IBM
Heron r2
133-156 qubit class · Heavy-hex
NATIVE GATE SET
RZZ, SX, RZ, X
SWAPs 0
Idle 0.50
PREDICTED INFIDELITY (95% CI)
0.0106
[0.0097, 0.0114]
VENDOR-AGNOSTIC TRANSPILER / v1.0
One circuit. Four vendors. Four cost surfaces.
Send one OpenQASM 3 circuit. The transpiler emits four vendor-native artefacts, each with a SWAP count, an idle-time fraction, and a predicted infidelity with a 95% confidence interval. Every artefact is pinned to a vendor SDK version and replayable byte-for-byte by canonical SHA-256.
SOURCE CIRCUIT / OPENQASM 3 / GHZ-3
One Hadamard, two CNOTs, three measurements. Seed 42.
IonQ
Tempo
~64 qubit class · All-to-all
NATIVE GATE SET
GPi, GPi2, MS
SWAPs 0
Idle 0.97
PREDICTED INFIDELITY (95% CI)
0.0083
[0.0076, 0.0090]
Quantinuum
H3-1
~96 qubit class · Zone graph
NATIVE GATE SET
U1q, ZZ
SWAPs 0
Idle 0.55
PREDICTED INFIDELITY (95% CI)
0.0032
[0.0030, 0.0035]
Pasqal
Orion β
~100 atom class · Programmable register
NATIVE GATE SET
CZ + Rabi-detuning
SWAPs 0
Idle 0.98
PREDICTED INFIDELITY (95% CI)
0.0164
[0.0150, 0.0177]
WORKED EXAMPLE / GHZ-3 / SEED 42
The same source circuit. Different topologies. Different costs.
| Vendor | Native gate set | Depth | 2-qubit gates | SWAPs | Idle fraction | Predicted infidelity (95% CI) |
|---|---|---|---|---|---|---|
| IBM Heron r2 | RZZ, SX, RZ, X | 2 | 2 | 0 | 0.50 | 0.0106 [0.0097, 0.0114] |
| IonQ Tempo | GPi, GPi2, MS | 2 | 2 | 0 | 0.97 | 0.0083 [0.0076, 0.0090] |
| Quantinuum H3-1 | U1q, ZZ | 2 | 2 | 0 | 0.55 | 0.0032 [0.0030, 0.0035] |
| Pasqal Orion β | CZ + Rabi-detuning | 2 | 2 | 0 | 0.98 | 0.0164 [0.0150, 0.0177] |
Same source circuit. Different topologies. Different costs. The manifest names every choice that produced these numbers, from the seed to the SDK version. Quantinuum H3-1 has the lowest predicted infidelity because the 2026 stub registry assigns it the highest published 2Q fidelity. Pasqal Orion β has the highest because the same registry assigns Rydberg-blockade 2Q the lowest. Neither line is a marketing statement. Both are properties of the registry the manifest cites.
VENDOR TOPOLOGIES
Connectivity is the cost.
SWAP overhead is a property of the coupling graph, not a vendor opinion. The four backends have four different graphs.
IBM Heron r2
Heavy-hex. Degree-3 nodes with weight-3 connections. SWAP overhead grows on long-range circuits.
IonQ Tempo
All-to-all. Every qubit connects to every other. SWAP overhead is structurally zero; the cost shifts into MS-pulse depth.
Quantinuum H3-1
Zoned racetrack. All-to-all within a gating zone. Cross-zone interactions trigger ion shuttling.
Pasqal Orion β
Programmable atom register. Blockade radius defines the effective neighbourhood. Out-of-blockade pairs route through intermediate atoms or analog segments.
COST MODEL
Three numbers per vendor.
The cost manifest is not a single score. The reader picks the column that matters for their workload. SWAP overhead dominates long-range circuits on heavy-hex. Idle fraction dominates sparse layers on all-to-all. Predicted infidelity dominates everything when error budget is tight. The manifest reports all three.
SWAP overhead
gates inserted by the routing pass
The router walks the source DAG and inserts SWAP gates wherever a 2Q operation crosses a non-edge in the vendor coupling map. All-to-all vendors report SWAP = 0 by construction. Heavy-hex on QFT-10 reports SWAP = 11.
Idle-time fraction
register-time spent idle vs computing
A 2Q gate occupies two qubits; the other qubits in the layer sit idle. Idle fraction is high on circuits that gate few qubits per layer. The number is per-vendor because layer scheduling is per-backend.
Predicted infidelity
bootstrap CI under stub-stage noise model
Pauli-Lindblad coefficients per vendor drive the prediction. Bootstrap resampling gives the 95% CI. The stub-stage CI is widened by a factor of 3; the fitted-stage CI lands when the mitigation spine reaches its measurement gate.
THE SIGNED MANIFEST
Replayable byte-for-byte by canonical SHA-256.
CANONICAL HASH / manifest.qapp
aa69456755e2027fda5ec65cc650683babd3af8e89623cc561ea3545b3dab738 Re-run the transpiler with the same seed, the same SDK version, the same optimization level. The hash recovers byte-for-byte. The reproducibility envelope captures seed, optimization level, basis gates, transpiler passes in order, SDK version, calibration timestamp.
The manifest carries the full hardware-run record at production: vendor, backend, job ID, vendor log URL, calibration timestamp, T1 and T2 at run time, transpiler passes in order, basis gates, mitigation primitive, statistical-equivalence test result. An EU auditor under DORA Art. 28-30 can replay the manifest byte-for-byte.
WHAT THIS DOES, AND WHAT IT DOES NOT
A ledger, not a pitch.
FOR 4
- Enterprise platform-engineering lead pricing vendor lock-in cost before a multi-year quantum program
- Quantum research lead choosing one of four hardware paths and needing apples-to-apples cost surfaces
- AI / HPC infrastructure architect at a Tier-1 EU bank running a procurement-grade vendor evaluation
- Auditor under DORA Art. 28-30 who needs a replayable artefact, not a vendor-authored chart
NOT FOR 4
- Vendor managed-service buyers (the IBM Quantum platform, Quantinuum Nexus, IonQ Cloud, Pasqal Cloud already run circuits as a service)
- Buyers expecting a single vendor ranking (the manifest reports per-column cost, not a composite score)
- Customers who want the tool to execute circuits on real QPUs (execution is the vendor runtime; the tool emits the submittable artefact)
- Workloads that require Atom Computing or any vendor not in v1.0 scope (held for amendment + registry row)
QUESTIONS
What buyers ask.
Q 01
Why four vendors, not one?
A single-vendor transpiler picks the cost profile that flatters its own backend. The buyer comparing hardware paths needs the same source circuit lowered into four native gate sets so the difference is in the topology, not the toolchain.
Q 02
Is this a managed-service runtime?
No. The tool emits four vendor-native artefacts plus a cost manifest. The buyer submits each artefact to the vendor runtime that owns it. Execution stays at the vendor; comparability stays at the manifest.
Q 03
Does the manifest pick a winner?
No ranking. The cost manifest reports SWAP overhead, idle-time fraction, and predicted infidelity per vendor. Where a vendor is genuinely worse on a workload, the manifest says so by number, not by adjective.
Q 04
How do I verify the prediction?
Every artefact ships under a signed .qapp manifest with a SHA-256 hash. The reproducibility envelope captures seed, optimization level, basis gates, transpiler passes in order, SDK version, and calibration timestamp. Two runs under the same envelope produce byte-equal artefacts.
Q 05
What is in v1.0 and what is not?
v1.0 covers IBM Heron r2, IonQ Tempo, Quantinuum H3-1, and Pasqal Orion β. Atom Computing is not in v1.0; adding it requires an amendment plus a registry row. The four-vendor scope is final at Gate 0.
→
One circuit, four cost surfaces, one signed manifest.
Send OpenQASM 3 or a Qiskit circuit. The four artefacts come back side by side. The cost manifest is the artefact a procurement team can read.
Atom Computing is not in v1.0 scope. Adding it requires an amendment plus a registry row. The four-vendor scope is final at Gate 0.