In short: Quantum Volume (QV) is a single hardware-agnostic number that rolls qubit count, connectivity, gate fidelity, and measurement error into one figure — a machine can only claim QV = 2n if it reliably runs a random "square" circuit on n qubits and n layers deep, so it rewards balanced quality over a raw qubit headline.
Why a raw qubit count lies to you
Ask a newcomer which quantum computer is "best" and they will quote qubit counts: 127 here, 1,121 there. But a 1,000-qubit machine whose two-qubit gates err 2% of the time cannot finish a circuit of any useful depth — errors compound multiplicatively, and the state decoheres into noise long before the answer appears. A 27-qubit machine with 99.9% gate fidelity and all-to-all connectivity may solve problems the bigger device cannot touch. We needed a yardstick that captures this, the way a CPU benchmark captures more than clock speed. In 2019 IBM proposed exactly that: Quantum Volume.
QV is defined on square circuits: the largest m for which a device can run a random circuit on m qubits to depth m and still beat a statistical threshold. The reported figure is 2m, so QV grows exponentially while the achievable width/depth grows only linearly.
How the protocol actually works
A QV test builds model circuits: for each of the m⁄2 layers, qubits are randomly paired, and each pair gets a Haar-random SU(4) two-qubit unitary. These are deliberately hard — no clever structure a compiler can exploit. You:
- Generate many random square circuits of width = depth = m.
- Compute the ideal output distribution on a classical simulator.
- Identify the heavy outputs — bitstrings whose ideal probability is above the median.
- Run each circuit on real hardware thousands of times and measure the heavy output probability (HOP).
A noiseless circuit yields a heavy-output probability that converges to about 0.85 (ln 2 + 1⁄2 asymptotically). The pass criterion: the device must exceed 2⁄3 (≈0.667) heavy outputs with 97.5% statistical confidence. The biggest m that passes sets QV = 2m. Because a transpiler must map the random SU(4) blocks onto the device's native gates and real topology, QV implicitly tests the whole stack — routing, calibration, crosstalk, and readout — not just isolated gate numbers.
Watching heavy-output probability converge
The intuition behind the 2⁄3 bar is best seen in motion: as you average more shots, the measured HOP settles. On a good device it parks above the pass line; on a noisy one it sinks toward the 0.5 "pure noise" floor.
Comparing real machines
Because QV is vendor-neutral, several hardware families report it directly. Trapped-ion devices, with all-to-all connectivity and long coherence times, have pushed QV far higher per physical qubit than early superconducting chips, which pay a routing tax for nearest-neighbor layouts.
| Platform (representative) | Qubit tech | Reported QV | Effective width m |
|---|---|---|---|
| IBM "Montreal"-class (2020) | Superconducting | 64 | 6 |
| Honeywell/Quantinuum H1 (2021) | Trapped ion | 1,024 | 10 |
| Quantinuum H2 (2023+) | Trapped ion | 219+ | 19+ |
Note the exponent: moving from QV 64 to QV 1,024 is only +4 qubits of usable square-circuit width, but a 16× larger reported number. This is why QV comparisons must be read as log scale — and why it eventually saturates for large, error-corrected machines.
Where QV stops being useful
QV requires classically simulating the ideal distribution to find heavy outputs, so it becomes impractical beyond roughly 30–50 qubits — you can no longer compute the ground truth. That ceiling is why IBM later complemented QV with CLOPS (circuit layer operations per second, a speed metric) and why the field increasingly reports application-level or error-corrected logical-qubit figures. QV remains the cleanest apples-to-apples number for today's NISQ-era machines, and a useful teaching tool for understanding how fidelity, connectivity, and depth trade off — the same balance-of-errors thinking that governs T1/T2 coherence budgeting.
Students coming from chip design often find the analogy natural: just as a tapeout's real limit is timing closure across the whole netlist — not the transistor count on the datasheet — QV measures what the full stack can actually execute. Teams that build those EDA skills at ChipXpert's VLSI training tend to grasp QV's systems view quickly.
Want to go from "what is a qubit" to reading QV benchmarks critically? Our quantum computing track on CourseTron builds the fidelity, connectivity, and error-model intuition this metric rewards.
Frequently asked questions
Is a higher Quantum Volume always better?
Generally yes among comparable NISQ devices, because QV captures balanced quality rather than raw size. But QV saturates and becomes unmeasurable past ~50 qubits, so for large or error-corrected systems you should look at logical-qubit counts, logical error rates, and speed metrics like CLOPS instead.
Why is the pass threshold 2⁄3 and not something higher?
An ideal square circuit's heavy-output probability asymptotes near 0.85, while a fully depolarized (noisy) output gives 0.5. The 2⁄3 bar, applied with 97.5% confidence, sits comfortably between those and marks the point where the device is provably doing better than noise.
Does QV depend on the device's connectivity?
Yes. The random SU(4) layers must be transpiled onto the real chip, so nearest-neighbor architectures incur extra SWAP gates and thus more error, lowering achievable QV. All-to-all platforms like trapped ions avoid this routing tax, which is a major reason they report higher QV per physical qubit.
Can I measure QV on a simulator?
You can run the full protocol on a noisy simulator — Qiskit even ships a QV experiment class — but a perfect noiseless simulator trivially passes every width, so QV is only meaningful against a device (or model) that carries realistic noise.