Skip to main content

Cross-Platform Comparison

Three Chips, One Suite

Identical quantum diagnostic circuits run on three different processors. Same algorithms, different hardware, different answers. All experiments designed and executed by an AI agent (Claude Opus 4.6).

2026-02-10·~105 hardware jobs·~253K shots·3 quantum processors

The Processors

Tuna-9

QV=8

Quantum Inspire / QuTech

Qubits 9
Edges 12
Topology Diamond lattice
Native H, CNOT

IQM Garnet

QV=32

IQM Resonance

Qubits 20
Edges 29
Topology Square lattice
Native PRX, CZ

IBM Torino

QV=32

IBM Quantum

Qubits 133
Edges Heavy-hex
Topology Heavy-hex
Native CZ, SX, RZ

Bell State Fidelity

Simplest entanglement benchmark: create (|00〉+|11〉)/√2 and measure. The spread between best and worst pairs reveals chip uniformity.

80%85%90%95%100%Best pair93.5%98.4%86.5%Worst pair85.8%91.2%86.5%Average90.2%96.3%86.5%
Tuna-9 IQM Garnet IBM Torino

*IBM used default transpiler qubit placement (not cherry-picked). IQM Garnet: 22/29 pairs swept (mean 96.3%, 7.2pp spread). All platforms have qubit-quality variation.

GHZ Fidelity vs. Qubit Count

How fast does entanglement quality degrade with circuit size? IBM Torino pushes to 50-qubit GHZ. Per-qubit error stays remarkably consistent at ~5%.

0%20%40%60%80%100%3q5q10q20q50q88.9%83.8%93.9%81.8%54.7%82.9%76.6%62.2%34.3%8.5%Tuna-9IQM GarnetIBM TorinoGHZ Fidelity

Key finding: qubit quality > circuit depth

On IQM Garnet, a GHZ-5 circuit routed through high-quality qubits (81.8%, depth 20, 9 CZ gates) outperformed one routed through the weak QB9 region (57.8%, depth 16, 7 CZ gates). Fewer gates on bad qubits is worse than more gates on good qubits.

Quantum Volume

Standard benchmark: heavy output fraction must exceed 2/3 for each circuit width n. QV = 2n_max.

2/30.50.60.70.80.91.0n=2n=3n=4n=5n=6n=7Tuna-9 (QV=16)IQM Garnet (QV=32)IBM Torino (QV=32)
8Tuna-9 QV

Passes n=2,3. Fails at n=4 (9 qubits, sparse topology).

32IQM Garnet QV

Passes n=2,3,4,5. 4x Tuna-9. 20 qubits, 29 connections.

32IBM Torino QV

Passes n=2-5, fails at n=6. Same as Garnet despite 6.7x more qubits. 133q, heavy-hex.

Noise Fingerprints

Bell state tomography in Z, X, Y bases reveals the type of noise, not just the amount. Dephasing (T2): ZZ > XX ≈ |YY|. Depolarizing: all equal. Amplitude damping (T1): asymmetric.

0.60.70.80.91.0⟨ZZ⟩0.9450.9750.729⟨XX⟩0.9020.9490.704|YY|0.8960.8810.675Tuna-9IQM GarnetIBM Torino

Tuna-9 — q4-q6 (best pair)

Dephasing. ZZ=0.945 > XX=0.902 ≈ |YY|=0.896. Phase degrades faster than population.

IQM Garnet — QB14-QB15

Dephasing. ZZ=0.975 > XX=0.949 > |YY|=0.881. Highest overall correlations. T2 is the bottleneck.

IBM Torino — default pair

Depolarizing. ZZ=0.729 ≈ XX=0.704 ≈ |YY|=0.675. All correlators within 5%. Uniform noise across all bases.

At a Glance

MetricTuna-9IQM GarnetIBM Torino
Qubits920133
Connectivity12 edges29 edgesHeavy-hex
Bell fidelity93.5%98.4%86.5%*
GHZ-388.9%93.9%82.9%
GHZ-583.8%81.8%76.6%
GHZ-10n/a54.7%62.2%
GHZ-20n/an/a34.3%
GHZ-50n/an/a8.5%
Quantum Volume163232
Dominant noiseDephasingDephasingDepolarizing
Per-qubit error3.5-6.2%1.4-5.0%4.6-6.1%
QPU time used~42K shots~70K shots44s / 10 min

Key Takeaways

IQM Garnet wins on gate quality — Highest Bell fidelity (98.4%), best GHZ-3 (93.9%), and strongest correlators. For circuits that fit in 20 qubits, Garnet delivers the cleanest results.

IBM Torino wins on scale — 50-qubit GHZ (8.5% fidelity) is noisy but measurable. GHZ-20 at 34.3% and GHZ-10 at 62.2% beat Garnet's GHZ-10 (54.7%). More qubits = more routing options = better qubit selection.

Tuna-9 punches above its weight — QV=16 on 9 qubits with basic gates. Best 5-qubit GHZ (83.8%) beats both IBM (76.6%) and Garnet (81.8%). Small but well-characterized.

QV=32 on both Garnet and Torino — More qubits doesn't mean higher QV. Per-gate error rate is the bottleneck, not qubit count. Both fail at n=6 where circuit depth overwhelms coherence.