Memory Access Prediction

Memory access prediction (MAP) aims to anticipate future memory requests to improve data prefetching and reduce memory latency in computer systems. Current research focuses on applying deep learning models, including transformers, attention-based networks, and multilayer perceptrons, to enhance prediction accuracy. A key challenge is balancing the accuracy of these complex models with their computational cost and storage requirements, leading to efforts in model compression techniques like knowledge distillation and tabularization to create more efficient and practical prefetchers. Improved MAP techniques have the potential to significantly boost the performance of various applications by mitigating a major bottleneck in modern computing.