confirming the accuracy of the established frontier.

Feed-Forward Minimally Shifts Single-Photon Frontier to 38.1

Even with error correction techniques applied, the single-photon frontier is shifted by a mere 0.2 inputs with feed-forward, bringing the threshold to 38.1 detected photons. This finding demonstrates that while feed-forward offers some benefit, it is not a substantial lever for extending the range of classically simulable quantum systems, according to work using an NVIDIA H100 GPU. Demultiplexed inputs achieve one sample per second on a single H100 GPU at between 31 and 50 photonic inputs, depending on the source type.

“For our samplers, coherence between the vacuum and single-photon components of a source makes simulation markedly harder for sources that emit a photon with the same probability, whereas feed-forward on a few modes barely changes the cost,” the researchers report. This meticulous approach highlights the challenges in definitively establishing the boundary between classical and quantum computational capabilities, even as photonic quantum computing continues to advance. Arithmetic precision and memory capacity now constrain photonic quantum computer simulation, even beyond the limitations of processing speed.

👉 More information
🗞 The simulability frontiers of photonic quantum computing
✍️ Jean Senellart, Benjamin Stott, Stephen Wein, Pascale Senellart, Shane Mansfield and Pierre-Emmanuel Emeriau (Quandela, Massy, France); Esperanza Cuenca Gomez (NVIDIA)
🧠 ArXiv: https://arxiv.org/abs/2611.12876v1

More like this