Researchers report the first chip-scale, space‑time multiplexed Gaussian boson sampling (GBS) processor. GBS is a photonic experiment that sends quantum light through a network and records which detectors click. It is of interest because the output correlations are believed to be hard for classical computers to simulate. The new device is built on thin‑film lithium niobate and runs at a 4 GHz clock rate, producing detection events as large as 11,059 photons seen in a single millisecond of sampling.

To make this, the team monolithically integrated high‑speed electro‑optic modulators, on‑chip delay lines and a time‑space interferometer network on a single wafer. The chip uses Mach‑Zehnder interferometer networks to mix light across spatial channels and two on‑chip delay loops to connect different time bins. Key device numbers reported include 250‑ps time bins, a modulator half‑wave voltage of 3.18 V, electro‑optic bandwidth above 67 GHz, mean facet loss about 0.9 dB at 1550 nm, and thermo‑optic interferometers with extinction above 35 dB. The light is finally measured by superconducting nanowire single‑photon detectors (SNSPDs) outside the chip.

In experiments the authors ran sampling at five pump powers and showed the distribution of total detector clicks move toward larger photon numbers as power increased. At the highest pump setting they recorded up to 11,059 clicks in a single 1‑millisecond sampling window. The paper reports that this scale places the device beyond the reach of precise classical reproduction of the full output distribution, and it calls the system “Zhiyuan 3.0” with the largest chip‑scale GBS mode number reported by the authors.

The work matters because moving GBS from free‑space or large fiber setups onto a single chip addresses practical limits of alignment, phase stability and programmability that have hindered scaling. Space‑time multiplexing uses both temporal and spatial modes to produce a much larger effective network without a proportional increase in physical components. The authors also reconfigured the same hardware as a GBS‑powered “world model” to learn and predict a fluid flow scenario (a Kármán vortex street), and they report lower prediction error with fewer trainable readout parameters than a classical echo state network baseline.