With the ever-growing demands of big data and AI applications, data transfers between the CPU and memory in traditional computer architecture lead to substantial energy dissipation. This issue, known as the von Neumann bottleneck, can be greatly alleviated by Processing-in-Memory (PIM) technology. Resistive Random Access Memory (ReRAM), a non-volatile device that can be organized as a crossbar structure, inherently supports matrix-vector operations required by AI applications. In this talk, I will introduce the ReRAM-based Processing-in-Memory architecture for AI acceleration, showcasing how it significantly improves data processing efficiency, paving the way for faster and more energy-efficient AI accelerators.