Nanocrystal-Si Flash Memory-based Engergy-efficient Multi-bit Compute-in-Memory Design for Edge Neural Networks

With the rapid advancement of artificial intelligence, the computational demands of neural network models have grown significantly. The traditional Von Neumann architecture separates memory and computation units and incurs substantial power overhead. Computing-in-Memory (CIM) has emerged as a promising paradigm to overcome this problem. However, most existing CIM systems still suffer from high memory cell operating currents and substantial power overhead from peripheral circuits. This work presents a novel CIM macro based on nanocrystal silicon (nc-Si) Flash memory. Leveraging the low-power characteristics of nc-Si Flash memory, we propose a multi-bit storage CIM macro in which digital-to-analog conversion for multi-bit inputs is performed within the nc-Si Flash CIM unit itself. Experimental results show that the proposed nc-Si Flash CIM array achieves an energy efficiency of 193.1TOPS/W. The classification accuracies of the LeNet-5 network on the MNIST and Letters datasets are 98.5% and 90.77%, respectively. The quantized VGG-9 achieves a classification accuracy of 76.2% on the CIFAR-10 dataset.