Quantum computational advantage is a race with an open finish: as soon as a quantum machine demonstrates a result, classical algorithms rush to catch up, sometimes reducing the gap to laughable margins. Around photonic experiments on Gaussian boson sampling (GBS), a whole zoo of simulators has emerged, from greedy samplers to methods based on matrix product states (MPS). The main Achilles' heel of photonics is photon loss, which, as recent works have shown, can radically simplify classical simulation. The new work joins this line but with a different scale: 1024 input squeezed states and 8176 output modes, an order of magnitude more than predecessors. Neighboring papers on the archive explore quantum algorithms, error correction, and energy efficiency, but here the bet is placed on sheer scale and suppression of losses at the source level.
The heart of the setup is four optical parametric generators pumped by laser pulses of 1.6 ns duration at a wavelength of 775.7 nm. They generate single-mode squeezed states, which then pass through three cascaded unbalanced Mach-Zehnder interferometers to filter out parasitic modes; the transmission of such a system reaches 99.8%, and the overall source efficiency is 92%. The squeezed states are directed into a programmable circuit with hybrid spatial-temporal encoding: three 16-mode interferometers are connected by two arrays of fiber delay lines with a step of 50 ns. This architecture provides a cubic growth in connectivity ($16^3$ = 4096) with a linear growth in physical resources. The output photons are detected by 16 superconducting nanowire detectors with an efficiency of 93% and a recovery time of 43 ns; the overall system efficiency is 51%. The phase of the entire circuit is stabilized to an accuracy of λ/200 at a wavelength of 1550 nm.
Experiments were conducted in three configurations: S64 (64 inputs, 4336 modes), M256 (256 inputs, 5104 modes), and L1024 (1024 inputs, 8176 modes). In the largest series, up to 3050 photon coincidences were registered, more than an order of magnitude higher than previous records. The photon number distribution matches the theoretical model of 'true' GBS and sharply diverges from all classical simulators: squeezed state, thermal state, and distinguishable photons. A Bayesian test against the squeezed state shows growing confidence in favor of the quantum origin of the sample as the considered subsystem size increases. Comparison of second- and third-order correlation functions rules out the greedy sampler, the independent pair and single photon sampler (IPS), and the treewidth-based sampler. For the most powerful classical method to date, the MPS algorithm proposed in 2024, it is shown that for any realistic parameters (bond dimension χ and effective number of squeezed photons $N_{eff}$), it cannot reproduce the experimental statistics. For the L1024 set, the required χ exceeds 8×$10^{21}$, and the estimated simulation time on the El Capitan supercomputer is more than $10^{42}$ years; the quantum processor produces a sample in 25.6 microseconds.
This work stands apart from its archive neighbors, which mostly explore algorithmic aspects of quantum computing or energy efficiency. Here, instead, is a hardware breakthrough: for the first time, GBS has been scaled to thousands of modes while maintaining an overall efficiency of 51% and preventing photon losses from devaluing the result. Unlike previous demonstrations where classical simulators narrowed the gap, Jiuzhang 4.0 leaves them no chance even in theory: the estimated speedup of $10^{54}$ times makes practical simulation impossible on any conceivable resources. This is not just another record but a qualitative leap in the reliability of evidence for quantum advantage in photonic platforms.
The next logical step is to use the same architecture to generate massive three-dimensional cluster states needed for measurement-based fault-tolerant quantum computing. Hybrid spatial-temporal encoding already allows controlling thousands of highly entangled modes, and further improvements in sources and detectors could push efficiency and photon numbers even higher. Perhaps new classical algorithms targeting this specific architecture will soon appear, so the race will continue, but now the quantum side has a solid safety margin.
The results directly affect the field of quantum computing and quantum information, setting a new standard for photonic experiments. They also stimulate the development of more sophisticated classical simulation algorithms and verification methods.
Immediate next steps include using the platform to generate cluster states and test elements of fault-tolerant computation. In parallel, we can expect new attempts at classical simulation targeting hybrid circuits.
For photonic quantum computing, the key unresolved problem remains combating losses and achieving the fault-tolerance threshold. This work shows that scaling at fixed efficiency is possible, but practical error correction is still far off: even lower losses and integrated sources are needed.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
This paper presents the photonic processor 'Jiuzhang 4.0' with hybrid spatial-temporal encoding, which achieved 3050 photon coincidences in a configuration with 1024 inputs and 8176 modes. The authors used high-efficiency squeezed light sources with 99.8% transmission and an overall system efficiency of 51%. Experiments showed that classical simulators, including MPS, do not reproduce the device's statistics, and the estimated simulation time on the El Capitan supercomputer exceeds 10^42 years.
Where its strength is
The key strength is the scale and loss suppression: a record number of photon coincidences with high source efficiency, which closes the loophole for classical simulation and strengthens the evidence for quantum advantage.
Where it could go next
Worth trying
Why this may matter
Such scaling of photonic platforms with loss suppression paves the way for practical quantum computing, and the potential simplification of sources and improved accuracy could make the technology more accessible.
Nearby works in our archive
The archive is sparse around this work: the closest review is only 0.52, against a usual 0.64. It looks like few have drilled nearby.
The number of possible photon configurations in the L1024 experiment is estimated at $10^{2461}$, which is larger than the number of atoms in the observable universe (about $10^{80}$).
The idea of quantum supremacy echoes Greg Egan's story 'The Halo', where characters use quantum computing to model consciousness. Here, though, the task is more down-to-earth but no less impressive: proving that a quantum machine can be faster than any classical one.