Access Interval Prediction For Tightly Coupled Memory Systems

Today, processing elements of embedded systems usually share some amount of the available on-chip memory in order to maximize the area utilization. However, sharing memory incurs runtime conflicts, which entail performance penalties. To ease this restriction, we introduce a method for memory Access Interval Prediction. It minimizes conflicts by predicting the interval between two consecutive memory accesses. In contrast to contemporary work, we do not rely on compile time information or other a priori knowledge. Standard benchmarks show that we can predict over 80% of all memory access intervals correctly, thereby significantly reducing the number of access conflicts.