| Title |
Implementation of Hardware Data Prefetching Technique Using Machine Learning Technique |
| Authors |
송경환(KyungHwan Song) ; 김강희(KangHee Kim) ; 최상방(SangBang Choi) |
| DOI |
https://doi.org/10.5573/ieie.2019.56.3.11 |
| Keywords |
hardware ; prefetch ; cache ; RNN ; GRU |
| Abstract |
The prefetch is a way to hide the memory latency between the processor and the memory and to resolve the performance gap, predicting and prepopulating the processor's frequently used or future data. As the processor's memory access pattern becomes longer and more complicated, the hardware structure becomes complicated and requires a lot of storage space. The LSTM (Long Short-Term Memory) prefetch technique, which is a prefetch method based on machine learning, performs learning and prediction using LSTM machine learning algorithm. However, since LSTM requires a large number of parameters for prediction, it requires a large amount of parameter storage space. In this paper, we propose a prefetch method that predicts the next access address by using GRU (Gate Recurrent Unit) which is a transformation algorithm of RNN as a predictor of prefetch and learning the memory access pattern of workload in GRU. The proposed method has excellent prediction performance. Reducing the number of parameters reduces the storage space for reading and writing parameters in the prefetcher design. This greatly reduces die area, resulting in high energy consumption efficiency. It also greatly reduces the computation time for the prediction and the time to generate the prefetch address. |