Quality Tunable Approximate Adder for Low Energy Image Processing Applications

In this paper, we present energy-efficient Tunable Approximate Adders (TAAs) based on power-gating. We propose two TAA designs: TAA1 and TAA2. TAAs are runtime configurable and are bit-tunable to support varying degrees of approximation with maximum bounded error. TAA designs have been implemented using UMC 65nm technology. On average, a single-bit TAA1 and TAA2 consume 15% and 47% lesser energy as compared to a single-bit exact mirror adder respectively. TAA1 is relatively lesser erroneous but consumes more energy as compared to TAA2. We also applied TAA1 and TAA2 on image processing applications: image addition, mean filtering and Laplacian filtering on multiple images. We have studied the trade-off between energy consumption and output quality by varying the number of approximate bits for the aforementioned applications. For TAA1 at 4-bit approximation, the average PSNR and average SSIM were 32.79 and 0.875 respectively with an energy saving of 6.67%. For TAA2 at 4-bit approximation, the average PSNR and average SSIM were 28.37 and 0.88 respectively with an energy saving of 28.9%.