Enhancing Image Segmentation Performance with MRAM based Processing-in-Memory Architecture

Image segmentation is essential for several computer vision applications, including object detection, autonomous driving, and medical imaging. Traditional image segmentation algorithms suffer due to computational inefficiency and memory bandwidth problems, limiting their real-time performance and accuracy. This work focuses on addressing these issues by enhancing image segmentation performance by leveraging a Processing-in-Memory (PIM) architecture based on Magnetic Random-Access Memory (MRAM) technology. The proposed approach optimizes data flow and memory access by integrating processing units into the MRAM memory array, resulting in better efficiency and real-time image segmentation capabilities. The proposed work introduces an MRAM-based PIM architecture to accelerate matrix multiplication operations in segmentation architectures. The study aims to assess and compare the effectiveness of the MRAMs for PIM based architecture for implementing segmentation models by analysing their performance. The computationally intensive segmentation models such as UNet demonstrate a competitive edge in generating images with high correlation to ground truth images. PIM architecture with Voltage Gated Spin-Orbit Torque (VGSOT) has demonstrated substantial reduction in terms of power, and latency by 99.97%, 35.34% respectively as compared to architectures using Phase Change Memory (PCM) crossbars in implementing segmentation inference operations.