Mapping the weights of an Artificial Neural Network (ANN) onto the resistance values of analog memristors can significantly enhance the throughput and energy efficiency of artificial intelligence (AI) applications, while also supporting AI deployment on edge devices. However, unlike traditional digital-based processing units, implementing AI computation on analog memristors presents certain challenges. The non-linear resistance switching characteristics and limited numerical bit precision, determined by the number of program levels, can become bottlenecks affecting the accuracy of ANN models. In this study, we introduce a resistance control method, a feedforward pulse scheme that enhances resistance configuration precision and increases the number of programmable levels. Additionally, we propose an evaluation method to explore the impact of setting multi-level resistance states on ANN accuracy. Through demonstrations on a TiO2−x-based memristor, our method achieves 512 states on a device with a high resistance state to a low resistance state ratio of just 1.19. Our approach achieves 95.5% accuracy on ResNet-34 with over 20 million parameters through weight transfer, thereby demonstrating the potential of analog memristors in AI model inference. Furthermore, our findings pave the way for future advancements in increasing resistance states, which will enable more complex AI tasks and enhance the in-memory computational capabilities required for AI edge applications.
