# Preprint says Rapids decodes distance-19 code in under a microsecond per round

Published 2026-11-20 · Quantum.world Team · Quantum.world

> A preprint describes Rapids, a neural decoder designed for FPGA hardware, processing more than one million surface-code syndrome rounds per second at distance 19. Its authors report lower logical error rates than existing real-time decoders, alongside latency figures that vary with the hardware configuration.

## FPGA implementation targets real-time decoding

A single FPGA can process more than one million syndrome rounds per second for a distance-19 surface code using the Rapids decoder, according to a new preprint. Submitted to [arXiv](https://arxiv.org/abs/2611.26447) on November 19, the paper by Chenyang Jin, Susanne F. Yelin and Andi Gu describes a streaming neural decoder built to meet the speed and accuracy demands of quantum error correction.

Rapids is designed to process syndrome measurements as an ongoing stream rather than waiting for a complete memory experiment. The authors co-designed its network for field-programmable gate array (FPGA) hardware, using operations that map to the device and training the network in integer arithmetic. They report that the distance-19 implementation keeps pace with a one-microsecond syndrome period on one FPGA.

## Speed and decision latency differ

Processing speed is not the same as decision latency, the time between readout and the decoder returning a correction. The paper gives a decision latency of 20.36 microseconds for the compact, single-FPGA distance-19 implementation, most of it arising from buffering syndrome rounds.

A faster version processes two sites per clock cycle and uses twice the FPGA hardware. The authors report that it processes a round in 0.49 microseconds and has a decision latency of 13 microseconds. They estimate that spreading the decoder across more FPGAs could reduce distance-19 latency only to about 8 microseconds, where chip-to-chip communication becomes the limiting factor.

## Reported accuracy across codes

In surface-code memory experiments under SI1000 noise at a physical error rate of 0.15%, the paper reports a logical error rate 240 times lower than union-find, a decoder the authors identify as previously demonstrated in real time at that code distance. They also report a rate nearly 10 times lower than minimum-weight perfect matching. These are results reported by the preprint, not measurements on a quantum processor.

The paper also evaluates Rapids on the Gross bivariate-bicycle code and a color code. For the Gross code, at a physical error rate of 0.1%, the authors say their most accurate decoder is 16 times more accurate than the real-time Relay-BP-1 decoder; they give that decoder a decision latency of 8.0 microseconds. For the color code, they report that a distance-15 build spanning two FPGAs processes a round in 0.17 microseconds, with a decision latency of 6.3 microseconds.

## Hardware scaling and scope

Rapids uses sparse neural-network layers and low-bit integer arithmetic to fit the FPGA’s hardware resources, according to the paper. Its convolutional network is pipelined across rounds, while a small recurrent network at the end accumulates information into a prediction of the logical error. The authors say the design can be divided across multiple FPGAs or adapted to custom silicon.

The preprint estimates that a custom-silicon implementation of the distance-19 decoder could process a round in 0.30 microseconds with 6.5 microseconds of decision latency. That estimate uses a 7-nanometer process node; it is not a reported measurement from a fabricated chip. The paper’s evaluations use FPGA implementations and simulated noise models for the surface, Gross and color codes.

## Sources

1. Original source · Preprint [arXiv](https://arxiv.org/abs/2611.26447) arXiv:2611.26447

## Related stories

- [Preprint details FPGA platform for testing QEC decoders](https://quantum.world/news/preprint-fpga-platform-qec-decoder-testing-cheq)
- [Google preprint links Willow’s lingering errors to third-state leakage](https://quantum.world/news/google-willow-temporally-correlated-error-suppression)
- [Preprint proposes code quilts for logical measurements without extra qubits](https://quantum.world/news/preprint-code-quilts-logical-measurements-no-extra-qubits)
- [IBM says million-round repetition code can run in one circuit](https://quantum.world/news/ibm-million-round-repetition-code-single-circuit)
- [Preprint extends GKP error correction to measurement sequences](https://quantum.world/news/dynamical-gkp-codes-measurement-sequences)

## Explore

[More articles](https://quantum.world/news)

---

Canonical: https://quantum.world/news/rapids-decoder-distance-19-sub-microsecond-qec
