MENOS

Up to 40.8× faster on Bernstein–Vazirani.

On identical circuits from 4 to 21 qubits, MENOS on a host CPU outpaced PennyLane lightning.gpu on an RTX 3050 for Bernstein–Vazirani and VQE workloads. The GPU pulls ahead on deep 20-qubit QAOA and QFT circuits.

Hardware: host x86_64 CPU (MENOS, 0 MB dedicated GPU VRAM) versus NVIDIA GeForce RTX 3050 Laptop GPU, 4 GB GDDR6 (lightning.gpu), on Ubuntu 24.04 LTS (WSL2), NVIDIA driver 592.82, CUDA 13.1, cuStateVec cu12. Methodology: median of 5 repetitions, 1 warm-up run, 1,024 shots per configuration, matched circuit definitions for both backends. These benchmarks were run on the developer's own machine; results depend on hardware, drivers and configuration. Datacenter GPUs and GPU-accelerated MENOS were not part of this scope.

Use cases

Plan first, then simulate.

Plan before you simulate

Analyze, cost, select, execute, and replan on overrun. The planner prices every candidate strategy under one cost model and sends the job down the cheapest feasible route; infeasible candidates are recorded with explicit refusal causes.

Six execution strategies

Dense, stabilizer, Pauli propagation, MPS, tensor network, and exact circuit cutting. Each is documented with its selection criteria, so you can read why a plan was chosen.

Strict budgets, exact results

Memory budgets default to 50% of physical RAM and are checked before anything is allocated. Time budgets are monitored at safe boundaries; on overrun the run halts cleanly, records the observed floor, and replans.

CPU or CUDA, by measured cost

For dense simulation the planner calibrates kernel launch and PCIe transfer costs on your machine and picks the GPU only when the estimated end-to-end time is lower. Ties go to the CPU.

Drop-in for Qiskit

MENOS ships as a native Qiskit BackendV2 and EstimatorV2. Existing circuits and estimator workflows run unchanged, with the planner behind them.

What the runtime does

• Circuit analysis and reduction: converts the circuit into its intermediate representation, removes operations outside the causal past via light-cone reduction, and identifies independent qubit subgraphs.
• Candidate strategy pricing: evaluates dense, stabilizer, MPS, Pauli propagation, tensor network, and cut plans under a unified cost model.
• Feasible plan selection: picks the cheapest feasible plan within memory and time budgets. Infeasible candidates are recorded with explicit refusal causes rather than hidden.
• Calibrated CPU and CUDA choices: selects GPU execution for dense simulation only when estimated end-to-end seconds, including data transfer latency, are lower than on the CPU.
• Strict exactness by default: every engine reachable from Python is exact. When an exact run cannot fit within budget, the job is refused rather than degraded.
• Safe monitoring and replanning: time budgets are checked at safe boundaries. On overrun, execution halts cleanly, records the observed runtime floor, and re-invokes the planner.
• Automatic factorization: qubits that never share an entangling operation are simulated as separate regions and recombined with no loss of exactness.

Getting started

MENOS is a Python package. Install it, then run a circuit through the Qiskit adapter.

pip install "menos[qiskit]"

python -c "import menos; print(menos.__version__)"

"""Bell state through the Qiskit BackendV2 adapter."""

from qiskit import QuantumCircuit
from menos.qiskit import MENOSBackend

qc = QuantumCircuit(2, 2)
qc.h(0)
qc.cx(0, 1)
qc.measure([0, 1], [0, 1])

backend = MENOSBackend()
result = backend.run(qc, shots=1000, seed_simulator=7).result()
counts = result.get_counts()
print(counts)

# A Bell pair only ever yields 00 or 11, and every shot is accounted for.
assert set(counts) <= {"00", "11"}
assert sum(counts.values()) == 1000

Specs

Version

1.2.0

License

Apache-2.0

Python

≥ 3.10

Verified on

Linux x86-64 · Windows x64

Interfaces

Qiskit BackendV2 and EstimatorV2

Engines

Dense (CPU and CUDA), stabilizer, Pauli propagation, MPS, tensor network, exact circuit cutting

Memory budget

50% of physical RAM by default; checked before allocation

GPU

NVIDIA CUDA for the dense engine, chosen by measured cost

• Benchmarks on this page were measured on a single developer machine against an entry-level laptop GPU. Numbers will differ on other hardware, drivers and configurations.

Questions

Resources

• Documentation
• Installation
• Qiskit integration
• Execution engines
• GPU execution
• Examples
• Source on GitHub