MettleQ / Apple Silicon / Quantum simulation
Local quantum-circuit simulation on Apple Silicon, with Qiskit and PennyLane integrations and reproducible CPU/GPU benchmarks.
20 qubits · depth 3
Full-state SDK calls, accuracy checked
Performance, with context
The crossover plot compares complete SDK calls. Read ratios together with width and numerical error: a faster call is useful only when the returned state also satisfies the accuracy check.
The GPU crossed Aer at 20 qubits and Lightning at 16 qubits in this measured workload. Smaller circuits can favor the CPU.
MettleQ brings quantum simulation to the Apple GPU through familiar Qiskit and PennyLane interfaces. It combines exact statevectors, bounded matrix-product states, MLX, and custom Metal kernels, with memory checks before execution.
| Qubits | Aer / MettleQ | Lightning / MettleQ |
|---|---|---|
| 16 | 0.234× | 1.453× |
| 18 | 0.801× | 3.488× |
| 20 | 1.162× | 8.371× |
Memory, contractions & other hardware
This microbenchmark compares CPU contraction, GPU-resident contraction, GPU readback, and CPU SVD across bond dimensions. At these sizes NumPy contraction is faster; SVD becomes the dominant cost. It is distinct from the full-state SDK benchmark above.
Both systems use the same six circuit families at 15–28 qubits with synchronized complex64 full-state output. MettleQ wins all six comparisons at 15 and 20 qubits; CUDA-Q is faster on all six at 28 qubits. This is a cross-machine comparison, not a claim about equivalent hardware.
The all-backend view includes CUDA-Q, Lightning GPU, Aer CPU, and Lightning CPU on the Windows machine. Compare like-for-like circuit widths and full-state contracts. The results are frozen on the MettleQ development branch.
In the supplied final crossover run, the 20-qubit Qiskit full-state call took 32.95 ms on MettleQ GPU versus 38.28 ms on Aer CPU.
The PennyLane full-state call took 19.29 ms on MettleQ GPU versus 161.52 ms on Lightning CPU at 20 qubits.
The goal is to keep existing quantum workflows usable while changing the execution engine underneath them. A Qiskit BackendV2 and registered PennyLane devices expose the simulator through each library’s normal circuit and result interfaces. Adaptive devices can select a CPU or Apple-GPU path based on the circuit, precision, and available memory.
An exact statevector stores every amplitude, so its memory cost grows exponentially with qubit count. Matrix-product states offer a different tradeoff: their useful range depends on entanglement and the allowed bond dimension. Asking an MPS backend to materialize a complete dense state removes much of that advantage, which is why this benchmark states its full-state contract explicitly.
The comparison record runs six circuit families at 15, 20, 24, 26, and 28 qubits on M3 Pro Metal and an RTX 3070 under Windows/WSL. Each uses one warm-up and three measured repetitions. The small-width results favor MettleQ, while CUDA-Q wins at 28 qubits; launch overhead and circuit structure affect the crossover. These are separate machines, and the full-state contract should be kept in view.
Build with MettleQ