Statistical memory management for digital signal processing

It is pointed out that, with the increasing gap between memory access time and processor cycle time, memory bandwidth has become a serious bottleneck in a growing number of digital signal processing applications. The authors present forward memory interleaving with the help of `smart allocation´ as an architecture level solution to this problem. They define an allocation algorithm, and test for performance using simulations for a large-vocabulary real-time speech recognition system. Within certain bounds, this allocation scheme has proven to be successful. It is emphasized that this sort of allocation is more successful as the memory access becomes more predictable, but not necessarily deterministic. This is the case for many digital signal processing applications, and therefore significant memory bandwidth improvement can be achieved with no extra hardware resources (as compared to generic algorithms)