What is a quantum error correction decoder and why is it a real-time classical problem?

TL;DR (status as of October 19, 2026): As of October 19, 2026, the only real-time decoder measured in a below-threshold quantum error correction experiment, in the dated record Rosetta Quantum compiled for this entry, is Google’s: it held a mean latency of 63 ± 17 µs while new syndrome rounds arrived every 1.1 µs on a 72-qubit superconducting processor (arXiv:2408.13687). A decoder is the classical algorithm that turns a quantum computer’s error-syndrome measurements into corrections, and it is a real-time problem by the experimenters’ own rule: decoding throughput «must keep pace» with the 1.1 µs cycle, or the syndrome backlog grows and the computation slows exponentially. The most accurate decoder published, Google DeepMind’s AlphaQubit, makes 6% fewer errors than tensor-network methods — and its authors state it is «still too slow» to correct errors in real time (blog.google, 2024-11-20). Rosetta Quantum reads the decoder race as the clearest classical bottleneck inside every quantum roadmap: at a million physical qubits, Riverlane’s CEO forecasts syndrome traffic near 100 TB/s (Ars Technica, 2024-07-09).

What is a decoder?

The classical algorithm that keeps a logical qubit alive. Quantum error correction splits the work in two: the quantum hardware runs stabilizer measurements every cycle and emits a stream of bits — the syndromes — that flag where errors may have struck; a classical computer then runs the decoder, an inference algorithm (minimum-weight perfect matching, union-find, belief propagation, or a neural network) that reads those syndromes and infers which correction to apply. The logical qubit a vendor advertises does not exist without this second half: the physical-to-logical ratio our ledger tracks assumes the decoder does its job, and our audit of who does the work in hybrid systems applies verbatim here — inside every «quantum» logical qubit there is a classical computer working at full speed. What makes the decoder unusual among classical problems is not its difficulty at rest but its deadline, and that deadline is measured.

Why is it a real-time problem?

Because the deadline repeats every microsecond. In Google’s below-threshold experiment, a new round of syndromes arrives every 1.1 µs, and the paper states the rule plainly: the throughput of transmitting, processing and decoding the syndrome information «must keep pace with the fast error correcting cycle time of 1.1 µs» — and its supplementary material adds that failure «causes syndrome backlog to accumulate», which can slow the computation exponentially (arXiv:2408.13687). For a memory experiment you can decode offline, after the run; but the moment a program branches on a logical measurement — feed-forward, the non-Clifford operations every useful algorithm needs — the correction frame must be resolved before the next step, so decoder latency becomes a floor under logical clock speed. The budget depends on the hardware’s own pace: superconducting qubits cycle in about a microsecond, while trapped ions run thousands of times slower (our architecture comparison), which is why Quantinuum can declare a real-time decode budget of roughly 2 ms for Helios (launch release, 2025-11-05, verified in our October 16 run — our Helios spec sheet) while Google’s decoder must answer in the tens of microseconds. Same problem, three orders of magnitude apart in deadline.

What is measured today?

One experiment anchors the real-time column. Google’s real-time decoder held a mean latency of 63 ± 17 µs, roughly constant across runs up to 1.1 seconds — a million cycles — at code distance 5 on a 72-qubit processor, while staying below threshold with an error-suppression factor Λ = 2.0 ± 0.1 (arXiv:2408.13687). Every other number in the dated record Rosetta Quantum compiled below measures a piece of the pipeline, not the closed loop: decoder chips, GPU throughput, transport latency, forecasts. One row per number, source beside each:

The numbers in this table are not mutually comparable — different codes, different systems, different definitions of latency — and printing them side by side is a reading aid, not a benchmark. The only closed-loop, below-threshold, real-time figure among them is the first row.

Why do accuracy and speed pull apart?

Because the same experiment measured both and printed the gap. Google’s real-time decoder reached a logical error per cycle of 0.35% ± 0.01% at d=5; its offline neural-network decoder, with no deadline, reached 0.269% ± 0.008% on the same data (arXiv:2408.13687). That spread — about 30% more logical error in exchange for answering in 63 µs — is the measured price of real time. AlphaQubit sits at the far end of the same trade: Google DeepMind reports it makes 6% fewer errors than tensor-network methods, which the team itself calls «highly accurate but impractically slow», and 30% fewer errors than correlated matching, the fast scalable baseline — and then states the limit in its own words: it is «still too slow to correct errors in a superconducting processor in real time» (blog.google, 2024-11-20). The trade matters economically, not just technically: a more accurate decoder raises the effective error suppression, which lowers the code distance — and therefore the number of physical qubits — needed per logical qubit. The decoder quietly moves the physical-to-logical ratio that converts roadmaps into dates.

How big does the classical side get at scale?

Big enough that a hardware industry is forming around it. Riverlane’s CEO put the scale forecast on record: «At a million physical qubits, we’ll be processing about a hundred terabytes per second, which is Netflix global streaming» (Ars Technica, 2024-07-09) — a forecast, not a measurement, and we print it as such. The engineering responses are already measurable. Riverlane’s Collision Clustering decoder runs at MHz decoding speed on an FPGA, and as an ASIC occupies 0.06 mm² and draws 8 mW while handling surface codes up to 1,057 qubits (arXiv:2309.05558, published in Nature Electronics; The Quantum Insider, 2025-01-07); its current Deltaflow 2 product page claims real-time error correction on up to 250 physical qubits (riverlane.com, accessed 2026-10-19). NVIDIA’s bet runs through general-purpose GPUs: CUDA-Q QEC ships GPU-accelerated Relay-BP decoding at 1.6 million and 500,000 iterations per second on the two gross qLDPC codes, plus a TensorRT path for AI decoders and sliding-window decoding (NVIDIA dev blog, 2025-12-17), with NVQLink as the transport whose measured round-trip is 3.96 µs (max) — a number obtained with no quantum processor in the loop, as our NVQLink entry documents; public production closed-loop QEC deployments through it stood at zero as of our October 17 run. Whether a decoder should live in an FPGA next to the fridge or in a GPU across a network link is, today, an open engineering argument with dated evidence on both sides — exactly the kind the Rosetta ledger exists to keep score on.

What should a reader do with a «logical qubit» announcement?

Ask three questions the press release usually skips. Was the decoding real-time or offline? — offline decoding is legitimate science and says nothing about running algorithms with feed-forward. What was the decoder’s latency against the machine’s own cycle time? — 63 µs is a result against a 1.1 µs cycle on superconducting hardware and would be trivial against a 2 ms trapped-ion budget. And which decoder produced the headline fidelity? — if the answer is a slow offline decoder, the real-time number, when it exists, will be worse, by a margin the Google experiment lets you estimate at tens of percent (arXiv:2408.13687). The future tense matters too: Riverlane’s roadmap aims at a million real-time quantum operations by 2026 (riverlane.com, accessed 2026-10-19), and NVIDIA’s architecture targets latency tolerances of tens of microseconds (arXiv:2510.25213) — both will count as measured when a closed loop on quantum hardware publishes its numbers, and not before.

What we don’t know

• AlphaQubit’s actual speed. Google states it is too slow for real time but publishes no latency figure in the announcement, so the size of that gap is unmeasured in the sources we could read.
• Whether a real-time, below-threshold decode has been published above distance 5. The d=7 result in the same Google paper (0.143% ± 0.003% per cycle) came from the offline neural-network decoder. Our scan in this run found no larger real-time demonstration — a statement about our search, not about the field.
• Cross-architecture comparability. The latencies in our table come from different codes, systems and definitions (mean latency, max round-trip, throughput). They share a page, not a benchmark.
• Whether GPU throughput survives a closed loop. The 3.96 µs round-trip had no QPU in the loop, and CUDA-Q throughput figures are decoder-only. The closed-loop number on real quantum hardware via that stack is not public.
• The 100 TB/s figure. It is a CEO forecast for machines that do not exist yet. We print it dated and attributed, not as a measurement.

Sources

• Google Quantum AI and Collaborators — «Quantum error correction below the surface code threshold» (arXiv:2408.13687; published in Nature): real-time decoder 63 ± 17 µs at d=5, cycle 1.1 µs — arxiv.org
• Google DeepMind / Google Quantum AI — «AlphaQubit tackles one of quantum computing’s biggest challenges» (2024-11-20) — blog.google
• Barber et al. — «A real-time, scalable, fast and highly resource efficient decoder for a quantum computer» (arXiv:2309.05558; Nature Electronics, DOI 10.1038/s41928-024-01319-5) — arxiv.org
• The Quantum Insider — «Riverlane Reports Decoder Tech Sets New Standard in Quantum Error Correction» (2025-01-07) — thequantuminsider.com
• John Timmer, Ars Technica (republished by Cambridge Innovation Capital) — «Why every quantum computer will need a powerful classical computer» (2024-07-09): Brierley, «about a hundred terabytes per second» — www.cic.vc
• Riverlane — Deltaflow Decode product page (accessed 2026-10-19): «real-time error correction on up to 250 physical qubits» — riverlane.com
• NVIDIA — «Real-Time Decoding, Algorithmic GPU Decoders, and AI Inference Enhancements in NVIDIA CUDA-Q QEC» (2025-12-17) — developer.nvidia.com
• «Platform Architecture for Tight Coupling of High-Performance Computing with Quantum Processors» (arXiv:2510.25213, v2 2025-11-03): round-trip latency 3.96 µs (max) — arxiv.org
• Quantinuum — Helios commercial launch press release (2025-11-05): ≈2 ms real-time decode budget (verified in our 2026-10-16 run; the page did not load in this run) — www.quantinuum.com
• Rosetta Quantum — related ledger entries: #33 logical qubits and the ratio · #20 superconducting vs trapped ion · #44 who does the work in hybrid systems · #82 Helios, measured vs promised · #83 what NVQLink actually connects

Rosetta Quantum publishes verdicts with raw, reproducible data. This is educational content, not a product claim. Quantum, independently measured.