Spiking neural networks (SNNs) promise energy-efficient edge computing, but they often suffer from lower accuracy and training instability on complex tasks like 3D point cloud recognition. In this work, we introduce a novel attention-based framework to enhance spiking point cloud processing while alleviating training instability in SNNs. Our key idea is a Sort-Mask-Convolve (SMC) paradigm that efficiently aggregates local neighborhoods by sorting points along each axis and applying a distance-based binary mask, leading to high processing speed and low memory footprint. We then develop a Point Attention Module (PAM) with local descriptor-guided spatial and channel attention to dynamically reweight features. Moreover, we deploy a Progressive Attention Decay (PAD) strategy that employs strong attention during training but gradually phases it out, yielding a pure spiking model at inference time. Our attention-based framework eliminates the need for explicit priors or approximations required by prior algorithms. Experiments on ModelNet40 show that our method improves Spiking PointNet’s accuracy by about 1.94% points (to 89.64%), slightly surpassing the ANN baseline (89.60%). Compared to conventional k-NN/ball query methods, our SMC achieves over 6 \(\times \) speedup and 90% memory savings.