arXiv:2611.04179
Hardware-Efficient Reinforcement-Learned Scheduling for Quantum LDPC Decoders
Mohsen Moradi, Vahid Nourozi, Wei Tang, David Mitchell
Abstract
Quantum low-density parity-check (QLDPC) codes are promising for fault-tolerant quantum computing, but their decoding algorithms present significant hardware implementation challenges. Reinforcement-learned sequential belief propagation (RL-S) improves the error-correction performance of conventional belief propagation (BP) through adaptive variable-node scheduling. However, its heap-based priority management introduces substantial hardware overhead. This paper presents hardware implementations of BP, RL-S, and bucket-based reinforcement-learned BP (BRL-S) for the [[288,12,18]] bivariate bicycle QLDPC code. The BRL-S architectures employ quantized priorities and first-in, first-out (FIFO) buckets to simplify scheduling. Three configurations with two, four, and eight priority levels are evaluated. FPGA synthesis results demonstrate that BRL-S reduces lookup-table utilization by approximately 68--72% compared with RL-S. ASIC synthesis further shows reductions of 63--67% in area and 64--69% in estimated power consumption. These results demonstrate that bucket-based scheduling substantially improves the hardware efficiency of reinforcement-learned QLDPC decoders.
AI Summary
The paper examines the hardware cost of adaptive scheduling in reinforcement-learned sequential belief propagation (RL-S) for QLDPC decoding. It implements flooding BP, heap-scheduled RL-S, and bucket-based RL-S (BRL-S) for the $[[288,12,18]]$ bivariate bicycle code. BRL-S quantizes pretrained scheduling scores into two, four, or eight priority levels and replaces max-heap maintenance with FIFO buckets implemented using linked-list pointers. RL-S and BRL-S share the same sequential sum-product message-update framework, so their hardware comparison primarily measures scheduling overhead. Depolarizing-channel simulations use at most 100 iterations, with each sequential iteration comprising up to 288 variable-node updates. The reported block-error-rate curves place four- and eight-bucket BRL-S close to RL-S, while two buckets produce some degradation, particularly at lower physical error probabilities. All sequential variants outperform flooding BP in the reported comparison. Synthesizable Verilog implementations are evaluated through FPGA and 45-nm ASIC synthesis. BRL-S uses 19,873–22,078 FPGA LUTs versus 69,737 for RL-S, a reduction of 68.3–71.5%; flip-flop counts fall from 10,536 to approximately 5,800. ASIC cell area decreases from $0.4564\,\mathrm{mm}^2$ to $0.1485$–$0.1712\,\mathrm{mm}^2$, and estimated power decreases from 3.766 mW to 1.175–1.366 mW. The evaluation is restricted to one code and three bucket counts. FPGA results are explicitly post-synthesis estimates rather than completed implementations; ASIC power is estimated, and clock periods are inferred from synthesis slack under a 50-ns constraint.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.