Memory-efficient MLSE with pre-decision-driven pruning and metric sharing for bandwidth-constrained baud-rate sampling coherent optical interconnects

The surging demand for intra-data center traffic necessitates optical interconnects that deliver high capacity, minimal latency, and exceptional power efficiency. Faster-than-Nyquist (FTN) signaling based on baud-rate sampling (BRS) offers a workable solution to power-efficient data center interconnects, yet its reliance on computationally intensive maximum likelihood sequence estimation (MLSE) poses significant challenges. To overcome these limitations, we propose the memory-efficient MLSE with pre-decision-driven pruning and metric sharing (MEPS-MLSE), a hardware-efficient algorithm that enhances computational efficiency through a synergistic three-dimensional optimization framework. By integrating state memory optimization, pre-decision region partitioning-based state pruning, and branch metric (BM) space sharing, MEPS-MLSE achieves a 94.06% reduction in computational complexity compared to conventional full-size MLSE, with negligible performance degradation. Experimental validation in a 40 GHz bandwidth-constrained FTN-16QAM system, operating at 80/90/100 Gbaud over 10-40 km transmission, underscores its robustness and efficacy. These results position the proposed MEPS-MLSE as an effective solution for power-efficient data centers, poised to meet the escalating scalability and energy demands of next-generation optical interconnects.