sizing spec

What a job costs, before it runs

The first version of this experiment was sized by hand and cost roughly 3,750.62 credits of work. The cost formula now lives in the code: gate counts are read off the compiled circuit, the estimate is printed before anything is uploaded, and a batch outside the envelope is refused with the number named rather than quietly trimmed. No job submitted, no shots bought. Gate counts and HQC estimates only; the AUROC column is an ideal statevector score, not a measured result.

Previous sizing
3,750.62

69 programs, 9 qubits, 1024 shots.

Current sizing
884.75

46 programs, 6 qubits, 256 shots.

Credits not spent
2,865.87

Same question, same cohort, same verdict discipline.

The formula

HQC = 5 + (N1q + 10*N2q + 5*Nm) / 5000 * C
N1q
one-qubit gates
N2q
two-qubit gates, weighted 10x
Nm
measurements, inits and resets, weighted 5x
C
shots

The +5 is per circuit, so for small circuits the program count dominates the bill before a single shot is taken.

Reviewer feedback from a Quantinuum engineer

The envelope below is not ours. It came from a review on the hardware side, and each point is written here as a rule the run now follows, with the line that enforces it.

Gate, qubit and shot limits
  • Two-qubit gates: target 100 per circuit, hard stop 1000, absolute ceiling 2000 once noise is counted.
  • Qubits: 4-20 for a demonstrator; 30+ is where classical simulation stops being easy.
  • Shots: 128-512 is the normal range. 1024 only with a stated reason.

The sizing preflight reads the compiled circuit and refuses over-width or over-gate batches by name before upload.

What actually drives the bill
  • The +5 is charged per circuit, so for small circuits the program count dominates before a single shot is taken.
  • A two-qubit gate counts ten times a one-qubit gate.
  • Cost is linear in shots — doubling shots doubles the bill.

Gate counts come off the compiled circuit, not a hand-typed estimate, and the per-program cost table prints at submit time.

Hardware reality
  • A million shots is not possible on Helios (98 qubits).
  • Allowable shots fall as the two-qubit gate count rises.
  • The two-qubit fidelity spec and the shot count belong in the same sentence.

Every quantum leg is labelled H2-Emulator, and the shot count is published next to the gate census rather than on its own.

Budget framing
  • 10,000 credits is a normal project; 100,000 is a showcase.
  • Cheaper is better — a smaller job that answers the question beats a large one that impresses.

A per-run budget is reserved up front; an over-budget batch is refused with the number named, never quietly trimmed.

The operating envelope

KnobLimitWhy
Two-qubit gates per circuittarget <= 100, hard stop 1000, absolute ceiling 2000 with noiseFidelity collapses past this on ion traps.
Qubits4-20 for a demonstrator; 30+ before classical simulation stops being easyBelow 20 a laptop reproduces the result.
Shots128-512 typical; 1024 only with a stated reasonCost is linear in shots.
Circuits per experimentKeep low — each carries +5 HQC69 circuits is 345 HQC before a single shot.
Budget10,000 HQC normal project, 100,000 HQC showcaseCheaper is better; the guard refuses rather than overspends.

A million shots is not possible on Helios (98 qubits). Allowable shots fall as the two-qubit gate count rises, so the two-qubit fidelity spec and the shot count belong in the same sentence.

Candidate circuits, costed side by side

Three feature maps for the same classification question. Gate counts come from the circuit itself; the score column is an ideal noiseless calculation used only to choose between designs, never reported as a measured result. Classical floor: Logistic regression with balanced class weightsAUROC 0.5.

DesignQubitsDepth1-qubit2-qubitCredits / programBatchIdeal score
Angle encoding + entangling ring615241012.88592.70.75
Second-order ZZ feature map635342018.52851.770.75
Hardware-efficient ansatz, 2 layerschosen623482019.23884.751

Generated Sun, 13 Sep 2026 01:48:42 GMT by Offline sizing analysis (pytket gate census + ideal statevector). Nothing on this page was submitted or billed.