Skip to main content
Back

Rapids Squeezes a 240x Better Quantum Error Decoder onto One FPGA

A co-designed FPGA neural decoder called Rapids keeps pace with superconducting qubits while cutting logical errors 240 times below prior real-time decoders.

Paper, Full Access
  • Rapids is a streaming neural decoder co-designed with an FPGA for real-time quantum error correction (paper).
  • At distance 19 on the surface code, it hits sub-microsecond per round and 20 µs latency on one FPGA.
  • Logical error rate is 240x lower than existing real-time hardware decoders like AQ2-RT, Willow RT and union-find.
  • On the [[144,12,12]] Gross code, Rapids is 16x more accurate than the real-time int4 Relay decoder.
  • Depthwise convs, Monarch matrices, 8-bit activations on DSPs and 4-bit weights in LUTs enable single-FPGA fit.
  • Routed FPGA builds are released on Zenodo; 7 nm ASIC estimated at 7.4 mm² and 10 W per decoder.

Fault-tolerant superconducting processors need classical hardware that can decode a new error-correction syndrome about every microsecond. In the Rapids paper, researchers report a streaming neural decoder that sustains that input rate on one AMD Versal Premium VP1802 FPGA. Their distance-19 surface-code implementation returns a decision in 20.36 µs and achieves a logical error rate 240-fold below the real-time hardware decoders evaluated in the study.

A 1 MHz queue has no slack

Quantum error correction encodes one logical qubit across many physical qubits, then repeatedly measures parity checks called syndromes. A decoder converts the resulting bit stream into a likely correction, usually recorded in a Pauli frame rather than applied directly to the hardware. For a surface code, the distance d determines the lattice size and the minimum length of an undetectable error chain. Larger distances improve protection while increasing the amount of syndrome data and decoding work.

A decoder serving 1 µs syndrome cycles must accept complete rounds at 1 MHz. Any lower sustained rate creates an unbounded queue that eventually stalls the processor. Two measurements describe the decoder’s performance:

  • Throughput: how often the decoder can accept another syndrome round.
  • Latency: how long one round takes to travel from input to decision.

Pipelining allows Rapids to accept a new round every microsecond while earlier rounds remain in flight, producing the reported 20.36 µs decision latency. Closed-loop operations still inherit that latency. Pauli-frame tracking can defer many corrections, while conditional operations and final readout may need a decoded result before proceeding.

Earlier real-time hardware decoders have favored algorithms such as union-find or small neural networks. An FPGA LSTM decoder reports 550 ns deterministic closed-loop latency, including 124 ns for neural inference, at distance 3. Scaling to distance 19 expands the syndrome lattice and temporal context, making that small-code result difficult to preserve.

An FPGA-shaped neural network

Rapids extends Cascade, a convolutional decoder with near-optimal benchmark accuracy, by adding a compact recurrent network. Convolutions process the spatial syndrome lattice, while recurrent state carries information between adjacent time windows. This parallel-window design consumes syndrome data as it arrives instead of waiting for an entire experiment to finish.