Physical errors are the disease. Traditional correction was often worse than no treatment. Our decoder is closer to an immune system.
Quantum devices are error-prone by nature. Physical error correction is what turns fragile physical qubits into reliable logical ones — and reaching roughly 10,000 physical qubits is what unlocks the first practically useful applications, such as Shor’s algorithm.
Errors accumulate. The computation dies before it produces an answer.
Hand-engineered decoding rules work, but can’t reach the logical error rates useful computation needs.
Learned decoding: the system recognises and corrects errors, and improves as it sees more data.
Four compounding advantages — not one trick.
Around 8 physical qubits per logical qubit, against roughly 100:1 for surface or colour codes.
The logical error rate must fall below 10-11. qLDPC alone does not get there with rule-based decoders — that is the bottleneck we attack.
Below 10K parameters, so inference fits comfortably inside real-time control.
Sub-microsecond latency: 0.37 µs atom rearrangement against roughly 60 µs in prior published work.
Published figures, stated as they are.
FPGA atom-rearrangement latency, versus ~60 µs prior art (arXiv:2210.10364) — roughly 160× faster.
Physical-to-logical qubit ratio with qLDPC codes, versus ~100:1 for surface codes.
Small enough for real-time deployment; research decoders run into millions and billions.
The threshold for useful, fault-tolerant computation.
Figures reflect internal benchmarks and published comparisons. Ask us for the technical details. Our Graph Transformer Decoder is open source.
Benchmarks on lifted-product qLDPC codes; baselines as labelled on each figure.
We work across modalities — we do not need to predict which one wins.
