A key issue to reduce the number of deadline misses and improve energy savings in embedded real-time systems is to accurately estimate the execution time of tasks as a function of the processor frequency. Existing execution time models however, use to rely on off-line analysis or on the assumption that the memory access time (quantified in processor cycles) is constant, ignoring that memory system components are not affected by the processor clock. In this paper, we propose the Processor-Memory (Proc-Mem) model, which dynamically predicts the execution time of the applications running in a multicore processor when varying the processor frequency. Proc-Mem approach is compared with a typical Constant Memory Access Time model, namely CMAT. Results show that the deviation of Proc-Mem is always lower than 6% with respect to the measured execution time, while the deviation of the CMAT model always exceeds 30%. These results turn in important energy savings for a similar number of deadline misses. Energy savings are on average by 22.9%, and up to 47.8% in the studied mixes.