mlx-quantum-sim

GPU-accelerated quantum circuit simulator for Apple Silicon, with real hardware noise models.

Why I Built This

I had a quantum algorithm idea. Wasn't sure if it worked.

The options were:

• Rent an A100 on the cloud → $2-5 per experiment, maybe $2,000/year
• Buy an RTX 4090 → $1,600 upfront
• Use my MacBook Pro (M1 Max, already on my desk) → $0

I chose the MacBook.

The idea might fail. Six of my first experiments did fail. If I'd rented cloud GPUs for those, I'd have wasted hundreds of dollars discovering dead ends.

Instead, I validated everything locally for $0. When something finally worked, then it made sense to consider scaling up.

This simulator exists because: you shouldn't have to pay to find out if your idea is bad.

What This Is

A quantum circuit simulator that runs on Apple Metal GPU via MLX. Includes calibrated noise profiles from Google Willow, IBM Heron, and QuTech Tuna-9 — extracted from publicly available calibration data.

Honest Comparison: Apple Silicon vs NVIDIA

All qubit counts assume complex64 (float32) precision. Subtract ~1 qubit for complex128.

	M1 Max (64GB)	M2 Ultra (192GB)	RTX 4090 (24GB)	A100 cloud (80GB)	Colab T4 (free, 16GB)
Max qubits (float32)	~32	~34	~31	~33	~30
Speed (30q circuit)	~5 ms	~4 ms	~0.3 ms	~0.3 ms	~1 ms
1000 circuits (30q)	~5 sec	~4 sec	~0.3 sec	~0.3 sec	~1 sec
Power per experiment	0.06 Wh	0.07 Wh	0.04 Wh	N/A	N/A
Extra hardware cost	$0 (if you own a Mac)	$5,000	$1,600	$1.10/hr	$0
Setup time	pip install mlx	Same	Install CUDA + drivers	Cloud setup + SSH	Open browser
Offline / private	✅	✅	✅	❌ (cloud)	❌ (cloud)
Other uses	Daily work	Daily work	ML / gaming	ML / training	ML (limited)

Where Apple Silicon Wins

• Zero marginal cost. If you own a Mac, every experiment is free. No cloud bills, no GPU purchase.
• Large unified memory. M2 Ultra 192GB → 34 qubits. No single consumer GPU matches this.
• Fast iteration. Edit code → run → results. No uploading, no environment setup, no waiting for cloud instances.
• Privacy. Circuits never leave your machine.

Where NVIDIA / Cloud Wins

• Raw speed. 10-15x faster per circuit. For production workloads, this matters.
• Ecosystem. cuQuantum, qsim, Qiskit Aer CUDA — mature and highly optimized.
• Scale. Multi-GPU and cluster computing. Apple can't do this.
• Free tier. Google Colab gives you a T4 GPU for free. Honest alternative for small experiments.

Why Not Just Use Google Colab?

Colab is a strong free option for small experiments. Use mlx-quantum-sim when you need:

• Offline access (no internet required)
• More than 16GB GPU memory (Colab T4 limit = ~30 qubits)
• Long-running experiments (Colab disconnects after ~12 hours)
• Privacy (proprietary quantum circuits stay local)
• Reproducibility (no session timeouts or random disconnects)

The Real Decision

"I have a quantum algorithm idea. Should I rent cloud GPUs?"

If you're EXPLORING (not sure it works):
  → Use your Mac. It's free. Most ideas fail. Save your money.

If you've VALIDATED (it works, need to scale):
  → Switch to NVIDIA / cloud. Pay for speed when you know it's worth it.

Apple Silicon = lab notebook (cheap, fast iteration)
NVIDIA = production machine (expensive, maximum throughput)

When to Use This vs NVIDIA

Scenario	Use mlx-quantum-sim	Use NVIDIA/cloud
Testing a new quantum algorithm	✅	Overkill
Homework / coursework	✅	Overkill
Prototyping with noise models	✅	Overkill
Publishing a paper (small circuits)	✅	Optional
Running 100K+ circuits	❌ Too slow	✅
Production / deployment	❌	✅
Training quantum ML models	⚠️ (small scale OK)	✅ (large scale)
33+ qubit simulation	✅ (M2 Ultra only)	❌ (need multi-GPU)

Features

• Metal GPU acceleration: All gate operations on Apple Silicon GPU — no CPU round trips
• Real hardware noise: Google Willow (105 qubits, 364 CZ pairs), IBM Heron, QuTech Tuna-9
• Stochastic trajectory simulation: Same approach as Google's qsim
• Up to ~33 qubits on M2 Ultra 192GB (statevector, O(2^n) memory)

Quick Start

pip install mlx

from mlx_quantum_sim import MLXQuantumSimulator
from noise_profiles import WILLOW_NOISE

# Ideal simulation
sim = MLXQuantumSimulator(3)
sim.h(0)
sim.cx(0, 1)
sim.cx(1, 2)
probs = sim.measure_probs()  # GHZ state

# With Google Willow noise
sim = MLXQuantumSimulator(3, noise_profile=WILLOW_NOISE, n_trajectories=100)
sim.h(0)
sim.cx(0, 1)
probs = sim.measure_probs()  # Fidelity < 1 due to realistic noise

Noise Profiles

Profile	Source	CX Error	T1	T2
WILLOW_NOISE	cirq-google calibration (willow_pink)	0.34%	70 μs	49 μs
HERON_NOISE	IBM published specs	0.50%	100 μs	100 μs
T9_NOISE	QuTech published specs	1.0%	50 μs	20 μs

Full per-qubit Willow calibration (T1, T2, readout errors, 364 CZ pairs) in willow_calibration.json.

Noise Channels

• Depolarizing (single and two-qubit)
• Amplitude damping (T1 decay)
• Phase damping (T2 dephasing)
• Thermal noise (gate-time dependent)
• Measurement readout error

Gates

H, X, Y, Z, S, T, Rx(θ), Ry(θ), Rz(θ), CNOT/CX, CZ, SWAP

API: sim.ry(theta, qubit) — angle first, qubit second.

Benchmark (M1 Max, 3 qubits, 33 gates)

Mode	Time/circuit
Ideal (no noise)	2.7 ms
Willow (100 trajectories)	495 ms

Mirror Circuit Fidelity

Depth	Ideal	Willow	Heron	T-9
1	1.000	0.959	0.966	0.922
5	1.000	0.942	0.865	0.722
10	1.000	0.865	0.861	0.490

Ordering matches hardware quality: Willow > Heron > T-9. ✓

Contributing

PRs welcome — especially:

• CUDA/PyTorch backend (for NVIDIA users)
• Additional noise profiles (Rigetti, IonQ, Quantinuum)
• Density matrix simulation
• Performance optimization

License

Apache License 2.0 (consistent with cirq-google). See LICENSE.

Citation

@software{mlx_quantum_sim,
  author = {Huang, Sheng-Kai},
  title = {mlx-quantum-sim: GPU-accelerated quantum simulator for Apple Silicon},
  year = {2026},
  url = {https://github.com/akaiHuang/mlx-quantum-sim}
}

Acknowledgments

Noise calibration data from Google's cirq-google (Apache 2.0). Trajectory-based noise simulation follows the approach of Google's qsim.