Hardware-efficient deep unfolding neural network for GLDPC decoding in high-speed optical fiber communications

Abstract

High-speed optical fiber communication systems require forward error correction decoders that combine strong error correction with low latency and efficient implementation. Generalized low-density parity-check (GLDPC) codes improve finite-length performance and lower the error floor, but their generalized check (GC) nodes require computationally intensive soft-input soft-output (SISO) decoding. We propose a hardware-efficient deep-unfolding decoder that maps outer belief-propagation (BP) iterations onto a sequence of trainable macro-layers. Each macro-layer contains a J-layer normalized offset Min-Sum subnetwork that approximates the component SISO extrinsic-information mapping. Simulation results show that the proposed architecture reduces the component-decoding core update proxy by 70% relative to exact maximum a posteriori (MAP) decoding while maintaining comparable bit error rate (BER) performance.