Noise reduction (NR) and hearing-loss compensation (HLC) are essential components of modern hearing-aid (HA) systems. While deep learning has significantly improved the efficacy of these modules, optimizing perceptual performance, model complexity, and algorithmic latency for real-time edge deployment remains a primary challenge. This study systematically explores the performance-complexity trade-offs across four processing paradigms: (1) a biophysically-inspired HA model; (2) a generic NR model; (3) a sequential Cascaded NR+HA pipeline; and (4) Individualized NR (INR), a unified framework that jointly performs NR and HLC. To bridge the gap between high-fidelity processing and strict real-time constraints, a teacher-student knowledge distillation strategy is employed to train compact, low-latency student models that approximate the behavior of high-capacity, non-causal teacher models. Experimental results demonstrate that the INR paradigm achieves superior HASPI and HASQI scores at a complexity comparable to standalone models, outperforming the cascaded approach while requiring only half the computational cost. Furthermore, knowledge distillation achieves up to 13.7x model compression, though the resulting performance is highly paradigm-dependent: HA models are largely robust to compression, whereas INR models exhibit the most significant performance degradation. These findings provide a data-driven foundation for selecting processing paradigms across diverse computational budgets, facilitating the deployment of advanced, biophysically-inspired processing on resource-constrained devices.