How your circuit is compiled

IonQ compiles a circuit submitted through the abstract gate interface before it runs, so the circuit on the hardware is not the one you wrote, and the rewrite is sometimes considerable.

The gate count you wrote is not the one submitted, and a circuit that looks small in your editor can still exceed a backend's limits.

What the compiler changes

Three transformations happen between pressing Run and the circuit executing:

• Multi-qubit gates are decomposed into two-qubit operations. A three-qubit gate does not exist on the hardware, so it becomes a sequence of two-qubit ones.
• Standard gates are converted to native equivalents. RX, RY, RZ and CNOT are all abstractions. Each is rewritten into the gates the machine physically performs.
• Redundancy is removed, then the whole thing is optimised again. Extraneous gates are optimised away or commuted into different places, then transpiled into native gates using special heuristics and optimised again. The sequence repeats rather than running once.

The first two usually add gates and the third removes them, so the net change depends on the circuit.

IonQ's native gates

Native gates are the ones IonQ hardware physically executes, by addressing ions with resonant lasers via stimulated Raman transitions. Everything else is a convenience that compiles down to these.

Qollab's IonQ hardware backends are Forte 1 and Forte Enterprise 1, so ZZ is the entangler behind your hardware runs. The Aria rows apply when you pick an Aria noise-model simulator from the Select QPU dialog.

The hardware has no physical Z gate. A Virtual Z advances or retards the phase of the following operation instead, which removes the RZ gate by folding its angle into the phase parameters of the gates that come after it. It costs no physical gate on the machine, so a circuit heavy in Z rotations can compile smaller than its written gate count suggests.

Why your gate count changed

Because the submitted circuit is a rewrite, the number in your editor is not what the backend counts against its limits. This is the mechanism behind two entries in the Error reference: TooManyGates fires on the compiled count, not the written one, and UnsupportedGate means a gate survived to submission that the target does not accept.

A circuit that compiles larger but stays under the limit runs, and comes back noisier: more gates means more opportunity for error. See Why your results look wrong.

Why a local transpile does not match

Running Qiskit's transpile() yourself costs nothing and submits no job, and its output is a version of this rewrite. It is not the circuit that runs.

IonQ compiles again on its own side after you submit, and those steps are proprietary transformations you cannot inspect. The compiler expands gates into native ones and also cancels or merges what it can, so the count you see locally can move either way on IonQ's side.

Should you write native gates yourself

Usually not. IonQ recommends native gates only when the abstract interface cannot do what you need: when you have to know exactly which circuit ran, or when you specifically want a circuit that is not optimised.

The examples IonQ gives are research work: investigating and characterising noise, designing noise-aware circuits, exploring new error mitigation techniques, and developing or testing methods for circuit optimisation and compilation.

For anything else, including a hackathon project, the abstract interface is the better choice. IonQ's own position is that if the goal is simply to run a quantum circuit for evaluation or algorithm investigation, native gates will often give lower-quality output than the abstract gate interface.

Submitting in native gates is currently the only way to bypass IonQ's compiler and optimiser, and it also turns error mitigation off by default, though debiasing can be turned back on. See Error mitigation and debiasing.

One more constraint: a circuit has to use one gateset throughout. You cannot mix native gates and abstract gates in the same circuit.

Related

• Compute backends
• Why your results look wrong
• Error reference
• Python / Qiskit projects