Modeling the Impact on Performance of Memory Pooling in Heterogeneous MPSoCs

Multiprocessor systems-on-chip with distributed memories and task processing are promising architectures to tackle processing demands of edge-cloud applications for autonomous vehicles. We present a novel model which allows estimation of the speedup when memory pooling is combined with prefetching. Processing time and data transfer time are both taken into account. In our scenario, memory pooling enables utilization of remote memories and prefetching hides the additional latency. The model shows speedups of up to 200% for data-intensive processing scenarios. Our approach shows that reasonable performance gains can be achieved when increasing flexibility of the memory architecture.