Challenges and opportunities for memristors in modern AI computing paradigms

The emergence of transformer-based artificial intelligence (AI) models has made a great impact on modern AI computing paradigms, where the attention mechanisms in transformer models dynamically generate weights and require computations involving global parameters. These requirements pose unprecedented challenges for memristor performance in in-memory computing, which demands memristor arrays with exceptional endurance, latency, energy consumption, and device uniformity. This perspective focuses on the alignment between AI computational requirements and memristor specifications, highlighting recent breakthroughs in materials, mechanisms, and applications for next-generation in-memory computing. Through this comprehensive analysis, our perspective provides critical insights into advancing memristor-based computing research toward practical AI applications. It also underscores key research priorities and the necessity for interdisciplinary collaboration to propel the future of AI innovation.