GPGPU Footprint Models to Estimate per-Core Power

We explore the problem of how to easily estimate the per-core power distribution of GPGPUs from the total power of all cores. We show that the dynamic energy consumption of a core for a given kernel, represented by its work footprint, is approximately proportional to the total time taken by all work units executing on that core, and the static power, represented by its core footprint, is proportional to the time that the core has assigned work. Footprints can be easily tracked using two hardware counters per GPU core. We also show how per-core power estimates can be used to compute power-performance pareto frontiers that identify opportunities for saving power and energy in cases of non-uniform work distribution by exploiting per-core DVFS support for GPGPUs.