A team led by researchers at Sandia National Laboratories introduced QUOPS, a benchmark for comparing quantum computers built on different hardware, in a preprint posted on September 10, 2026. After scoring processors from Google, IBM, and Quantinuum, the authors concluded that “current QUOPS scores are insufficient by roughly five orders of magnitude” for the two challenge problems they analyzed.
Those problems are factoring a 2048-bit RSA key and estimating an energy level of FeMoco, the metal cluster at the core of the enzyme that bacteria use to turn atmospheric nitrogen into ammonia. The authors calculated that factoring RSA-2048 would require a QUOPS score of about 250 million, against 1,504, the best score they measured without discarding any runs, on Quantinuum’s Helios-1.
Running QUOPS on error-corrected logical qubits, the authors measured a score of 40 on the same Helios-1 hardware.
Who ran the tests
Timothy Proctor of Sandia’s Quantum Performance Laboratory in Livermore, California, is first author. Robin Blume-Kohout, based at Sandia in Albuquerque, New Mexico, is senior author. The paper’s 30 authors list affiliations with Sandia, Quantinuum, NVIDIA, and the University of New Mexico. Twenty of them list Quantinuum, whose machines had the highest scores among the physical-qubit processors.
The authors thanked Ryan Byrne and Eliott Rosenberg for running the experiments on Google’s Willow. The team ran its IBM experiments on the Air Force Research Laboratory’s IBM Quantum resources.
The authors wrote that, to the best of their knowledge, this is “the first cross-platform benchmarking experiment to directly compare physical- and logical-qubit quantum computations on an identical footing.”
How QUOPS works
A QUOPS test circuit, a sequence of operations on qubits, alternates layers of single-qubit rotations with layers of CNOT gates, a standard two-qubit operation. Its width is the number of qubits it acts on. Its size is the number of rotations plus twice the number of CNOT gates.
Under the rules in the preprint, a circuit shape, meaning a given width and size, passes if its mean process polarization is at least 1/√e, about 0.61, with 95% confidence. Process polarization is a measure of output quality on which 1 is perfect and 0 is pure noise. The QUOPS score is the size of the largest passing circuit among shapes in which size is between the square and the cube of the width.
The QUOPS rate is the number of operations a machine executes per second, with corrections for imperfect output and for runs discarded after error detection. The team estimated process polarization with mirror circuits, in which a circuit is followed by its reverse so that the correct output is known in advance.
The scores
On Google’s Willow, a superconducting processor with 105 physical qubits, the team measured a score of 216 QUOPS at 20 million QUOPS per second. IBM’s ibm_boston, a Heron r3 processor with 156 programmable qubits, scored 204 at 310,000 QUOPS per second.
Quantinuum’s H2-1 trapped-ion processor, with 56 physical qubits, scored 1,320 at 353 QUOPS per second. Quantinuum’s Helios-1, with 98 physical qubits, scored 1,504 at 303 per second.
Helios-1 reached 1,824 at 247 per second when the team discarded runs in which leakage was detected. Leakage is an error in which a qubit leaves the two states that encode its information.
On Helios-1, the team also encoded up to eight logical qubits in the Steane code. In that code, each logical qubit is stored in seven physical qubits so that any single-qubit error can be corrected. The logical-qubit processor scored 40 QUOPS at a width of four logical qubits and a rate of 4.9 QUOPS per second.
The authors wrote that they chose the code for its simplicity, without optimizing for QUOPS performance. They also wrote that the experiment “does not saturate Helios-1’s fault-tolerant capability.”
Targets and trend
Using Craig Gidney’s 2025 estimate for RSA-2048, the authors calculated a target of about 250 million QUOPS and a required rate of about 5,700 QUOPS per second to finish in five days. For FeMoco, they used a 2025 estimate by Guang Hao Low and colleagues and calculated about 340 million QUOPS at about 800 per second.
The authors wrote that today’s physical-qubit processors have, or nearly have, the necessary rate. The five-order-of-magnitude shortfall they identified is in QUOPS scores, a measure of the size of circuits a machine can run.
In their conversion, the authors assumed four T gates, a costly class of logic operation, for each Toffoli gate, the three-qubit operation in which both resource estimates are counted. With seven, the targets would rise to about 420 and 590 million, they wrote.
The team also reconstructed QUOPS scores for IBM and Quantinuum processors from published results on quantum volume, an earlier benchmark. The authors found that scores doubled about every 1.4 years for Quantinuum and every 2.1 years for IBM.
Extending that trend, the authors calculated that machines would not reach the capability required for useful challenge problems until 2050–2070. The dates extend the fitted trend and assume that it continues unchanged. They also wrote that the trend would have to accelerate fourfold to meet hardware roadmaps with major scientific capabilities projected for the early 2030s.
For physical-qubit operation on the two architectures they modeled, the authors calculated a ceiling of about 330 QUOPS with each qubit linked to four neighbors, however many qubits a machine had. With every qubit linked to every other, they put the ceiling at about 10,000.
They wrote that achieving much beyond 10,000 QUOPS with physical-qubit operations “would be both remarkable and surprising.”
The next milestone
A fault-tolerant machine that outscores every physical-qubit machine would be “a major milestone for the field,” the authors wrote. They projected that a machine encoding 20 to 30 logical qubits in 1,000 to 5,000 physical qubits could “comfortably exceed” 10,000 QUOPS.
Update, November 12, 2026: Quantinuum wrote in a September 14 blog post that Sandia developed QUOPS with input from Quantinuum and NVIDIA. In an October 19 release, Sandia said useful workloads “will require 270 to 370 million QUOPS,” figures that differ from the preprint’s 250 million and 340 million targets.
