| name | fpga-quantum-error-decoder |
| description | FPGA-based real-time quantum error correction decoding architecture. Combines hardware-integrated NN decoders on FPGA with superconducting quantum processors for low-latency closed-loop QEC. Use when: (1) Designing real-time QEC control systems, (2) Implementing FPGA-based syndrome decoders, (3) Building low-latency feedback loops for fault-tolerant quantum computing, (4) Analyzing closed-loop latency budgets for QEC cycles, (5) Implementing mid-circuit Pauli-frame corrections in non-Clifford logical circuits. Trigger: FPGA QEC decoder, real-time quantum error correction, low-latency syndrome decoding, hardware integrated QEC, surface code FPGA, closed-loop quantum feedback.
|
FPGA-Based Real-Time Quantum Error Decoder
Hardware-integrated control architecture for real-time surface-code QEC using
FPGA-based neural network decoders, achieving deterministic closed-loop latency
of 550 ns (124 ns NN decoding) within a 1.25 μs QEC cycle.
Core Architecture
Quantum Processor → Syndrome Measurement → FPGA Controller → NN Decoder → Feedback Correction
(superconducting) (repeated) (Xilinx/Intel) (124 ns) (within 1.25 μs)
Key Metrics (from arXiv:2605.04892)
| Component | Latency | Description |
|---|
| NN Decoding | 124 ns | FPGA-based neural network inference |
| Total Closed-Loop | 550 ns | End-to-end syndrome to correction |
| QEC Cycle | 1.25 μs | Full error correction cycle period |
| Code Distance | d=3 | Surface code demonstration |
Hardware Design Principles
- Deterministic Latency: Fixed-latency FPGA pipelines avoid timing jitter
- On-Chip NN Inference: Neural network weights stored in FPGA BRAM/LUTs
- Closed-Loop Feedback: Decoder output directly drives quantum control pulses
- Mid-Circuit Correction: Supports Pauli-frame updates during logical operations
Implementation Workflow
Step 1: Neural Network Quantization for FPGA
import torch
model = load_trained_decoder()
model_int8 = torch.quantization.quantize_dynamic(
model, {torch.nn.Linear, torch.nn.Conv2d}, dtype=torch.qint8
)
torch.onnx.export(model_int8, dummy_syndrome, "decoder_int8.onnx")
Step 2: FPGA Pipeline Design
Input Layer (syndrome bits) → [FPGA registers]
↓
Convolution Layer → [DSP slices + BRAM for weights]
↓
Activation (ReLU) → [LUT-based lookup]
↓
Output Layer (correction bits) → [Pipeline registers]
↓
Control Signal → Quantum processor feedback
Latency Budget Breakdown:
- Syndrome readout: ~100 ns
- NN inference (INT8): 124 ns
- Correction computation: ~50 ns
- Control signal routing: ~100 ns
- Total: 550 ns
Step 3: QEC Cycle Integration
def qec_cycle(fpga_controller):
syndromes = measure_stabilizers()
corrections = fpga_controller.decode(syndromes)
if corrections.needs_feedback:
apply_physical_correction(corrections)
else:
update_pauli_frame(corrections)
assert elapsed_time() < 1250
Critical Design Considerations
Latency vs Accuracy Tradeoff
| Decoder Latency | Logical Error Rate | Use Case |
|---|
| < 200 ns | ~0.1% degradation | Real-time feedback |
| < 1 μs | ~0.01% degradation | Near-real-time |
| > 10 μs | Baseline (offline) | Post-processing |
Scaling to Larger Distances
For distance-5 and distance-7 surface codes:
- Syndrome data size grows as d²
- NN model size grows proportionally
- FPGA resource utilization: LUTs, DSPs, BRAM
- Pipeline parallelism is key: decode multiple syndrome patches concurrently
Mid-Circuit vs Pauli-Frame Correction
- Pauli-frame updating: Software-only, zero latency, works for Clifford circuits
- Mid-circuit feedback: Required for non-Clifford gates (T-gates, magic state distillation)
- FPGA decoder enables mid-circuit correction when Pauli-frame alone is insufficient
Error Model Adaptation
The FPGA-based NN decoder adapts to varying error conditions:
Training: Offline NN training with diverse noise models
Deployment: FPGA inference adapts to real-time error statistics
Adaptation: Periodic model updates based on observed syndrome patterns
Robustness Features
- Error condition variation: Maintains performance across different physical error rates
- Calibration drift: Retraining triggers when logical error rate exceeds threshold
- Crosstalk compensation: NN naturally learns spatial error correlations
Comparison with Traditional Decoders
| Feature | MWPM | BP | FPGA-NN Decoder |
|---|
| Latency | ms-scale | μs-scale | 124 ns |
| Hardware | CPU/GPU | CPU/FPGA | FPGA (dedicated) |
| Adaptability | Fixed model | Iterative | Learned patterns |
| Circuit-level noise | Requires modification | Limited | Native support |
| Real-time feedback | No | Marginal | Yes |
Resources
- arXiv:2605.04892: "Real-time Surface-Code Error Correction Using an FPGA-based Neural-Network Decoder"
- arXiv:2605.04459: "Triage: Adaptive Parallel Window Decoding Scheduler"
- Related skill:
neural-decoder-quantum-error-correction (algorithm-level)
- Related skill:
quantum-fault-tolerance-verification (verification methods)