“Qubit fidelity” is a concept that comes from the gate-model world, where one asks “if I try to apply a π-pulse (or a CNOT) how likely am I to get the state I expect?”.
D-Wave processors do not apply gates; they realise an analogue, time-dependent Ising Hamiltonian

H(s)=A(s)∑iσi​x+B(s)⎛⎝∑i hiσi​z+∑i<j> Jijσi​zσj​z⎞⎠ ,   0≤s≤1,

and the question is therefore not “how accurate is a gate?” but

• How accurately do the programmed hi, Jij appear on the chip?
• How uniform is the transverse energy A(s) and the final energy scale B(s)?
• How much classical/quantum noise is present during an anneal?

So the usual fidelity tools used for gate machines (randomised benchmarking, interleaved RB, etc.) are not applicable, and D-Wave does not publish “single-qubit fidelity” numbers because they are not meaningful for an annealer.

What is calibrated – and what is publicly known

Every D-Wave system is shipped with an automated calibration package that is run by the system itself (typically every night or when temperature cycles occur). Internally the processor measures, qubit-by-qubit and coupler-by-coupler:

1. Persistent-current Ip of each flux qubit.
2. The linear map between the dimensionless user parameters (hi,Jij, anneal-offset, etc.) and the physical flux biases applied by the Φ-DACs.
3. Crosstalk coefficients between every flux line and every qubit/coupler.
4. The annealing trajectory A(s),B(s) produced by the global bias coils.
5. Low-frequency (1/f) flux noise and high-frequency relaxation times.

Those data are stored in the controller and are used to pre-distort the control pulses so that, after correction, the effective Hamiltonian on the chip is as close as possible to what the user asked for. End-users do not see the raw calibration tables, but several aggregate numbers can be queried through the Ocean SDK:

sampler.properties['h_gain_schedule'] # mapping h -> flux
sampler.properties['j_settle_rms'] # RMS residual coupling error
sampler.properties['qubit_anneal_offset_ranges']
…

(The exact key names differ slightly between 2000Q and Advantage systems.)

Typical published numbers

The following figures are taken from the technical documentation of recent generations (Harris et al. Phys. Rev. B 2010, Lanting et al. PRX 2014, King et al. arXiv:1806.08422, and the “Advantage Performance” white paper, 2021):

• RMS residual local field after calibration: ≤0.004 (dimensionless, i.e. <0.4 % of full programmable range).
• RMS residual coupling error: ≤0.02 on 2000Q, ≤0.007 on Advantage (~0.7 % of full range).
• Persistent-current spread (σ/μ) over all qubits: ≲2 %.
• Flux noise at 1 Hz: 1–2 μΦ₀ /√Hz (comparable to other large-area flux qubits).
• Energy-relaxation time T₁ ≃ 10–20 µs, dephasing time T₂* ≃ 100 ns.

If one tried to convert those coherence times into gate fidelities, you would obtain numbers far below the 99.9 % seen in transmon processors; however, gate fidelity is simply not the performance metric that matters for quantum annealing.

How does one estimate performance?

Because the dominant error mechanism is control error (static mis-specification of h or J), users typically

• apply “gauge averaging” (re-run the problem under random σz → −σz transformations);
• request multiple anneals and choose the best energy sample;
• occasionally exploit per-qubit “anneal offsets” supplied by D-Wave to compensate for residual inhomogeneity.

All of these techniques assume that the calibrated control errors are small and slowly varying—which is precisely what the automated calibration procedure guarantees.

Is there a special benchmarking method?

Yes, but it is quite different from RB:

• “Ham-error benchmarking” (Lanting & Przybysz, D-Wave Tech. Rep.) – repeatedly program small, exactly-solvable Ising instances and fit the distribution of measured energies to infer mean and variance of control errors.
• “Thermalisation benchmarking” – measure how the success probability varies with anneal time to extract an effective system temperature and freeze-out point.
• “Gauge error spectroscopy” – measure residual field errors by exploiting that gauges flip the physical sign of hi,Jij but not of the error terms.

These in-house procedures produce the calibration tables mentioned above and they scale linearly with the number of devices, so they remain practical for 5000+-qubit processors.

Bottom line

• There is no single-qubit “fidelity” number for a D-Wave machine because it does not run quantum gates.
• What matters is how closely the implemented Ising Hamiltonian matches the requested one; that is characterised by per-device calibration data that D-Wave keeps internal but summarises through RMS error numbers of order 10⁻³–10⁻² of full scale.
• The processor constantly re-calibrates itself, and users can mitigate the remaining static errors by gauge averaging and anneal-offset techniques.

For more detail you can consult:

– R. Harris et al., “Calibration of an analog quantum annealer”, Phys. Rev. B 81, 134510 (2010).
– T. Lanting et al., “Entanglement in a quantum annealing processor”, Phys. Rev. X 4, 021041 (2014).
– J. King et al., “Performance of a quantum annealer under programmable disorder”, arXiv:1806.08422 (2018).
– D-Wave White Paper, “Advantage™ system – Performance improvements”, 2021.