“QPU hours” on the devices¶
Euro-Q-Exa: EuroHPC time is billed in wall-clock QPU seconds, the \(t\) below. LRZ also defines a unit for its quantum systems, the equivalent qubit-second: wall-clock time on the device, scaled by qubit count and by one-qubit gate time. The formula is as follows:
\(t\): wall-clock time on the QPU.
\(n(S)\): number of qubits in system \(S\).
\(T_{1qg}(S)\): one-qubit gate time of \(S\); Q-Exa is the reference system.
For LRZ’s definition of the unit, open: LRZ: Definition of the computing unit for the LRZ Quantum Systems
VLQ bills device execution time, measured between execution start and end on the hardware itself. Classical overhead, staging and result handling fall outside the meter, so a slow client costs you nothing.
IT4Innovations publish the billing formula, the per-shot cost, a minimum charge of 0.9 seconds per job, worked examples, and an explicit statement that queue latency is not billed. Their own figures imply a useful rule: below roughly 2,250 shots you pay the floor anyway, so submit few large jobs rather than many small ones.
The formula is as follows:
\(N_\text{jobs}\), \(N_\text{shots}\): number of jobs, and shots per job.
\(c\): cost per shot, 400 µs as the worst case, or 350 µs plus the actual circuit time.
0.9 s: the minimum charge per job.
To estimate consumption on VLQ, open: IT4Innovations: Estimating Consumption on VLQ
Piast-Q bills device time, which the site defines as time on the machine including ion loading and re-initialisation, and excluding queue wait, network round-trips and time your client spends waiting.
Classical resources are in node hours, as we are all used to. Sites could discuss them in GPU hours.