C-AFA: A Conditionally Approximate Full Adder for Efficient DNN Inference in CIM Arrays

Digital compute-in-memory (DCIM) architectures are becoming crucial for real-time and accurate deep neural network (DNN) inference due to their capacity for precise computations. However, traditional DCIM systems often struggle to balance precise data processing with computational efficiency. In scenarios where exact computations are not always necessary, especially in DNNs featuring sparse data, existing multi-bit operations frequently fail to achieve an ideal balance between accuracy and efficiency. In this paper, we introduce a novel approach to DCIM architecture that selectively skips computations by using probabilistically determined values, effectively reducing the computational load. Our simulation results demonstrate that this method significantly reduces computing cycles, achieving a speedup of up to 1.5× compared to traditional methods, while maintaining accuracy with only a minimal decrease of 1.5%.