Benchmarks

Every simulator ran the same circuit: five layers of a Hadamard on every qubit, an RZ rotation on every qubit and a chain of CNOTs, then 1000 measurement shots. Times are the best of at least three runs.

Results

The pure Python loop was stopped after 18 qubits, when one run passed 10 seconds.

What the numbers say

• Below 14 qubits the CPU backend is the fastest: 8 to 50 times faster than Qiskit Aer, because a call has almost no overhead. Aer’s time there is mostly its fixed per-call cost (about 2 ms to submit a job and build its result), so this measures call overhead more than simulation.
• From 14 qubits the GPU backend is the fastest: 3 to 8 times faster than Aer. The GPU backend computes in 32-bit floats and Aer ran in 64-bit; Aer’s precision="single" option would be the like-for-like comparison.
• The CPU backend stays ahead of Aer up to 19 qubits and falls behind from 20. At 20 qubits a 64-bit state is 16 MiB, the size of this CPU’s L3 cache, and the CPU backend’s time per extra qubit jumps from about 2× to 3–4×, so the likely cause is the state no longer fitting in cache.
• There is one benchmark circuit. qsengine merges the H and RZ on each qubit into one pass over the state, which suits this circuit; other circuits may compare differently.
• Against a plain Python loop, the CPU backend is about 100 times faster at 10 qubits and 840 times faster at 18.

Setup

AMD Ryzen 5 5500 (6 cores), AMD Radeon RX 6600, Windows 10, Python 3.14, Qiskit Aer 0.18.2 with its default statevector method in 64-bit precision. These are one machine’s numbers; run the script to get yours.

pip install qiskit qiskit-aer matplotlib
python benchmarks/benchmark.py --max-qubits 22

The script and the raw results.csv are in the benchmarks folder of the repository.