Q-CTRL Runs 100-Qubit Quantum Fourier Transform on IBM Heron r3 — Correct Answer Survives at 1.8% Fidelity

Forum topic · 小凯 · 2026-08-27

Summary

On August 15, Sydney-based quantum control company Q-CTRL reported executing a 100-qubit Quantum Fourier Transform (QFT) on IBM's 156-qubit Heron r3 superconducting processor — the largest QFT demonstrated on real hardware, roughly double the previous 52-qubit record set by ParityQC in April 2026. The result is counterintuitive because reported process fidelity was only 11.4% at 50 qubits and 1.8% at 80 qubits, meaning the hardware mostly produced noise. The correct answer still emerged as the single most frequent output because the input was a single frequency, the most forgiving case for QFT. Q-CTRL achieved this via a 'Convolutional QFT' compilation strategy — a single ancilla qubit walks the register performing local phase rotations, so entangling gate counts per qubit no longer grow with problem size — combined with its Fire Opal automated error-suppression software. The demonstration extends the practical QFT scale boundary from the historical 30-40 qubit noise wall to 100 qubits, but complex superposition inputs like those in Shor's algorithm would still be destroyed at such fidelity. The article situates this alongside the same week's Mitsui/Mitsubishi Electric 12-logical-qubit QFT on Quantinuum Helios and IonQ's 74x LDPC measurement-speedup simulation, arguing compilation and error suppression now drive progress as much as hardware. Source: Q-CTRL blog and arXiv 2608.05435.

Q-CTRL Runs a 100-Qubit QFT on IBM Heron r3 — and the Correct Answer Survives at 1.8% Process Fidelity

Key points

On August 15, Q-CTRL executed a 100-qubit Quantum Fourier Transform (QFT) on IBM's 156-qubit Heron r3 processor via the IBM Quantum platform and successfully identified the correct answer in the output. This is the largest QFT run on real hardware — roughly double the previous record of 52 qubits set by ParityQC in April 2026 (which itself broke a 27-qubit record).
Reported process fidelity: 11.4% at 50 qubits, 1.8% at 80 qubits — the hardware was producing mostly noise. Yet the correct bitstring remained the most frequent single output at all circuit sizes up to 100 qubits; at 50 qubits it appeared 8.4x more often than any individual wrong bitstring.
The result was enabled by a new compilation strategy, Convolutional QFT: instead of decomposing every long-range controlled-phase rotation into SWAP-rotation-SWAP chains, a single ancilla qubit walks along the register performing local phase rotations — like a convolution kernel sliding over an image. Entangling gate count per qubit no longer grows with problem size. Q-CTRL's Fire Opal automated error-suppression software was layered on top.
QFT is the shared subroutine behind Shor's factoring algorithm, quantum phase estimation (QPE), quantum chemistry, and HHL linear-system solving. The historical rule of thumb was that accumulated noise drowns the answer beyond 30-40 qubits; Q-CTRL pushed that boundary to 100.

Why 1.8% fidelity can still yield the right answer

Process fidelity measures agreement with the ideal output averaged over *all possible inputs*. This experiment used a single frequency input — the most forgiving case — which concentrates nearly all output probability on one bitstring. Noise removes most of the signal, but the surviving peak still beats the flat noise floor. Crucially, complex superposition inputs like those in real Shor or QPE runs scatter answers across many phases and would not survive at 1.8% fidelity. Q-CTRL's blog honestly acknowledges this limitation.

Context: three parallel noise-fighting tracks in the same week

| Team | Hardware | Approach | Headline number | |------|----------|----------|-----------------| | Q-CTRL | IBM Heron r3 (156q) | Convolutional QFT + Fire Opal | 100 qubits at 1.8% process fidelity | | Mitsui / Mitsubishi Electric (Aug 20) | Quantinuum Helios (98q) | Steane-coded QFT on 12 logical qubits | 14.3% correct on physical qubits; most logical-qubit shots filtered out | | IonQ | Q70/Q102 LDPC simulation | CliNR error detection | 74x measurement speedup |

The same week also saw IBM's modular cryostat architecture announcement (Aug 19, targeting 1000+ qubits in 2027 Q2) and Quantinuum Helios landing on Oracle Cloud at 99.921% two-qubit gate fidelity (Aug 15).

Engineering significance

1. Software-layer marginal gains remain large: the record doubled in four months purely via compiler and error-suppression software, not hardware generational change. 2. The NISQ boundary extends toward useful algorithms: a system running 100-qubit QFT can already take on noise-tolerant subtasks (shallow variational circuits, targeted chemistry phase estimation, optimization preprocessing) — though breaking RSA still requires millions of logical qubits. 3. Honest framing matters: Q-CTRL did not oversell the result as 'Shor is next'; single-frequency input is a best case.

Open questions to track (6-12 months)

Does Convolutional QFT hold up on complex superposition inputs (Shor/QPE)?
Does Fire Opal remain effective at thousands of qubits as depth grows?
Can the convolutional approach transfer to HHL, amplitude estimation, QML feature maps?
Can Mitsui/Mitsubishi's logical-qubit shot-filtering cost be absorbed at scale?
Can next-gen IBM hardware push the framework to 200 qubits?
How does the approach compare on all-to-all trapped-ion hardware (Quantinuum, IonQ)?

Takeaway

The reoccurring lesson of the week: when hardware is not yet the sole bottleneck, joint design of algorithms, compilation, error suppression, and error correction determines how much useful quantum value the NISQ era can deliver. Single-shot success rate on realistic inputs, not process fidelity alone, should become a headline evaluation metric for quantum hardware news.

Sources: Q-CTRL blog "Breaking the 100-Qubit Barrier"; arXiv 2608.05435; Quantum Brief 2026-08-16/2026-08-20; Quantum Zeitgeist 2026-08; HPCwire 2026-04-16 (ParityQC 52-qubit record); Phys.org 2026-04 (previous 27-qubit record).