Resource Aware power modelling for compute-bound tasks on GPUs

Scientific computing workloads require significant amount of energy when executed on Graphical Processing Units (GPUs). To make the execution of such workloads energy-optimal, power models that can accurately estimate power, hence total energy, are required. In this paper, we propose a GPU power model as a linear function of core frequency, number of active GPU lanes, and number of Thread Blocks (TBs) for compute-bound kernels; these are kernels that require minimal memory access and CPU-GPU transfers. We implement and validate our model using kernels from different domains. Since we use common architectural components such as core frequency, lanes, and a general programming paradigm such as TBs, we build a model on a Pascal GPU which can be easily extended to future GPU generations. We find that the constant power term in our model is the primary component. We achieve a Mean Absolute Percentage Error (MAPE) of 5% and 8% on two different kernels from the CUDA SDK samples and Rodinia benchmark, during validation on a Pascal GPU. We also obtain a MAPE of 14.8% from our evaluation of the GICOV kernel from the Rodinia benchmark on an Ampere GPU; we find these initial results encouraging, and look forward to extend our power model further over a wider range of kernels and architectures.