We Measured a Photonic Quantum Computer

A photonic quantum computer has no qubits. It has modes, which are paths light can take, and photons that travel down them. You build a circuit from beamsplitters and phase shifters, send single photons in, and the answer is which modes they come out of. There is no register to initialise, no gate to apply, and nothing to decompose into a gate set.

We ran one. Three times, on Quandela's Belenos processor in Paris, for about a euro each. What follows is what came back.

The experiment, and why it is the right one

Every modality has a canonical test that is simple enough to have an unambiguous answer. For gate machines it is a GHZ state, where a perfect device returns only all-zeros or all-ones. For photonics it is Hong-Ou-Mandel interference, and it is one of the strangest results in physics.

Send two identical photons into the two inputs of a 50:50 beamsplitter. Classically each photon independently goes left or right, so a quarter of the time one exits each side. Quantum mechanically the two paths that would produce that outcome cancel exactly, so it never happens. The photons always leave together.

That gives a clean fidelity measure. Ideal output is half `|2,0>` and half `|0,2>`, and the coincidence `|1,1>` at exactly zero. Any coincidence you observe is the device telling you the photons were not perfectly identical.

What came back

                     run 1        run 2        run 3
shots               10,000       10,000        5,000
|1,0,0>              4,985        5,033        2,425
|0,1,0>              4,752        4,693        2,442
|2,0,0>                117          133           61
|0,2,0>                113          121           65
|1,1,0>                 11            7            3
two-photon events      241          262          129
HOM visibility      95.44%       96.95%       97.67%
wall clock           42.5 s       44.7 s      120.8 s

Visibility above 95% on every run, and the three agree within shot noise: the gap between the first and second is 1.51 points against a combined standard error of 1.71. They do not repeat exactly, and a sampled distribution that did would be more worrying than one that does not.

For context, the same class of measurement on the gate hardware we resell returned 96% on IQM Garnet, 95% on IonQ and 80% on Rigetti. The photonic device is in the same band as the best of them.

The number that is not on any spec sheet

Look at the row for two-photon events. Out of 10,000 detections, only 241 had both photons arrive. The overwhelming majority of the time, one photon was lost somewhere between the source and the detector.

This is the defining fact about photonic hardware and it does not appear in a qubit count. Photons are lost in every component they pass through, and the loss compounds. It is also why the economics are inverted: Quandela charges EUR 0.000001 per shot, roughly four hundred times less than the cheapest superconducting shot we can buy, because a machine that discards most of its attempts has to allow you an enormous number of them.

Two figures are worth stating precisely, because they are the ones we can defend. Of detections where at least one photon arrived, 2.4% had both (241 of 10,000, and 262 of 10,000 on the repeat). And the device separately declares a transmittance of 0.053, meaning roughly one photon in twenty survives the trip.

What we cannot tell you is the overall attempt count. Those 10,000 are detected events after a filter that discards empty shots, not 10,000 attempts, so the true number of photon pairs sent in is larger and we have not established it.

There is also a discrepancy we do not fully understand, and it seems more useful to say so than to leave it out. If two photons each survive independently with probability 0.053, both arriving should happen about 0.28% of the time. We observe 2.4%, nearly ten times that. The likely explanations are that the filter changes the denominator, or that transmittance is defined per-component rather than end to end, or that the two photons' fates are not independent. We have not determined which, and any loss figure quoted from these numbers alone would be guesswork dressed as measurement.

The device tells you how good it is

This is the part that surprised us. Every job returns the processor's own noise parameters alongside the counts:

g2                     0.013     single-photon purity
indistinguishability   0.881     how identical the photons are
transmittance          0.053     photon survival

No gate processor we have access to does this. A superconducting or trapped-ion device returns counts and nothing else; if you want to know how much of the deviation was the machine, you have to work it out yourself from an ideal reference.

Belenos hands you the inputs. It states the parameters, we measure the outcome, and a fidelity certificate compares the two. That makes it the cleanest certification target in our fleet, and it is the reason a photonic engine was worth building rather than merely interesting.

Running it yourself

Both halves are live. `photonic.slos.cpu` computes the exact distribution locally, which is the reference a hardware certificate is measured against, and `qpu.quandela.belenos` submits to the device.

import perceval as pcvl
from perceval.components import BS, PERM

# photons enter modes 0 and 2; bring them together, then interfere
circuit = pcvl.Circuit(3) // (1, PERM([1, 0])) // (0, BS.H())
input_state = pcvl.BasicState([1, 0, 1])

One constraint worth knowing before you submit: Belenos has photon sources on specific modes only, `0, 2, 4, 6, 8, 9, 12, 13, 16, 18, 20, 22`. A circuit putting photons into modes 0 and 1 is refused. Our engine checks that locally and rejects before submission rather than letting the provider refuse a job you have already queued.

Pricing is Quandela's own, passed through: 0.3 credits per job plus 0.000001 per shot, where a credit is one euro. A 5,000-shot run costs 0.305 credits. The euro-to-dollar rate used is recorded in every result, so the charge can be checked against the invoice that produced it.

What this does and does not tell you

It does not rank photonic against superconducting. Hong-Ou-Mandel and GHZ are different experiments on different machines, and a single number comparing them would be invented rather than measured. Two photons on a beamsplitter is also about the simplest thing a photonic processor can be asked to do, and it says nothing about how the device behaves at twelve photons across twenty-four modes, which is where the interesting work is.

What it does show is that a photonic processor is reachable, affordable, and accurate enough on this test to sit alongside the gate hardware, and that it reports enough about itself to be certified rather than merely trusted. The wider comparison across all four modalities, and what it means that no single circuit can run on all of them, is the next thing we are writing.