Hardware-Aware Mathematical and Numerical Evaluation of LIF, Izhikevich, and Digital Spiking Neuron Models for Low-Cost Neuromorphic Implementation

Low-cost neuromorphic computing requires neuron models that combine relevant spiking dynamics with numerical stability and technological feasibility. This work compares three representative spiking-neuron families: Leaky Integrate-and-Fire (LIF), Izhikevich, and a simplified digital ShiftLIF-type approximation. A unified, reproducible, and hardware-oriented framework was developed using a synthetic stimulus bank, homogeneous MATLAB simulations, and controlled variations in temporal discretization, noise, quantization, and parametric perturbations. The evaluation combined dynamic, numerical, and hardware-aware metrics with multicriteria aggregation across application scenarios. Izhikevich achieved the lowest mean first-spike error (0.809 ms) and the highest noise robustness (0.981), whereas LIF attained the lowest normalized-state RMSE (0.017) and the lowest mean computation time (1.875 ms). Although ShiftLIF did not lead the local temporal-accuracy metrics, it showed the lowest discretization sensitivity (0.177), parametric sensitivity (0.0195), quantization sensitivity (0.3084), and arithmetic cost per iteration (4 operations). It also obtained the highest score in the balanced multicriteria scenario (0.691) and ranked first in all four evaluated scenarios. By contrast, under 8-bit quantization, LIF completely lost spiking activity in 70.6% of the runs. Overall, no model was universally superior; however, within the adopted framework, ShiftLIF provided the most favorable trade-off for low-cost digital neuromorphic implementations.