Investigating parallel execution of quantum Machine Learning circuits on superconducting hardware

Quantum Machine Learning algorithms generally rely on hybrid implementation with a classical learning loop running on a CPU and small quantum circuits running on a Quantum Processing Unit (QPU) for serialized processing batches of data. Considering the scarcity of quantum resources, parallelizing the execution of multiple instances of small quantum circuits instead of allocating a single circuit would better use quantum hardware resources. However, exploiting all available qubits in QPUs could result in a noisier operating condition and disrupt or slow the learning mechanism. This paper investigates the parallelization of QML algorithms on QPU for data clustering based on trainable generative models. We design a parallel macro-circuit with the Qiskit framework and measure their performance on an IBM superconducting quantum computer. A theoretical performance model is then proposed to determine the expected level of acceleration. Finally, we measure the impact of the QPU's occupancy on the loss functions of our QML algorithm, enabling us to identify the most reliable and exploitable acceleration range.