MEMPower: Data-Aware GPU Memory Power Model
This paper presents the MEMPower power model. MEMPower is a detailed empirical power model for GPU memory access. It models the data dependent energy consumption as well as individual core specific differences. We explain how the model was calibrated using special micro benchmarks as well as a high-resolution power measurement testbed. A novel technique to identify the number of memory channels and the memory channel of a specific address is presented. Our results show significant differences in the access energy of specific GPU cores, while the access energy of the different memory channels from the same GPU cores is almost identical. MEMPower is able to model these differences and provide good predictions of the access energy for specific memory accesses.
Published in: Architecture of Computing Systems – ARCS 2019 : 32nd International Conference, Copenhagen, Denmark, May 20–23, 2019, Proceedings, 10.1007/978-3-030-18656-2_15, Springer