Skip to main content
Back to timeline
arXivSource publication:

Quantum error correction decoding no longer has to choose speed over accuracy: TCBP-JOBS matches Tesseract's logical error rates on GPU, with serial speedups up to thousands of times

Synopsis

Decoders for general quantum error-correcting codes have long traded speed against accuracy: fast decoders lose accuracy, while most-likely-error decoders are too slow for real-time correction. TCBP-JOBS uses a small multilayer perceptron to learn a calibrated probability for each fault from the full belief-propagation trajectory, then starts GPU-parallel searches over many orderings from those probabilities. Across 72 surface, color and bivariate-bicycle code circuits it reaches logical error rates comparable to or lower than Tesseract, by up to 4.5 times on the largest codes, with serial decoding 10 to 7,400 times faster than sparsified Tesseract.

Source-provided article image: TCBP-JOBS: GPU-Accelerated Quantum Error Correction Decoding and Noise Learning with Near-Optimal Accuracy
arXiv

Interpretation

General quantum error-correcting codes can now be decoded at GPU-parallel speed while keeping accuracy close to most-likely-error decoding, so speed and accuracy are no longer an either-or choice. Previously, general-code decoders were either fast or accurate: belief propagation plus ordered statistics decoding (OSD) is fast, but belief propagation often fails to converge on quantum codes and OSD searches around a single ordering, leaving an accuracy gap; most-likely-error decoders such as Tesseract are accurate but need seconds per round on bivariate-bicycle codes because of their exponential search space. Across 72 surface, color and bivariate-bicycle code circuits under SI1000 noise at physical error rates of 0.1% to 0.4%, against sparsified short-beam Tesseract, 17 of 36 code/noise settings have significantly lower logical error rate and one significantly higher, with no significant difference in the rest; the largest reduction, 4.5 times, is on the gross code at SI1000 noise.

The method works in two steps: it first corrects the fault ranking that belief propagation supplies, then runs a bounded parallel search around that ranking. Trajectory-calibrated belief propagation (TCBP) feeds each fault's marginal log-likelihood ratios over all belief-propagation iterations, plus four appended features, into a multilayer perceptron with two hidden layers of 16 units, which outputs a calibrated fault probability; the model is fitted once per code family on simulated shots in about 10 seconds and needs no experimental data. Jittered OSD beam search (JOBS) starts from these probabilities, forms 16 perturbed orderings per shot, solves OSD-0 in each and refines every solution with a bounded beam search, all on the GPU; on the 24.2-million-shot Tesseract benchmark, one GB200 GPU finishes in 45 minutes, against 130,257 CPU thread-hours for Tesseract.

Perspective

The demonstrated scope is 72 surface, color and bivariate-bicycle code memory and logical-surgery circuits at physical error rates of 0.1% to 0.4% under SI1000 noise, with serial and batched decoding on one GB200 GPU; the noise-learning data are simulated from the same detector error model used as the oracle.

At lower physical error rates, rare-event splitting shows Tesseract is more accurate on the gross code, and TCBP-JOBS fails on some four-fault configurations; the noise-learning study measures recovery of a known model rather than model mismatch on hardware; the splitting estimates lack independent verification below the range reachable by direct Monte Carlo.