Quantum Error Correction: IonQ Decoder Runs in Real Time on One CPU

A decoder on one CPU

On September 22, 2026, IonQ announced what it calls the industry's first end-to-end real-time quantum error correction decoder, and said the decoder runs on a single off-the-shelf CPU.7 The company's case is about a classical problem. In conventional setups, the computers that interpret error signals can fall behind, and when that happens the quantum processor has to pause until they catch up.7 IonQ says one standard processor can handle that work continuously in the background, so the quantum machine keeps running at full speed.7

The numbers come from an arXiv paper. IonQ tested a dual-decoder architecture on benchmark circuits that simulated up to 408 logical qubits spread across 88 memory blocks and magic-state factories, running more than 31.5 million quantum operations, which puts the test at the "MegaQuOp" scale.7 Under standard noise, the decoder added as little as 0.02% "stretch" time, which means decoding barely slowed the overall computation.7 IonQ says the result supports its proprietary Walking Cat architecture and its roadmap beyond 256 physical qubits toward platforms with thousands of qubits.7

Some enthusiast coverage calls the result a "monumental leap" that removes a major barrier to large-scale fault tolerance.1 That outlet does concede that the work does not mean fault-tolerant machines exist yet.1 The more careful reading is that IonQ validated the classical half of the problem against simulated circuits, not against hundreds of physical logical qubits running on its own hardware. That still matters. Commodity processors could keep pace with error correction without exotic, custom control electronics, and the cost of the classical stack would not grow out of control as quantum systems get bigger.7

Why the decoder matters

Error correction spreads one logical qubit across many physical qubits. The system cannot measure the quantum state directly, so it pulls out "syndromes" by entangling data qubits with auxiliary qubits and measuring the auxiliaries.5 A classical decoder then turns that stream of syndromes into corrections.4 If the decoder is too slow, it builds a backlog that cancels out the benefit of correcting errors at all.

Microsoft made the same point in a September 22 essay that defines "scalable logical qubits." Its bar is logical qubits kept alive by repeated error correction, able to run a universal set of fault-tolerant operations with low-latency real-time decoding and feedback, and replicable into the hundreds or thousands.5 Microsoft names the trade-off between decoder accuracy and latency as a core engineering tension. Some decoders correct more accurately but cost more compute, and others give up accuracy for speed.5 The company also says utility-scale workloads need logical error rates around 10^-12 to 10^-15.5

Other groups are working on the same bottleneck from different directions. Researchers at RWTH Aachen and the University of Basel proposed "Cluster-As-You-Go," a modified Union-Find decoder that starts clustering errors while syndromes are still being measured. In surface-code simulations it cut post-measurement idling by more than 30%, at a small cost in accuracy.9 The authors say hardware timing benchmarks are still needed to confirm it works in practice.9 Google and DeepMind have taken a machine-learning route with AlphaQubit, which reportedly beat established decoders on certain benchmarks.4

Better codes as well as faster decoders

The second line of progress is the codes themselves. A UC Berkeley team built quantum CSS codes from classical "repeat multiple accumulate" codes. The encoding circuits have logarithmic depth and a linear gate count, and after four accumulation rounds the codes come within 0.001 of the CSS Gilbert–Varshamov bound.3 Tencent Quantum Laboratory reported the first rate-optimal "phantom" qLDPC code families. Some logical gates in these codes can be performed just by permuting physical qubits.10 The Tencent researchers also showed that checking whether a code has this phantom property is as hard as the Graph Isomorphism problem, and the coverage says the results do not yet show practical thresholds at larger sizes.10

These are theory results, and they should be read that way. Still, they point the same way as the hardware roadmaps: less overhead per logical qubit. Japan's Moonshot program reports that applying error mitigation to logical qubits can cut physical-qubit needs by as much as 80% in favorable cases. It also reports a concatenated-code architecture whose space and time overheads are smaller than any other known architecture.6

Where the roadmaps stand

The decoder news arrives at a busy point for vendor milestones. IBM is betting on qLDPC "bivariate bicycle" codes instead of the surface code most of the field uses. Its plan runs through the Kookaburra module in 2026 to Starling, a 200-logical-qubit, 100-million-gate machine in Poughkeepsie due in 2029.11 As of late September, IBM had not announced that Kookaburra was delivered, and it has usually unveiled processors at its November developer conference.11 Separately, IBM and the University of Chicago reported a July 2026 demonstration with 70 logical qubits on a problem they described as beyond practical classical simulation.12 The advantage claims had not finished peer review at last report.18

Google's Willow chip still anchors the field's belief that error correction works, because it showed logical errors falling as the code grew.11 Google has not declared its next milestone, a long-lived logical qubit, complete. In March 2026 it added a neutral-atom hardware team in Boulder.11 Quantinuum reported a September 2026 "Helix" result in which logical Clifford operations beat physical ones by 4.28x without post-selection, at 4.6×10^-5 error per logical qubit per correction cycle.19 It also lists Sol for 2027 and Apollo, its fault-tolerant system, for 2029.19 IonQ's own stated targets are 800 logical qubits in 2027 and 80,000 logical qubits on 2 million physical qubits by 2030.18

Logical-qubit counts depend on who is counting

The biggest disagreement in the coverage is over how many logical qubits anyone has actually built. Several outlets repeat Quantinuum's 48 error-corrected logical qubits on Helios and QuEra's 96.1718 A stricter leaderboard notes that Quantinuum's 48 came from runs that kept only 14% to 31% of attempts, and it credits the company with only 3 counted error-correction logical qubits as of October 2026.13 Even the QuEra paper describes its deep-circuit results as "dozens" of logical qubits, so 96 is the peak of a single experiment.14 One commentator puts it bluntly: a qubit count that does not say "logical, with error rate" is marketing.20

That dispute is why Microsoft's definition matters. It is also why a result like IonQ's should be judged on whether it holds up on real hardware, not on the size of its simulated benchmark.

The cryptography stakes

Cheaper error correction does not stay inside physics labs. Published estimates of the physical qubits needed to break RSA-2048 fell from about one billion in 2012 to 20 million in 2019, to roughly 900,000 in 2025, and to between about 11,000 and 471,000 in 2026, depending on assumptions.21 Google researchers estimated that 256-bit elliptic-curve cryptography could be broken with fewer than 1,200 logical qubits and under 500,000 physical qubits.28 An Oratomic–Caltech team put the floor at around 10,000 reconfigurable atomic qubits, with much longer runtimes.23 The largest neutral-atom array demonstrated so far holds about 6,100 atoms and has not been used for computation.23

Policy has responded with dates. A June 2026 U.S. executive order requires post-quantum key establishment for high-value federal systems by the end of 2030 and post-quantum signatures by the end of 2031.23 Google, Cloudflare and Microsoft have all set 2029 targets for their own migrations.28 Actual deployment lags behind. About 54% of the top million websites look quantum-safe, but the figure drops to 22% once Cloudflare-hosted domains are removed.22 Only 7% of organizations have deployed quantum-safe cryptography across most of their certificates.25

Assessment

The most useful way to read this month's news is that fault tolerance is now mostly an engineering problem in systems integration. IonQ's CPU-based decoder, the latency work at Aachen, and the new code constructions all reduce the overhead that sits between physical qubits and useful logical ones. None of them shows a working fault-tolerant computer yet, and the disagreement over logical-qubit counts means vendor claims need independent checking. For security teams the conclusion is simpler: the regulatory deadlines are already set, so they should not wait for the hardware question to be resolved.