Superconducting vs trapped-ion: I ran the same circuit on two quantum computers

· quantum, ibm-quantum, ionq, trapped-ion, hardware, error-mitigation, zne, honest-benchmarks


We took the exact same trained quantum circuit and ran it on two physically different quantum computers: IBM’s, which stores qubits in tiny superconducting loops chilled near absolute zero, and another built from single trapped ions — individual atoms held still by lasers. Same recipe, two utterly different machines. Then we compared each machine’s answer to the perfect simulator answer, and tried to clean up the noise on each — honestly reporting when the cleanup helped and when it backfired.

Same model, two machines

We never retrain per machine — that would burn hardware time for zero learning. We train once on the simulator, reconstruct the circuit natively, prove it matches the trained model to ~1e-15, then run inference on each backend. The superconducting run (ibm_fez) used 2048 shots; the trapped-ion run (ionq:forte-1) used 100. Both aim at the same exact simulator value, +0.7240.

The twist: better raw, worse mitigation

The headline you’d expect — ‘superconducting vs trapped-ion, who wins?’ — has an honest answer with a sting. The trapped-ion machine’s raw answer was closer to the truth. But zero-noise extrapolation, the same cheap trick that helped on the superconducting run, hurt here: at only 100 shots the noise-scaled points came out non-monotonic ([+0.60, +0.80, +0.52] at scales [1, 3, 5]), so the extrapolation ran the wrong way. That’s the whole lesson — mitigation is a tool with a failure mode, not a free ‘make it better’ button.

Results

backendtechnologyshotssim ⟨Z₀⟩raw ⟨Z₀⟩mitigated ⟨Z₀⟩raw errormit error
ibm_fezsuperconducting2048+0.724+0.486+0.8700.2380.147
ionq:forte-1trapped-ion100+0.724+0.600+0.3200.1240.404

Figures

grouped sim/raw/mitigated bars: superconducting vs trapped-ion

Same circuit, two machines (on-brand, generated): the shared simulator truth (gold line), then each backend’s raw (red) and ZNE-mitigated (green) ⟨Z₀⟩. The trapped-ion raw sits closest to the gold line — but its green mitigated bar overshoots past it, the misfire the numbers actually show.

Honest caveats

This is a cross-hardware reproduction of one trained model on two qubit technologies — not a quantum-advantage claim (D1). The honest finding has two halves, and the second is the interesting one:

Reproduce it

Every number traces to cross_hardware.json / summary.json / equivalence.json (D7). Open Quantum job id(s) 4f0ed05e-49f3-4ddb-ab75-03d00ec7998d, 5e453e57-fb83-4441-91af-5af8a3e654af, b2f1ed57-ef17-4a47-ab39-e67f7f36ba4d; IBM job id(s) d97qoc52su3c739if17g. experiments/20260711-090430_p3b-openquantum.