Exponential cost, measured
Runtime and memory roughly double per added qubit. Around 30 qubits an exact state vector needs 16 GB; at 50 it would need 16 PB.
Benchmark Lab · live
Benchmark your own device.
Every qubit doubles the memory an exact simulation needs. Run the draft QAISS Benchmark Protocol here and watch where your machine hits the wall — with every number labelled by hardware, version, workload, precision, and date.
Live · runs locallyNo data leaves your device
Protocol · qaiss-bench-v0-draft
brickwork: H layer, then 4 × (RY(θ) on every qubit + CZ on alternating pairs); θ from seeded RNG (seed 2026). 3 repetitions, median wall-clock of state evolution (sampling excluded; small sizes batched to ≥ 4 ms).
Ready. Larger sizes can take a few seconds and use up to ~32 MB.
Results
| Qubits | Gates | Median runtime | State-vector memory |
|---|---|---|---|
| Run the benchmark to fill this table. | |||
Compare devices
Load up to two benchmark records: from a phone, another laptop, a colleague, or the Python connector (python -m qaiss_cuquantum bench --backend qpanda3, also reference, tn-reference, gpu, gpu-tn). Only records produced with protocol qaiss-bench-v0-draft are accepted — anything else is not comparable.
Run the benchmark above, then load records to compare.
WHAT IT SHOWS
WHAT IT DOESN’T
Not a ranking
Results depend on your browser, power settings, and other tabs. JavaScript is far slower than optimised C++ or GPU libraries. Compare runs only with matching metadata.
WHY IT MATTERS
Why backends differ
This wall is why GPU-accelerated simulation, tensor-network methods, and real quantum hardware exist — and why QAISS wants to compare them with honest metadata.
Technology foundations