- IonQ researchers say a decoding pipeline on one Apple M4 Max laptop chip kept up with error data from a simulated 408-logical-qubit workload.
- It’s a simulation and an arXiv preprint (posted August 25, 2026), not a live quantum computer and not peer reviewed.
- The trick works because trapped-ion machines give a decoder milliseconds per cycle. Superconducting machines don’t.
- Live-hardware testing is targeted for 2027, per Post Quantum.
A quantum computing startup that could never afford a rack of custom decoding hardware may not need to. IonQ researchers tested whether the kind of chip that runs a video-editing app could keep up with one of quantum computing’s biggest bottlenecks, and in simulation it did.
What IonQ actually ran
Quantum computers only work if they catch errors before those errors pile up. Physical qubits are noisy, so a working system must constantly check for mistakes and fix them without falling behind. That decoding step has often demanded serious hardware: custom chips built just to keep pace. IonQ’s paper describes a decoding pipeline that runs on a single Apple M4 Max chip, the same processor found in a high-end laptop. The team used 12 of its 16 cores: eight for error decoders and four for outcome decoders. No custom silicon, and no bank of FPGAs. (IonQ’s own announcement came later, in late September, according to Post Quantum.)
The largest benchmark simulated 408 logical qubits, using 68 memory blocks plus 20 magic state factories. The paper estimates that configuration would need 11,680 physical qubits, but no such machine exists. In that run, the decoder handled about 555,000 T gates and about 1.31 million logical measurements. A separate 102-qubit workload had a larger count of about 1.08 million T gates and 1.10 million measurements. The decoder kept up with the incoming stream of error data in both.
Keeping up is the point. Quantum processors generate a constant stream of “syndrome” data, signals about which qubits might have flipped. If the classical decoder can’t keep up, errors pile up faster than the system can correct them and the computation stalls. In IonQ’s simulations, the decoder added less than 0.3% delay at a two-qubit error rate of one in 10,000, and at most about 11.5% delay when the error rate was five times higher. Two details are easy to miss. The 1-to-5-millisecond cycle time is an assumed budget for trapped-ion hardware (1 ms for the 102-qubit runs, 5 ms for the 408-qubit run), not something measured. And all of the error data came from a noise simulation, not a live device, with timings that leave out the controller electronics and the transfer of data from them.
Why a laptop chip can do this, and how it compares
It comes down to timing. Trapped-ion machines like IonQ’s complete an error-checking cycle in milliseconds. Superconducting chips, like Google’s, take about a microsecond. That gives IonQ’s decoder roughly 900 to 4,500 times more time per decision, according to Post Quantum’s analysis. The same laptop chip wouldn’t keep up with a superconducting machine, which is why those groups use specialized hardware. Within that limit, the result suggests classical decoding may not be the bottleneck for trapped-ion systems at this scale. IonQ estimates that three 32-core CPUs could eventually decode errors for 1,000 logical qubits, though that’s a projection, not a demonstration.
| Approach | Hardware | What was shown | Status |
|---|---|---|---|
| IonQ’s new pipeline | One Apple M4 Max (12 of 16 cores) | Up to 408 logical qubits, simulated error data | arXiv preprint, simulation |
| IonQ’s earlier beam-search decoder | One Apple M3 Pro core | Under 1 ms at the 99.9th percentile; 17-fold lower logical error rate than a standard decoder on one code family | IonQ blog, Jan 2026 |
| Google (2024) | Real-time decoder on live superconducting chip | 63 microseconds average latency at code distance 5 | arXiv paper |
| Riverlane (2024) | FPGA decoder on live superconducting qubits | Under 1 microsecond mean decode time per round, 8 qubits | arXiv paper |
| IonQ’s 1,000-qubit estimate | Three 32-core CPUs | Projected, not demonstrated | IonQ estimate |
The “17-fold” error reduction in IonQ’s January research announcement belongs to a different, earlier decoder: a beam-search decoder tested on one family of codes against the industry-standard BP-OSD method. It is not the decoder in the new paper.
Other teams are chasing the same problem from different angles. We recently covered an AI decoder trained on synthetic data that tops matching benchmarks, the company says, and Photonic’s new error-correcting codes that Photonic says cut the physical qubit overhead a system needs. Decoding speed, code efficiency and qubit overhead are three sides of the same scaling problem. IonQ’s announcement ties the result to its Walking Cat architecture and a roadmap that starts with a 256-physical-qubit machine, which are company plans. Nicolas Delfosse, who leads quantum research at IonQ and co-wrote the paper, calls the single-CPU result “a practical path to commercial-scale fault-tolerant quantum computing.” That is IonQ’s own framing, not an independent finding.
Why It Matters to You
You can’t use this today, and there’s no product. If the approach holds up, it could lower the hardware bar for trapped-ion quantum systems, so a lab without a budget line for specialized decoding hardware (picture a hypothetical computer-science student) could lean on the same kind of chip already sitting on a desk. Researchers and quantum hardware builders would feel any benefit first, and only after live-hardware tests, which IonQ targets for 2027 on its Superion 10K system.
What’s still unproven
- It’s a preprint. The paper hasn’t been peer reviewed, and the press coverage mostly traces back to IonQ’s own announcement.
- It’s a simulation. The error data and the 408 logical qubits are simulated.
- It shows speed, not accuracy. The paper reports no logical error rates for these three workloads, and the benchmarks are single steps of larger algorithms, not full simulations.
- Headline numbers are best cases. The 0.02% figure is the lowest of 15 settings, and the “31.5 million operations” figure counts error-checking cycles, not gates. At the 99.9th percentile, decoding stalled for up to 32 cycles (about 160 milliseconds).
- It isn’t the first real-time decoder. Google and Riverlane have already run real-time decoding on live hardware. “First” is IonQ’s own description.
- Independent review is thin. The most detailed outside critique we found is from analyst Marin Ivezic at Post Quantum. We found no named academic commenting.
- 10,000 qubits is an extrapolation. The paper concludes decoding shouldn’t block progress toward trapped-ion machines of 10,000-plus physical qubits, but that wasn’t tested.
- Fault tolerance isn’t solved. The result suggests the classical decoding load looks manageable at this scale, while hardware error rates, accuracy at scale and live integration remain open.
Quick answers
What is a decoder, and why must it be fast?
Quantum computers constantly measure clues about errors. A decoder is the classical software that reads those clues and works out what went wrong. If it falls behind, data piles up and the computation slows or stalls.
Why does a laptop chip work for IonQ but not for Google?
Trapped-ion cycles take milliseconds, while superconducting cycles take about a microsecond. That’s roughly 900 to 4,500 times more time per decision for IonQ.
Was this run on a real quantum computer?
No. It used simulated noisy data fed into the decoder. IonQ targets live-hardware tests in 2027.
Sources
- IonQ: IonQ Demonstrates Industry’s First End-to-End Real-Time Quantum Error Decoder (Sep 22, 2026)
- arXiv 2608.25027: Real-time decoder for a MegaQuOp quantum computer using a single CPU (Aug 25, 2026; revised Sep 3)
- The Quantum Insider: IonQ Demonstrates Industry’s First End-to-End Real-Time Quantum Error Decoder (Sep 23, 2026)
- Post Quantum: IonQ real-time decoder analysis (Sep 23, 2026)
- Quantum Computing Report: IonQ demonstrates real-time QEC decoding (Aug 27, 2026)
- Pulse 2.0: IonQ real-time decoder on a single CPU (Sep 23, 2026)
- IonQ: Breaking the decoding bottleneck (Jan 8, 2026)
- Riverlane: Why the decoder is the beating heart of QEC (Apr 14, 2026)
- arXiv 2410.05202: Riverlane real-time decoding on live hardware (Oct 2024)
- arXiv 2408.13687: Google, quantum error correction below the surface code threshold (Aug 2024)
- arXiv 2502.17368: Preskill, megaquop machines (Feb 2025)
