Title:

What Circuits Do You Actually Need? On-Demand Benchmark Generation with MQT Bench

Abstract:

Rigorous benchmarking underpins progress across the quantum computing stack: compiler developers need comparable input circuits, hardware designers need realistic workloads for connectivity and noise analysis, and machine-learning workflows need large, well-annotated circuit datasets. Existing benchmark suites, however, are typically fixed libraries, or target only a single abstraction level, and adapt poorly as algorithms, devices, and research questions evolve.

This talk presents the second major version of MQT Bench, which replaces the fixed corpus of its predecessor with programmable, unbounded circuit generation. Users compose three modular building blocks—algorithmic templates, gatesets, and device models—to synthesize exactly the circuits a given study requires, at four abstraction levels ranging from abstract logical circuits to executable, device-mapped implementations. All components are user-extensible, and circuits are available both through the Python package mqt.bench and on demand via the web frontend at https://mqt-bench.app.

We walk through three representative use cases: comparing quantum compilers on identical inputs, evaluating a custom hardware architecture before it is built, and generating device-labeled datasets for machine-learning models.

We close by noting a class of programs that cannot be captured as flat circuits at all—those with dynamically-bounded loops or conditionals on measurement results—and what this implies for benchmark design.