PiC-NDM: A 65nm processing-in-CAM-based binary CNN Accelerator for natural disaster management

Binary convolutional neural networks, BCNNs, in which weights and activations are constrained to binary values, offer a promising approach for energy efficient inference in distributed and resource constrained sensing platforms. However, most existing BCNN implementations still rely on full precision computation in critical network components, including the first input processing layer and the final output classification layer. This reliance reduces the benefits of binarization and complicates deployment in extreme edge and mission critical systems.This work presents PiC-NDM, a silicon based processing in memory (PiM) accelerator that implements a binary convolutional neural network in-content addressable memory (CAM). The proposed architecture exploits intrinsic CAM operations to realize binary convolution and classification using XNOR and analog POPCOUNT, eliminating the need for full precision arithmetic or auxiliary digital processing units. By leveraging statistical aggregation effects enabled by approximate matching, PiC-NDM achieves accurate classification without batch normalization, softmax, or multi bit accumulation. PiC-NDM is designed and fabricated in a commercial 65nm CMOS process and evaluated using representative binary classification workloads. Measured silicon results demonstrate accuracy comparable to software baselines, achieving 95.2% accuracy on the MNIST dataset and 93.5% accuracy on a hand gesture dataset. The prototype delivers a throughput of 560K inferences per second and an energy efficiency of 703M inferences per second per watt, demonstrating suitability for real time environmental hazard detection in distributed natural disaster management systems.