An Accurate Energy Estimation Framework for VLIW Processor Cores

Abstract

In this paper, we present a comprehensive energy estimation framework for software executing on very long instruction word (VLIW) processor cores. The proposed energy model is used as an average energy estimator coupled to the instruction set simulator (ISS) of the processor. The base energy of an execution set is computed as the NOP energy added with incremental energies of each instruction in the execution set. The inter execution-set energy is accurately modeled with a new approach as a linear equation of three factors -functional to functional instruction switching; functional to NOP or prefix instruction switching; and variability in the length of the execution set. This reduces the characterization complexity of the model to 0(N), where N is the total number of instructions in the instruction set of the processor. We have introduced the concept of "functional separability" in the energy model, wherein the energy of each high-level function of the processor core is distinctly mapped to only one component in the model. The model is also capable of handling control codes with branches and predicated execution. The average error magnitude of the framework when applied on the StarCore processor with a large suite of DSP and control benchmarks is 2.5%, whereas the maximum error is less than 6.0%.