A study on the energy-precision tradeoffs on commercially available processors and SoCs with an EPI based energy model

Abstract

Energy-efficiency is a critical concern for many computing systems. With Moore's law showing its limits, new hardware design and programming techniques emerge to pursue energy scaling. Among these, approximate computing is certainly the most popular in current works. It has been shown that reducing precision using software techniques can show significant energy savings on commercially available processors. In this paper, an energy model based on Energy Per Instruction (EPI) has been built in order to understand which mechanisms enable those savings. EPIs of various instructions have been measured and data movement has been identified as being the major consumer. The energy model has been evaluated on a computationally intensive Newton method for solving nonlinear equations using double-precision and single-precision floating-point data types. For all the cases, the model predicts the consumption with less than 10 % error. The energy breakdown reveals that arithmetic operations consume less than 6 % of the total energy and that savings are mainly achieved by reducing the amount of data transferred between registers, cache and main memory. With these conclusions, implementing approximate arithmetic circuits in this type of architecture would have a very limited impact on the consumption. It is however shown that specialized hardware implemented on an FPGA interconnected with a processing system can lead to an additional 20 % energy savings on the Newton method using the same single-precision data type.