Go further · Chapter 11 · 11b · A step deeper
Why doesn't qubit count tell you how good a quantum computer is?
The short answer (tap to open)
A quantum computer's usefulness depends on how many operations it can run before errors swamp the answer, not just on how many qubits it has. A small machine with very accurate qubits can outperform a big noisy one. Benchmarks like quantum volume try to capture both size and quality in one number.
Why can fewer qubits win?
Every operation risks an error (chapter 8). A program on 1,000 qubits at 99% accuracy fails almost at once: a few hundred operations and the answer is noise. A program on 50 qubits at 99.9% can run ten times longer. Useful work needs both enough qubits and enough operations, and the second depends on accuracy.
Machine A: 1,000 qubits at 99%
8% chance of a clean run: the answer is mostly noise.
Machine B: 50 qubits at 99.9%
78% chance of a clean run: it works.
Illustrative machines. A clean run means no error in any two-qubit operation; about one operation per pair of qubits per layer.
Three program sizes, in words
- Qubits the program uses: 100; Layers of operations: 1. Machine A: 1,000 qubits at 99%: 61% chance of a clean run: it works. Machine B: 50 qubits at 99.9%: Can't run it: too few qubits.
- Qubits the program uses: 50; Layers of operations: 10. Machine A: 1,000 qubits at 99%: 8% chance of a clean run: the answer is mostly noise. Machine B: 50 qubits at 99.9%: 78% chance of a clean run: it works.
- Qubits the program uses: 20; Layers of operations: 50. Machine A: 1,000 qubits at 99%: 1% chance of a clean run: the answer is mostly noise. Machine B: 50 qubits at 99.9%: 61% chance of a clean run: it works.
Illustrative machines. A clean run means no error in any two-qubit operation; about one operation per pair of qubits per layer.
What do benchmarks measure?
- Quantum volume: the biggest square program (equal qubits and layers) that runs reliably. It rewards balance. Record: 2²⁵ (Quantinuum, September 2025).
- Error per layered gate: IBM's measure of the average error across a layer of operations on many qubits.
- Algorithmic qubits: IonQ's measure across a set of real-style programs. Record: #AQ 64 (September 2025).
- Application benchmarks: independent suites, such as the US QED-C's, which run small versions of chemistry, optimisation and simulation tasks.
- Logical qubits and logical error rate: the measures that will matter most once error correction scales (chapter 8).
What should you be wary of?
Benchmarks a company designs itself, records on tasks chosen to suit the machine, and numbers without dates or error bars. Prefer benchmarks other labs can run on any machine.
Check
Two quick ones
Collect
Your aha card
Accuracy buys depth, and depth buys usefulness.
Count operations you can trust, not qubits.