IQM Researchers Benchmark Projected Quantum Kernels Across 45 Qubits Using Cross-Device Referee-Challenger Protocol

IQM and QbitSoft benchmark projected quantum kernels on up to 45 qubits using a cross‑device referee‑challenger protocol.

IQM Researchers Benchmark Projected Quantum Kernels Across 45 Qubits Using Cross-Device Referee-Challenger Protocol

Benchmarking Projected Quantum Kernels on 45 Qubits

IQM Quantum Computers, together with QbitSoft, have released a technical preprint describing an empirical study of projected quantum kernels (PQKs) on their superconducting processors. The work evaluates PQKs on 15, 30, and 45 qubits using the 54‑qubit Emerald and 20‑qubit Garnet chips, employing a novel cross‑device referee‑challenger framework to guard against over‑fitting to device‑specific noise.

Cross‑Device Referee‑Challenger Protocol

The protocol designates one processor as a referee that assigns labels to data points, while a separate processor (or a spatially isolated qubit chain) acts as the challenger that attempts to classify the same data. By forcing the two devices to compete, the method reduces the risk that a quantum machine‑learning model merely learns the idiosyncrasies of a single chip’s noise profile.

Key Findings

  • 15‑qubit shallow circuits (3 Trotter steps, depth 45) achieved 81 % quantum accuracy versus 51 % for the best classical baseline, a statistically significant 30‑point gap.
  • 15‑qubit deep, unmitigated circuits (10 Trotter steps, depth 129) saw accuracy drop to 68 % versus 64 % classically, narrowing the gap to an insignificant 4 points.
  • Error‑mitigation (zero‑noise extrapolation, adaptive Pauli twirling) restored much of the lost performance; ZNE added 14 points, while APT contributed 17 points and halved feature‑noise.
  • 30‑ and 45‑qubit runs maintained a 17‑ to 28‑point advantage over 11 classical models. A matrix‑product‑state (MPS) simulation of the 45‑qubit circuit matched the hardware result (72 % vs. 73 %).

The study encodes Fashion‑MNIST images into a second‑order Trotterized 1D Heisenberg Hamiltonian, following a classification protocol originally proposed by Google researchers. The authors note that because the 45‑qubit circuit can be simulated quickly on a classical MPS simulator, future work will need to target more complex 2D dynamics to approach genuine quantum advantage.

Practical Takeaways

  • Cross‑device validation appears effective at exposing over‑fitting to hardware noise, a common pitfall in quantum machine‑learning experiments.
  • Error‑mitigation techniques such as ZNE and adaptive Pauli twirling can substantially recover performance, but their impact on the labeling process must be carefully accounted for.
  • Scalability: While the current benchmarks show promising gaps over classical baselines, the ability to simulate the circuits classically suggests that larger, more entangled workloads are required to demonstrate a clear quantum advantage.

The full preprint and accompanying blog post provide detailed methodology, circuit specifications, and raw data for interested researchers.

--- This article synthesizes the information provided in the original briefing without adding external quotations or unpublished data.

Hype check

Partly

The benchmark is presented in a preprint without peer review and lacks independent replication, though the data appear concrete.

  • Peer-reviewed paper, or press release only?Press release only — Only a technical preprint on arXiv is mentioned
  • Error-corrected (logical) qubits, or physical qubits only?Logical / not a qubit claim — The claim concerns kernel performance, not qubit counts
  • A new milestone, or a rebrand / restatement?New — First reported cross‑device PQK benchmark up to 45 qubits
  • Independent confirmation?Not yet — No external confirmation or reproduction cited
  • Does the headline claim match what the source shows?Matches — Headline accurately reflects the described benchmark

An editorial assessment of the claim as reported, not a judgment of the company or researchers. How we rate hype →

Source

Read original source

Why we picked this

Accepted from curated QuantumUnzipped source (relevance scoring disabled).

← Back to all articles