cuStateVec
State-vector simulation with optimised gate kernels and memory management.
Connector backend written (gpu.custatevec.fp64). Not yet executed on NVIDIA hardware.
PrototypeGPU path · cuQuantum connector
Run anywhere. Verify in the browser.
A Python connector re-runs any Workload Passport on NVIDIA cuStateVec, NVIDIA cuTensorNet or a CPU reference — and the result is checked bit-for-bit by the verifier in the Lab. Portability you can prove, not just claim.
CPU state vector · parity verifiedCPU tensor network · parity verifiedcuStateVec · code ready, untested on GPUcuTensorNet · code ready, untested on GPU
One passport, any backend
The portability loop
- 01
Build in the browser
Design a circuit in the Lab and download its Workload Passport.
- 02
Run on a backend
The connector executes it on cuStateVec or cuTensorNet (GPU), or a CPU reference.
- 03
Emit a new passport
Same circuit hash, backend recorded, output hashed the same way.
- 04
Verify in the browser
Paste it into the Lab verifier: it re-runs the workload and compares every count.
Evidence
Parity, measured
Reference runs exported from the browser simulator, re-computed by the Python connector. Identical output hashes mean identical results, count for count.
| Workload | Qubits | Shots | Seed | Noise | Output SHA-256 (browser = CPU state vector = CPU tensor network) | cuStateVec | cuTensorNet |
|---|---|---|---|---|---|---|---|
| bell | 2 | 2048 | 42 | none | ✓ 0f4333419a1da909… | not run | not run |
| ghz | 3 | 2048 | 42 | none | ✓ 61f7e8a04335409b… | not run | not run |
| rotations | 3 | 4096 | 7 | none | ✓ 9097b7eee36bfca0… | not run | not run |
| noisy_ghz | 3 | 1000 | 123 | readout 0.05 | ✓ dd1947c0eca8176a… | not run | not run |
Why the GPU columns are empty. The connector’s cuStateVec and cuTensorNet backends follow NVIDIA’s public Python API, but they have not been executed yet — the QAISS team had no NVIDIA GPU available. It will be marked verified only after the GPU parity test passes on real hardware, with the device and driver recorded here. Have a GPU? Help us run it.
Passport → cuQuantum
See the exact GPU code
Pick or edit a circuit. This page builds the same objects the connector hands to NVIDIA’s libraries: a gate-by-gate cuStateVec program, or the tensor network cuTensorNet contracts. Generated in your browser; nothing runs on a GPU here.
Memory planner
How far can a GPU go?
An exact state vector stores 2ⁿ complex amplitudes, so memory doubles with every qubit. Upper bounds only: libraries also need workspace. Tensor-network memory depends on the circuit’s entanglement and contraction path, not just its width.
—complex128 (fp64, what the connector uses)
—complex64 (fp32)
| Device memory | Fits fp64? | Max qubits fp64 | Max qubits fp32 |
|---|---|---|---|
| 16 GB GPU | — | 29 | 30 |
| 24 GB GPU | — | 30 | 31 |
| 48 GB GPU | — | 31 | 32 |
| 80 GB GPU | — | 32 | 33 |
| 141 GB GPU | — | 33 | 34 |
| 192 GB GPU | — | 33 | 34 |
Use
Three commands
Pure-Python reference with no dependencies; the GPU extra installs NVIDIA’s cuQuantum Python bindings and CuPy.
# CPU reference (any machine)
python -m qaiss_cuquantum run passport.json --backend reference -o out.json
# Same circuit as a tensor network (CPU, NumPy)
pip install ".[tn]"
python -m qaiss_cuquantum run passport.json --backend tn-reference -o out.json
# NVIDIA GPU (Linux, CUDA 12/13)
pip install ".[gpu-cu12]"
python -m qaiss_cuquantum run passport.json --backend gpu -o out.json # cuStateVec
python -m qaiss_cuquantum run passport.json --backend gpu-tn -o out.json # cuTensorNet
# Time it — load record.json at /benchmarks to compare with your browser
python -m qaiss_cuquantum bench --backend gpu --max-qubits 26 -o record.json
# Prove it — then paste out.json into /lab → Verify
python -m unittest discover -s testsNVIDIA cuQuantum libraries
What each one could do for QAISS
Library descriptions summarise NVIDIA’s public documentation. QAISS usage is our plan, with honest status.
cuTensorNet
Tensor-network contraction with path optimisation and approximate methods.
Connector backend written (gpu.cutensornet.fp64); the network it contracts is verified on the CPU. Next: amplitudes and samples beyond state-vector memory.
PrototypecuDensityMat
Density-matrix simulation with arbitrary operator actions.
Planned: richer noise models than the current readout bit-flip.
PlannedcuPauliProp
Pauli propagation for efficient expectation values.
Planned: VQE / QAOA energy estimates in the Quantum AI Lab.
PlannedcuStabilizer
Clifford-circuit simulation for error-correction research.
Planned: stabilizer benchmarks at hundreds of qubits.
PlannedOfficial resources
Read NVIDIA’s docs
Independent project. NVIDIA, cuQuantum, cuStateVec and CUDA-Q are trademarks of NVIDIA Corporation. The connector does not redistribute NVIDIA software; users install it from NVIDIA’s packages under NVIDIA’s licence terms. No partnership or endorsement is implied.