Quantum kernel feature map: which simulator is fastest
Data-encoding circuit for kernel methods
8 simulators ran this circuit at 8 to 16 qubits, 12 workloads in all. Qulacs was the fastest in 4, TensorCircuit in 4 and Qiskit Aer (MPS) in 4. The tables below hold every time as it was measured, with the fastest in each row marked. Open this circuit in the interactive tool to see the hardware that could run it and what a run would cost.
Expectation value, global observable (all_z), depth 2
Time per run, by qubit count
| Qubits | PennyLane | PennyLane Lightning | Qiskit | Qiskit Aer | Qiskit Aer (MPS) | Google qsim | Qulacs | TensorCircuit |
|---|---|---|---|---|---|---|---|---|
8 qubits | PennyLane 8.00 ms | PennyLane Lightning 4.26 ms | Qiskit 2.04 ms | Qiskit Aer 1.51 ms | Qiskit Aer (MPS) 1.81 ms | Google qsim 2.02 ms | Qulacs 107 µs | TensorCircuit 288 µs |
12 qubits | PennyLane 12.7 ms | PennyLane Lightning 5.92 ms | Qiskit 4.12 ms | Qiskit Aer 3.10 ms | Qiskit Aer (MPS) 2.58 ms | Google qsim 2.92 ms | Qulacs 895 µs | TensorCircuit 871 µs |
16 qubits | PennyLane skip | PennyLane Lightning 12.9 ms | Qiskit skip | Qiskit Aer 27.6 ms | Qiskit Aer (MPS) 3.42 ms | Google qsim 5.13 ms | Qulacs 16.9 ms | TensorCircuit 30.7 ms |
Expectation value, global observable (all_z), depth 2, batch of 32
Time per run, by qubit count
| Qubits | PennyLane | PennyLane Lightning | Qiskit | Qiskit Aer | Qiskit Aer (MPS) | Google qsim | Qulacs | TensorCircuit |
|---|---|---|---|---|---|---|---|---|
8 qubits | PennyLane 10.6 ms | PennyLane Lightning 40.2 ms | Qiskit 64.7 ms | Qiskit Aer 47.6 ms | Qiskit Aer (MPS) 57.9 ms | Google qsim 64.6 ms | Qulacs 3.36 ms | TensorCircuit 1.28 ms |
12 qubits | PennyLane 76.1 ms | PennyLane Lightning 66.1 ms | Qiskit 132 ms | Qiskit Aer 98.4 ms | Qiskit Aer (MPS) 82.5 ms | Google qsim 93.4 ms | Qulacs 28.3 ms | TensorCircuit 52.5 ms |
16 qubits | PennyLane skip | PennyLane Lightning 262 ms | Qiskit skip | Qiskit Aer 869 ms | Qiskit Aer (MPS) 107 ms | Google qsim 165 ms | Qulacs 539 ms | TensorCircuit 1.09 s |
Expectation value, global observable (all_z), depth 2, batch of 256
Time per run, by qubit count
| Qubits | PennyLane | PennyLane Lightning | Qiskit | Qiskit Aer | Qiskit Aer (MPS) | Google qsim | Qulacs | TensorCircuit |
|---|---|---|---|---|---|---|---|---|
8 qubits | PennyLane 30.3 ms | PennyLane Lightning 301 ms | Qiskit 521 ms | Qiskit Aer 383 ms | Qiskit Aer (MPS) 463 ms | Google qsim 521 ms | Qulacs 26.9 ms | TensorCircuit 8.45 ms |
12 qubits | PennyLane 652 ms | PennyLane Lightning 501 ms | Qiskit 1.04 s | Qiskit Aer 781 ms | Qiskit Aer (MPS) 657 ms | Google qsim 739 ms | Qulacs 226 ms | TensorCircuit 428 ms |
16 qubits | PennyLane skip | PennyLane Lightning 2.19 s | Qiskit skip | Qiskit Aer 6.98 s | Qiskit Aer (MPS) 855 ms | Google qsim 1.32 s | Qulacs 4.31 s | TensorCircuit 8.81 s |
Expectation value, global observable (all_z), depth 2, batch of 1024
Time per run, by qubit count
| Qubits | PennyLane | PennyLane Lightning | Qiskit | Qiskit Aer | Qiskit Aer (MPS) | Google qsim | Qulacs | TensorCircuit |
|---|---|---|---|---|---|---|---|---|
8 qubits | PennyLane 110 ms | PennyLane Lightning 1.41 s | Qiskit 2.12 s | Qiskit Aer 1.51 s | Qiskit Aer (MPS) 1.84 s | Google qsim 2.10 s | Qulacs 109 ms | TensorCircuit 79.4 ms |
12 qubits | PennyLane 7.92 s | PennyLane Lightning 2.25 s | Qiskit 4.16 s | Qiskit Aer 3.14 s | Qiskit Aer (MPS) 2.64 s | Google qsim 2.98 s | Qulacs 915 ms | TensorCircuit 1.79 s |
16 qubits | PennyLane skip | PennyLane Lightning 8.28 s | Qiskit skip | Qiskit Aer 27.9 s | Qiskit Aer (MPS) 3.44 s | Google qsim 5.29 s | Qulacs 17.3 s | TensorCircuit error |
Which quantum computers can run it
At depth 3, 1,000 shots
| Quantum computer | Can run it up to | Borderline up to | Estimated cost at that size |
|---|---|---|---|
| Can run it up to 20 qubits | Borderline up to 32 qubits | Estimated cost at that size $200.30–$600.90 via AWS Braket | |
| Can run it up to 32 qubits | Borderline up to – | Estimated cost at that size $200.30–$600.90 via AWS Braket | |
| Can run it up to 12 qubits | Borderline up to – | Estimated cost at that size $23.80–$71.40 via AWS Braket | |
| Can run it up to 96 qubits | Borderline up to – | Estimated cost at that size Not priced this vendor publishes no rate card | |
| Can run it up to 48 qubits | Borderline up to – | Estimated cost at that size Not priced this vendor publishes no rate card | |
| Can run it up to 48 qubits | Borderline up to – | Estimated cost at that size Not priced this vendor publishes no rate card | |
| Can run it up to – | Borderline up to – | Estimated cost at that size – | |
| Can run it up to – | Borderline up to – | Estimated cost at that size – | |
| Can run it up to – | Borderline up to – | Estimated cost at that size – | |
| Can run it up to 32 qubits | Borderline up to 96 qubits | Estimated cost at that size from $0.015 via IBM Quantum. 0.01s at maximum throughput, so this is a floor | |
| Can run it up to 64 qubits | Borderline up to 156 qubits | Estimated cost at that size from $0.421 via IBM Quantum. 0.26s at maximum throughput, so this is a floor | |
| Can run it up to 32 qubits | Borderline up to 96 qubits | Estimated cost at that size from $0.421 via IBM Quantum. 0.26s at maximum throughput, so this is a floor | |
| Can run it up to 4 qubits | Borderline up to 32 qubits | Estimated cost at that size $0.725–$2.17 via AWS Braket | |
| Can run it up to 24 qubits | Borderline up to 48 qubits | Estimated cost at that size $1.90–$5.70 via AWS Braket | |
| Can run it up to 16 qubits | Borderline up to 20 qubits | Estimated cost at that size $1.75–$5.25 via AWS Braket |
Sizes are the grid's own: 4, 8, 12, 16, 20, 24, 32, 48, 64, 96, 128, 156 qubits. A device can run a circuit when a clean run is more likely than not, and is borderline when it is better than one in fifty. The cost is for one run of 1,000 shots at the largest size in the row, as a band and not a quote.
The simulators
Each one across every circuit family
How this was measured
One host, one harness
Every simulator ran the same circuits on one host, single-threaded and core-pinned, driven through one common harness, so any overhead the harness adds is charged to the simulator. Each result was checked against a reference implementation: a simulator that ran fast but returned the wrong answer is left out of the ranking. Nothing is interpolated or rescaled. A time was measured, or it is not shown.
Run 2026-08-10, 8 cores, threads=1. numpy 2.4.6, scipy 1.18.0, torch 2.13.0, numba 0.66.0, pennylane 0.45.1, pennylane_lightning 0.45.0, qiskit 2.5.1, qulacs 0.6.14, tensorcircuit 1.8.0, cudaq 0.15.1. One host and one thread make the comparison reproducible. It is not a prediction for your laptop.
The interactive tool shows the same measurements for any workload you pick, with the quantum hardware that could run it and what a run would cost.