Tightly integrating a GPU and a QPU for fast calibration of multi-qubit circuits

With the scaling up of quantum processing units, their many calibration protocols and the forthcoming of large-scale quantum error correction circuits, integrating scalable and flexible classical computer resources within quantum sequences has become a priority.

In this work, we demonstrate the tight integration of a reinforcement learning agent, residing in a Grace Hopper superchip, with a real superconducting quantum computer. We show that the agent enables learning the optimal circuit drive and readout policies for the multiqubit device, resulting in the reduction of overall circuit execution errors.

Any improvement in the fidelities associated with QEC stabilizer circuits will result in exponential reductions in logical qubit errors. This strongly motivates the continuous calibration of such circuits on timescales shorter than the hardware drift rates, allowing for reductions in QEC overheads, a necessary step towards large scale logical quantum systems.