A New Low Power Schema for Stream Processors Front-End with Power-Aware DA-Based FIR Filters by Investigation of Image Transitions Sparsity

Abstract

Data stream processors and accelerators, due to the outstanding energy performance, run on hardware more than any time in modern designs. The general model for these processors comprises massive shift register arrays with the largest share in energy dissipation and processing elements (PE). In this paper, a new gated flip-flop is designed and utilized in shift register arrays, to decrease power consumption. Distributed arithmetic (DA) is an efficient method for calculating the inner product and FIR filters. DA-based FIR filter consists of two parts of shift register and PE array. Due to the significant share of power in shift register, in this paper, DA-based FIR filter is employed to show the improvement of the proposed gated flip-flop. Investigation of statistical properties of input in image processing applications, utilization of implicit clock gating, and multi-vdd techniques are three main approaches we used in this study to increase energy efficiency. It is shown that the transition density (TD) in 50% of static images of target databases is lower than 0.5. A set of random data with different TDs is generated, fed to the gated flip-flop in 180 nm technology, and the results show a 62–2% improvement in dynamic power consumption. Further optimization of 29–21% is achieved when the multi-vdd is applied on the wrapper circuit of the gated flip-flop. Likewise, using the proposed flip-flop in the shift register unit of the DA-based FIR filter has improved the power consumption by 15–40% compared to the conventional flip-flop.