Abstract
Object detection, as a fundamental task in computer vision, mainly performs the classification and localization of objects in images or videos. However, traditional edge computing platforms fall short of meeting the demands for state-of-the-art object detection model size and computing power. Here, a 128 Mb phase change memory chip is fabricated with a high memory yield of 99.99999% in a 40 nm node and utilized for efficient in-memory vector-matrix multiplication and in-memory max computation. In particular, in order to mitigate the significant programming energy overheads for large-scale memristor arrays and the reliance on high-precision analog-to-digital-converter (ADC) in compute-in-memory operations, a novel mixed-precision weight mapping strategy is adopted. Compared with traditional schemes, the ADC modules achieve up to a 22.3× reduction in energy consumption while maintaining equivalent network performance. Ultimately, this memristive in-memory object detection system demonstrates 4,180× higher energy efficiency and 228× greater computational throughput compared to GPU implementations.