Quantinuum announced QUOPS (the Quantum Universal Operations Performance System) at the recent Quantum World Congress in Maryland. The benchmark was developed by Sandia National Laboratories’ Quantum Performance Laboratory, with contributions from Quantinuum and NVIDIA.

QUOPS reduces quantum computer performance to two numbers, a score (Q) that captures the largest random circuit a machine can execute successfully and a rate (Ω) that captures the effective operations per second it delivers.

The underlying paper, posted to arXiv just ahead of IEEE Quantum Week, reports measured results for Google Willow, IBM’s ibm_boston, and Quantinuum’s H2-1 and Helios-1, including a fault-tolerant logical-qubit implementation on Helios-1.

The benchmark addresses a gap in how the industry measures progress:

• Qubit counts, two-qubit fidelities, and gate speeds describe components.
• Quantum Volume, the field’s previous flagship system metric, becomes impractical to verify once circuits exceed what classical computers can simulate.
• QUOPS uses mirror-circuit techniques that scale past that limit.
• QUOPS measures physical and logical qubit systems on an equal footing, which no prior cross-vendor benchmark has done.

The initial results favor Quantinuum’s trapped-ion QCCD architecture by roughly sevenfold in capability, while superconducting systems lead in throughput by three to five orders of magnitude.

The more significant finding is the remaining gap. The authors calculate that today’s best machines are about five orders of magnitude below the circuit sizes required for canonical utility problems, such as RSA-2048 factoring and FeMoco energy estimation, a gap that physical-qubit scaling alone cannot close.

Technical Details

QUOPS defines a family of randomized benchmark circuits composed of alternating layers of arbitrary-angle single-qubit Pauli rotations and CNOT gates. Each circuit has independently variable width (qubit count) and size (rotation count plus twice the CNOT count).

The benchmark runs these circuits across a grid of shapes, determines which shapes a machine can execute successfully, and maps the boundary of that machine’s capability region.

The methodology rests on several core elements:

• Success threshold: A circuit shape passes if its mean process polarization reaches at least 1/√e (roughly 61%) at 95% statistical confidence.
• Utility cone: The Q score is the largest passing circuit size in the region where size falls between width squared and width cubed. The cone excludes narrow or shallow circuits that classical computers simulate efficiently and brackets the expected shape of first-utility workloads.
• Scalable verification: Mirror circuit fidelity estimation measures polarization without classically simulating the circuit, eliminating the exponential verification cost that capped Quantum Volume.
• Throughput metric: Ω reports effective operations per second, discounted for the sampling overhead of error mitigation and for shots discarded via postselection or error detection. The authors argue that this makes Ω more computationally relevant than raw rate metrics such as IBM’s CLOPS.
• Anti-gaming controls: Compilation barriers placed throughout the circuits limit how vendors may optimize benchmark workloads, so scores reflect general-purpose compilation.
• Error-mitigated capability: A lower polarization threshold produces a distinct error-mitigated capability region and score, with the corresponding sampling cost reflected in a reduced Ω.
• Modularity: The default circuits directly proxy rotation-heavy workloads, such as Hamiltonian simulation. The framework allows other circuit families, including Toffoli-heavy proxies for chemistry and factoring, to be substituted.

The paper reports the following measured results (all entries except the final row execute directly on physical qubits):

Platform	Architecture	Q (QUOPS)	Ω (QUOPS/s)	Width at Score
Google Willow	105 superconducting qubits, degree-4	216	2.0 × 10⁷	6
IBM ibm_boston	156 superconducting qubits, degree-3	204	3.1 × 10⁵	6
Quantinuum H2-1	56 trapped-ion qubits, all-to-all	1,320	353	12
Quantinuum Helios-1	98 trapped-ion qubits, all-to-all	1,504	303	16
Helios-1, leakage postselected	98 trapped-ion qubits, all-to-all	1,824	247	16
Helios-1, Steane [[7,1,3]] logical	Up to 8 logical qubits	40	4.9	4

At a 1% polarization threshold, error mitigation raises scores by three to eleven times while reducing the rate by three to four orders of magnitude. Under that regime, Helios-1 reaches 20,442 QUOPS at 0.29 QUOPS/s, and Willow reaches 651 QUOPS at 1.2 × 10⁴ QUOPS/s.

The paper translates published resource estimates into QUOPS targets:

• RSA-2048 requires approximately 2.5 × 10⁸ QUOPS.
• FeMoco requires approximately 3.4 × 10⁸ QUOPS.

Completing either in five days requires rates of 5,700 and 800 QUOPS/s, respectively, which today’s physical machines already meet or nearly meet. Circuit size is the binding constraint.

Using historical Quantum Volume data, the authors estimate that QUOPS capability has doubled every 1.4 years at Quantinuum and every 2.1 years at IBM. Extending those trends suggests utility between 2050 and 2070, whereas vendor roadmaps promising utility in the early 2030s require a fourfold acceleration.

The fault-tolerant experiment is the most candid result in the paper.

Quantinuum encoded up to eight logical qubits in the Steane code using fully fault-tolerant Clifford+T compilation, increasing QUOPS from 24 to 40 and QUOPS/s from 1.3 to 4.9 across three software iterations on unchanged hardware. The logical architecture still lags the same machine’s physical qubits by more than an order of magnitude. 

The authors project that a fault-tolerant system encoding 20 to 30 logical qubits in 1,000 to 5,000 physical qubits would exceed 10,000 QUOPS, a level the authors would find surprising for any physical-qubit architecture.

Reference implementations are available on GitHub, and NVIDIA published a QUOPS implementation in CUDA-Q alongside its CUDA-Q Logical announcement.

Analysis

Quantinuum benefits from QUOPS on two fronts :

• The first published results show that its trapped-ion systems score roughly seven times higher on Q than the leading superconducting systems, a lead that follows directly from all-to-all connectivity and high two-qubit fidelity.
• QUOPS also natively measures logical-qubit systems, aligning with Quantinuum’s argument, repeated at Quantum World Congress, that fault tolerance must be evaluated at the integrated system level. 

Quantinuum and its predecessor, Honeywell Quantum Solutions, built years of marketing around successive Quantum Volume records, and QUOPS gives the company a successor metric that extends into the fault-tolerant era and starts from a position of strength.

It’s a position carrying several risks:

• Independence is the benchmark’s central selling point, yet the authorship undermines it. Sandia holds first and senior authorship, yet 20 of the paper’s 30 authors are Quantinuum employees, and Quantinuum is promoting a benchmark on which it leads.
• Google and IBM did not co-author the framework. Google staff conducted the Willow experiments, and IBM hardware was accessed via Air Force Research Laboratory resources. Neither company has committed to reporting QUOPS.
• The throughput dimension exposes Quantinuum’s weakest attribute. Willow’s Ω is about 66,000 times that of Helios-1, and competitors will emphasize Ω as much as Quantinuum emphasizes Q.
• Quantinuum’s own logical result scores 40 QUOPS. The benchmark that validates its physical-qubit lead also quantifies how far behind its fault-tolerant implementation trails.

Practitioner Impact

QUOPS matters most to the small but growing population of buyers making seven- and eight-figure quantum decisions, including national laboratories, HPC centers, and enterprises evaluating on-premises systems.

Those buyers currently compare vendors using metrics each vendor selects. QUOPS gives them a vendor-neutral number they can include in an RFP and verify after delivery (and Quantinuum explicitly encourages agencies to set QUOPS thresholds in procurement). 

Practitioners should weigh several considerations before relying on it.

• Q and Ω answer different questions: A chemistry team running deep circuits on modest qubit counts cares about Q, while a team running high-shot sampling workloads cares about Ω. Neither number alone ranks systems.
• The utility cone embeds an assumption about workload shape: Workloads outside it require reading the full capability region, which the benchmark reports but the single score obscures.
• Running QUOPS requires expertise in mirror circuits, compilation barriers, and statistical testing: Most enterprise teams will rely on vendor-reported or third-party results for the foreseeable future.
• The default circuit family is optimized for rotation-heavy simulations: Buyers with Toffoli-dominated workloads, including factoring and much of quantum chemistry, should request the substituted circuit families the framework permits.
• Cloud quantum pricing typically tracks shots or machine time: Ω provides procurement and finance teams with a basis for estimating the cost per useful result across vendors.

Competitive Impact

QUOPS enters a crowded field of quantum performance metrics, most of which originate with a single vendor and few of which span both physical and logical systems.

IonQ and neutral-atom vendors, including QuEra, Pasqal, and Atom Computing, are absent from the initial results. Their participation will determine whether QUOPS becomes an industry standard or remains a Quantinuum-favored scorecard.

The following table summarizes the principal alternatives buyers and agencies use today.

Alternative	Primary Proponent	Model / Approach	How It Compares to QUOPS
Quantum Volume	IBM, widely adopted including by Quantinuum	Largest square random circuit passing a heavy-output test	QUOPS adds non-square shapes, logical-qubit support, and scalable verification. QV’s classical verification cost grows exponentially with width.
Algorithmic Qubits (#AQ)	IonQ	Qubit-count figure derived from QED-C application circuits	Simpler to communicate. Vendor-defined and tied to small application circuits far below utility scale.
CLOPS and EPLG	IBM	Raw circuit-layer throughput and layered-gate error rate	Measure speed and component quality. Ω adds mitigation and postselection overhead, and Q ties results to successful circuit size.
Logical qubit counts and logical error rates	Google, Microsoft, QuEra, Quantinuum, others	Encoded qubit count plus per-cycle or per-operation logical error	Standard in fault-tolerance announcements, with no link to computational capability. QUOPS shows eight Helios-1 logical qubits delivering 40 QUOPS.
QED-C Application-Oriented Benchmarks	Quantum Economic Development Consortium	Open suite of algorithm-level benchmarks	Broad and vendor neutral. Application circuits that fit on current hardware remain far smaller than utility workloads.
DARPA Quantum Benchmarking Initiative	U.S. Department of Defense	Government verification of whether any architecture reaches utility scale by 2033	Evaluates designs and roadmaps through expert review. QUOPS supplies a measured-capability metric such programs could adopt.

QUOPS differentiates itself most clearly in three areas: scalable verification, native handling of logical qubits, and an explicit link to resource estimates for named challenge problems. No competing metric covers all three.

Its weakest points are adoption and perceived neutrality. Quantum Volume gained legitimacy because IBM originated it and because Quantinuum and others adopted it. QUOPS needs Google, IBM, and at least one neutral-atom vendor to publish results before buyers treat it as a standard.

Final Thoughts

Sandia, Quantinuum, and NVIDIA have produced the most rigorous cross-platform quantum benchmark published to date. QUOPS measures integrated system capability, scales beyond classical verification limits, treats physical and logical qubits identically, and maps measured performance to published resource estimates for utility-scale problems.

Open implementations and CUDA-Q integration lower the barrier to independent replication.

Governance is the biggest open question. A benchmark promoted by the vendor that leads it needs broad buy-in to avoid the appearance of self-interest, and the authorship balance gives competitors an easy reason to decline. Sandia’s stewardship is the asset that matters most here, and moving QUOPS under a neutral body, such as a standards organization or a DOE-led consortium, would strengthen its claim to independence.

The most important number in the QUOPS paper belongs to no vendor. Every quantum computer measured today sits roughly five orders of magnitude below the capability needed for the first utility-scale problems, and physical-qubit improvements alone appear to close that gap decades too late.

QUOPS turns the fault-tolerance debate into a measurable race, and the first logical-qubit system to post a QUOPS score above that of every physical-qubit machine will mark the moment when quantum computing’s utility timeline becomes credible.