The first QubitBenchmark release
Published results for 21 quantum computers and research systems, from 10 makers, show two designs pulling ahead in different directions, and a large gap still to close.
Key points
- Trapped-ion machines run the biggest test circuits; superconducting machines run them tens of thousands of times faster.
- Not every “logical qubit” is error-corrected. Several headline counts come from codes that can only detect errors.
- Most results come from the machines’ makers. Only a handful were measured independently.
Summary. This first release records published benchmark results for quantum computers from Google, IBM, Quantinuum, IonQ, Rigetti, IQM, D-Wave and several neutral-atom groups. Every result links to its original paper or technical document, carries an evidence grade, and is only compared with results measured the same way.
1 Bigger or faster
QUOPS, a new benchmark from Sandia National Laboratories with Quantinuum and NVIDIA, measures two things at once: how large a random circuit a machine can run successfully, and how quickly it runs them. In the first results, Quantinuum’s trapped-ion Helios-1 scored 1,504 on circuits 16 qubits wide, and H2-1 scored 1,320. Google’s Willow (216) and IBM’s ibm_boston (204) ran smaller circuits, 6 qubits wide, but tens of thousands of times faster.
The authors estimate that breaking RSA-2048 encryption would need a score of about 250 million, roughly five orders of magnitude beyond today’s machines. Because the paper is not yet peer-reviewed, and because one maker helped design the test, these results are grade B.
2 What counts as a logical qubit
Error-corrected (“logical”) qubits are the route to large, useful machines, and the headline counts are growing fast. But the codes behind them differ. Quantinuum reports 48 logical qubits on Helios with a code that corrects errors. The Harvard–MIT–QuEra, Atom Computing and Infleqtion results used smaller codes that can detect errors but not correct them, and some kept only a small fraction of their runs. We list the two kinds separately and never rank one against the other.
The clearest error-correction result in this release is Google’s: a distance-7 surface code on Willow that fails 0.143% of the time per correction cycle, published in Nature in 2024.
3 Gate errors
The best two-qubit operations now go wrong around once in a thousand tries on trapped-ion machines, and a few times in a thousand on superconducting and neutral-atom machines. Makers report these figures in different ways (median, average or a single best pair), so the chart shows them together only to give a sense of scale.
4 How sure can we be?
Most results in this release come from the makers themselves, in papers or technical documents that explain their method (grade B). A few were measured independently: a Los Alamos study of quantum volume on several cloud machines, a TNO study of D-Wave annealers, and the Metriq project’s tests of IBM and Quantinuum machines. Announcements without a method are kept, with a warning, but never ranked.