Total: 1
This paper presents a memristor-based compute-in-memory hardware accelerator for on-chip training and inference, focusing on its accuracy and efficiency against device variations, conductance errors, and input noise. Utilizing realistic SPICE models of commercially available silver-based metal self-directed channel (M-SDC) memristors, the study incorporates inherent device non-idealities into the circuit simulations. The hardware, consisting of 30 memristors and 4 neurons, utilizes three different M-SDC structures with tungsten, chromium, and carbon media to perform binary image classification tasks. An on-chip training algorithm precisely tunes memristor conductance to achieve target weights. Results show that incorporating moderate noise (<15%) during training enhances robustness to device variations and noisy input data, achieving up to 97% accuracy despite conductance variations and input noises. The network tolerates a 10% conductance error without significant accuracy loss. Notably, omitting the initial memristor reset pulse during training considerably reduces training time and energy consumption. The hardware designed with chromium-based memristors exhibits superior performance, achieving a training time of 2.4 seconds and an energy consumption of 18.9 mJ. This research provides insights for developing robust and energy-efficient memristor-based neural networks for on-chip learning in edge applications.