QUOPS: Helios Hits 1,824. Chemistry Still Wants 250 Million

NVIDIA spent last week shipping CUDA-Q Logical, a compiler stack for fault-tolerant circuits that barely exist. This week the same company shows up on a scorecard.

Paul Arnold at Phys.org covered it on 17 September. Quantum Computing Report named the collaboration: Sandia National Laboratories, Quantinuum, NVIDIA. The paper is on arXiv as 2609.12146, titled “Benchmarking the computational power of quantum computers.” The metric is QUOPS, Quantum Universal Operation Performance System. It is trying to replace “we have N qubits and a nice fidelity slide” with two numbers you can put next to a competitor.

Those numbers are unkind to the slide.

What Q and Ω actually count

QUOPS does not grade a single gate. It runs randomized layers of arbitrary-angle single-qubit rotations, RP(θ), and CNOTs, across different circuit widths and depths. Quantum Computing Report says the full-system path includes compilation, error correction, syndrome decoding, and error mitigation. That is the point. A chip that looks clean in isolation can still fail when the compiler, the decoder, and the sampling overhead sit in the same job.

Two top-level scores:

• Q, the QUOPS score: the largest circuit size s = 2 × w × depth that still succeeds inside a utility-shaped volume w² ≤ s ≤ w³, at a mean process polarization of at least 1/√e, about 61%.
• Ω, the QUOPS rate: operations per second after you count the extra shots that error mitigation and postselection eat.

Phys.org’s plain-language version: the size of the largest computationally relevant circuit the machine can finish, and how fast it finishes it. Randomized patterns stand in for the shapes used in some scientific calculations. Canonical challenge problems sit on the same plot as stars. First utility, the authors argue, lives at large width and large gate count, not at a 50-qubit demo with depth 4.

I am not going to pretend 61% polarization is a number you should tattoo. I will say it is a threshold they published, which is more than most vendor “quantum volume” footnotes give you.

The leaderboard, with the speed column attached

Quantum Computing Report printed a table. I am copying the numbers, not the adjectives.

Quantinuum Helios-1, trapped-ion QCCD, 98 physical qubits: Q = 1,504, or 1,824 with leakage postselection. Ω = 303 QUOPS/s at width 16. Phys.org rounds the headline to “more than 1,500” and credits the all-to-all layout. Ions can talk to any other ion. Moving them takes time. That is the slow column.

Quantinuum H2-1, 56 qubits, same modality: Q = 1,320 (1,392 with postselection), Ω = 353 at width 12. Slightly faster than Helios, smaller circuits. The family resemblance is obvious.

Google Willow, 105 transmons on a 2D grid: Q = 216 at width 6. Ω = 2.0 × 10⁷ QUOPS/s. Phys.org says 20 million operations per second. That is the other personality. Rigid neighbors, small circuits, a clock that makes the ion machines look parked.

IBM ibm_boston, 156 transmons, heavy-hex: Q = 204 at width 6, Ω = 3.1 × 10⁵ QUOPS/s. Same neighborhood as Willow on Q, two orders of magnitude slower than Willow on Ω. A 156-qubit chip that cannot grow width past 6 in this test is the sentence procurement people should keep.

Helios again, this time as a logical machine: up to 8 Steane-encoded logical qubits, Clifford+T, active magic-state injection, syndrome extraction. Q = 40. Ω = 4.9 QUOPS/s. Width 4. That is the fault-tolerant column, and it is tiny. CUDA-Q Logical last week was software for this layer. QUOPS is the first public score I have seen that puts a logical stack on the same axes as Willow.

IonQ’s Superion 256 pitch is a factory. It is not in this table. Absence is not a ranking. It is a reminder that the people who showed up for QUOPS are the people who were in the paper.

1,824 versus 250 million

Phys.org is blunt about applications. Breaking RSA-2048, or modeling the molecules chemistry actually cares about, wants roughly 250 million to 340 million QUOPS. The best physical score in the study is 1,824. That is about 100,000 times short. The authors write that computational capability must grow by five orders of magnitude, which is why they keep pointing at fault tolerance.

Do the division yourself. 250,000,000 / 1,824 is in the 10⁵ range. Marketing that says “we are in the NISQ utility era” has to live next to that fraction. IBM and UChicago’s verifiable advantage paper was a different claim, on a different task. QUOPS is trying to be the yardstick you can rerun when the press release lands.

Willow’s 20 million operations per second looks like a rebuttal until you remember Q is 216. Fast at the wrong size is still the wrong size. Helios can hold a bigger circuit and then wait on the ion transport. Logical Helios can run a tiny encoded circuit and wait even more. None of these three personalities is 250 million.

Error rates, more physical qubits, or a design that wastes less connectivity: Phys.org lists those as the closing moves. None of them is a new logo. They are the same three knobs this industry has had since the first transmon paper.

Why Sandia put NVIDIA in the author list

CUDA-Q Logical was an orchestration layer. QUOPS needs compilation and decoding in the loop or the score cheats. NVIDIA’s presence here is not a GPU cameo. It is the admission that the benchmark is a software stack plus a chip. If you only measure the chip, IBM’s 156 qubits look like a win over Helios’s 98. On Q, Helios is not close. On Ω, Willow lapping everyone is the other warning: throughput without circuit volume is a demo.

The paper and, per Quantum Computing Report, public software repositories are the part that matters for anyone who is not Quantinuum. A closed benchmark is a press cycle. An architecture-agnostic one with a polarization threshold is something a lab can fail in public. Government procurement is named in that writeup as a use. That is the audience. Roadmaps that list “logical qubits” without a Q and an Ω are about to look incomplete, which is the point of publishing a score that makes today’s logical Helios a 40.

I will not turn this into investment advice. QUBT and IONQ and IBM will keep printing slides. QUOPS is a preprint. Preprints move. If Helios’s 1,824 depends on leakage postselection, the 1,504 without it is the number I would put in a procurement appendix. If Willow’s 2×10⁷ Ω is measured at width 6, do not compare it to Helios at width 16 and call it a race.

What to do with the table this month

If you run a quantum group, clone the repo when you can find it and try to reproduce one row. If you cannot reproduce Willow’s 216, the metric is not ready to be a contract. If you can, start asking vendors for Q and Ω at a stated width, with and without postselection, physical and logical.

If you write algorithms, the cone w² ≤ s ≤ w³ is the shape they think utility lives in. A circuit that is wide and shallow, or deep and skinny, may be a valid experiment and still sit outside the score. That is a design choice. It is also how a benchmark steers hardware.

If you just needed a translation of this week’s paper: trapped ions currently win the largest successful circuit. Superconducting grids currently win the clock. Encoded logical qubits currently look like a science project on the same axes. The applications on the poster need five more zeros. NVIDIA will sell you the compiler either way.

The CUDA-Q post from four days ago and this scorecard are the same week in two moods. One ships tools for machines that do not exist. The other measures the machines that do, and the gap is not subtle.

How the test is supposed to stop cheating

Component metrics are easy to game. Publish two-qubit fidelity on the best pair. Publish qubit count on a chip that cannot keep the last twenty coherent. Publish a “quantum volume” from a circuit family you compiled by hand. QUOPS tries to make the cheat visible by forcing randomized RP(θ) and CNOT layers through the same pipeline a user job would hit.

Polarization at 1/√e is a pass/fail, not a beauty contest. Either the process looks live enough at that width and depth, or the point is outside the capability region. Phys.org’s figure caption, lifted from the paper, draws that region in blue and green and puts the QUOPS score at the largest successful size inside the cone. Stars mark where first utility is “likely.” RSA and chemistry sit far up the plot. Helios’s 1,824 is a dot near the origin if you plot it on a log axis with 250 million.

Postselection is the remaining soft spot. Helios jumps from 1,504 to 1,824 when leakage postselection is on. Ω already claims to include sampling overhead. If a vendor quotes only the 1,824, ask what they discarded. If they quote Ω without width, they are comparing Willow’s 20 million at w=6 to an ion clock at w=16. That comparison is how a superconducting team wins a meeting and loses a chemistry circuit.

Logical Q=40 is the other honesty check. Eight Steane qubits with magic-state injection is real work. It is also two orders of magnitude below physical Helios on Q, and two more below Willow on Ω. Fault tolerance is not a free upgrade. It is a tax you pay to stop the 10⁵ gap from being a 10⁷ gap later. CUDA-Q Logical exists because that tax is currently a pile of scripts. QUOPS exists because someone wanted the tax on a chart.

IEEE Quantum Week is sitting in Toronto this same week, 13 to 18 September, which is why IonQ was busy handing out paper awards. Different circus. QUOPS is the one with a table you can argue with. If the software repositories actually land, argue with them. If they do not, treat the preprint as a well-sourced press cycle and wait for v2.

The applications on the poster still want 250 million. Nobody in the table is close. That is the story. The rest is which architecture you prefer to be far away in.