A Comparison of Approaches for High-Level Power Estimation of LUT-based DSP Components

Abstract

We compare two approaches for high-level power estimation of DSP components implemented in FPGAs for different sets of data streams from real-world applications. The first model is a power macro-model based on the Hamming distance of input signals. The second model is an analytical high-level power model based on switching activity computation and knowledge about the component’s internal structure, which has been improved to also consider additional information on the signal distribution of two consecutive input vectors. The results show that the accuracy of both models is, in most cases, within 10% of the low-level power estimates given by the tool XPower when cycle-bycycle input signal distributions are taken into account, and that the difference between the model accuracies depends significantly on the nature of the signals. Additionally, the effort required for the characterization and construction of the models for different component structures is discussed in detail.