Analysis of Hardware Prefetchers Suitable for CNN Applications

Employing CPUs for convolutional neural networks (CNNs) is of interest to many users due to its availability. However, to make CPUs more competitive, it is necessary to bridge the performance gap between CPUs and other accelerators such as GPUs and TPUs. Date prefetching is the promising optimization technique because CNN operations tend to have predictable data access patterns. This paper examines and analyzes the usefulness of various hardware prefetchers in CNN applications.