Researchers report the design and experimental tests of RP000, a small quantum processor that runs at room temperature and is built with standard silicon microchip manufacturing. The device encodes three quantum bits, or qubits, in properties of single photons that travel through the chip. The team says their measurements and simulations show RP000 outperforms classical neural networks of similar size on several machine‑learning tasks, and that it is more tolerant to noise than an equivalent superconducting device in their comparisons.

The processor is made on a 220‑nanometre silicon‑on‑insulator platform using complementary metal–oxide–semiconductor (CMOS)‑compatible fabrication. On the chip the authors place Mach–Zehnder interferometers (small on‑chip beam splitters and combiners) and thermo‑optic phase shifters (metal heaters that change optical phase when powered). Light is coupled in and out through edge couplers matched to polarization‑maintaining fiber arrays. Single photons come from a heralded C‑band source; the herald photon signals when its partner photon is on the chip. Output photons are detected by indium‑gallium‑arsenide single‑photon avalanche detectors and timed by a time‑tagger module. The package includes a thermoelectric cooler regulated by a proportional‑integral‑derivative (PID) feedback loop to keep the chip thermally stable.

To run tasks the team programs a parametric quantum circuit — sometimes called an Ansatz — made of layers of single‑qubit rotations (labeled Ry and Rz) and controlled NOT gates (CNOTs) that create entanglement. They calibrated each phase shifter by sweeping control powers, fitting the phase versus power relation, and iterating until the whole chip was under stable control. To validate calibration they set up unitaries that route a single photon deterministically to each of the eight computational basis states. For those tests the measured probability of the target output exceeded 98% for every configuration, with an average of 99.04 ± 0.01% and residual probabilities on non‑target channels below 1%. In a continuous 64‑hour campaign the target probabilities fluctuated within ±2% at three standard deviations and no recalibration was needed.