Maestro: Intelligent Execution for Quantum Circuit Simulation

TL;DR

The paper tackles the fragmentation of quantum circuit simulators by introducing Maestro, a unified, extensible interface that orchestrates multiple backends via a common intermediate representation and a predictive runtime model. It integrates state-vector, MPS, tensor-network, stabilizer, GPU-accelerated, and p-block methods, with GPU paths and distributed computing support, optimized for HPC environments. A data-driven backend-prediction engine selects the fastest simulator for each circuit, enabling efficient batched and distributed workloads. Benchmark results show significant throughput gains across single and multi-circuit scenarios, validating Maestro’s utility for hybrid quantum-classical workflows and large-scale HPC deployments.

Abstract

Quantum circuit simulation remains essential for developing and validating quantum algorithms, especially as current quantum hardware is limited in scale and quality. However, the growing diversity of simulation methods and software tools creates a high barrier to selecting the most suitable backend for a given circuit. We introduce Maestro, a unified interface for quantum circuit simulation that integrates multiple simulation paradigms - state vector, MPS, tensor network, stabilizer, GPU-accelerated, and p-block methods - under a single API. Maestro includes a predictive runtime model that automatically selects the optimal simulator based on circuit structure and available hardware, and applies backend-specific optimizations such as multiprocessing, GPU execution, and improved sampling. Benchmarks across heterogeneous workloads demonstrate that Maestro outperforms individual simulators in both single-circuit and large batched settings, particularly in high-performance computing environments. Maestro provides a scalable, extensible platform for quantum algorithm research, hybrid quantum-classical workflows, and emerging distributed quantum computing architectures.

Figures (8)

• Figure 1: Architectural overview of the Maestro interface. User circuits are ingested into a common Intermediate Representation (IR). The Prediction Engine analyzes circuit features (e.g., entanglement entropy, gate density) to dynamically route execution to the optimal simulation backend.
• Figure 2: Predicted vs actual optimal simulators confusion matrix. To account for possible variance in execution time due to the run time environment in cases where multiple simulators are of the same quality, we choose to validate simulations where Maestro was off by less than $0.01$ s, with that error being no more than $10\%$ of the optimal runtime.
• Figure 3: Statevector-simulable task runtimes by backend, for a selection of circuit types. The GPU simulation used a NVIDIA T4 GPU and 4 vCPUs with 15 GB of memory, while the rest were run exclusively on 16 vCPUs with 124 GB of memory. 1,000 shots were captured.
• Figure 4: Runtimes of MPS simulations for different simulators. The fidelity threshold for acceptance is set at $0.95$. For each circuit, the bond dimension is initialized at four. It is then iteratively doubled in subsequent runs until the acceptance threshold is met. Only the runtime of the final, successful iteration is recorded. 1,000 shots are taken for all simulations.
• Figure 5: Time required to execute the whole circuit batch for each of the policies. A fidelity threshold of $0.95$ was required for MPS simulations, and run times include sampling $1000$ shots per circuit.