Comparative Measurements

Comparative performance measurements against commonly used quantum simulators. These results are intended to make performance characteristics reproducible and transparent across representative workloads, not to rank projects: each simulator makes different trade-offs, and a ratio here describes one workload on one host under the controls listed below.

Every simulator replays the same gate list for each circuit, built from the prism_q::circuits generators and hashed so a rerun can prove it did the same work. The timed region is: execute the circuit and materialize the full 2^n probability vector; building the circuit from the gate list is excluded on every side. Ratios are the comparator's median over PRISM-Q's median, so a ratio above 1.00x means PRISM-Q finished sooner and below 1.00x means the comparator did; a ratio within 10% of 1.00x is reported as within band, because cross-process timing noise on one host is of that order.

Host and versions

Per-call overhead of driving a comparator from Python on a one-gate circuit, recorded so small rows can be read correctly: aer-statevector 549 us, aer-automatic 492 us, qsim 130 us.

Results

Reading the ratios: the comparators differ in design, and the design explains most of a gap. QuEST and Spinoza apply every gate as its own pass over the state and carry no gate fusion, so their time grows with the gate count; PRISM-Q, Aer and qsim fuse gates before execution, which pays most on the deep families (HEA, QV). Under automatic, Aer routes a Clifford circuit (GHZ) to its stabilizer method, as PRISM-Q routes it to its stabilizer backend, so that row compares two tableau simulations plus the dense read-out rather than two statevector runs. The GHZ rows at every size are dominated by materializing the 2^n probability vector, not by the gates. The QFT rows run the expanded textbook sequence on every simulator, PRISM-Q included, so PRISM-Q's block FFT path (which the Benchmarks page measures through the generator's QftBlock) is not exercised here and its QFT times are higher than on that page.

Summary

Counted over circuits with at least 16 qubits. A row where the comparator did not run is not counted.

Equivalence

Every simulator reproduced the reference probability vector (aer-statevector) to within 1e-05 total variation distance on every circuit. The tolerance separates a wrong answer, which lands near 1e-1, from rounding; the max TVD column shows the measured distance, and a value near 1e-6 is the single-precision comparator.

Reproducing

The harness, its adapters, and the build steps for each comparator are in comparison/README.md in the repository. The run that produced this page is committed as comparison/results/reference-<host>.json; --check against it re-runs the comparison and exits non-zero when a circuit hash, an output, or a verdict no longer matches.

python comparison/run_comparison.py --iterations 5 --qubits 16,20,24 \
  --out comparison/results/run.json --markdown comparison/results/run.md

The Benchmark Methodology page states the controls these numbers were taken under and what a comparison needs before it can be cited.