Statistical analysis and modeling for error composition in approximate computation circuits

摘要: Aggressive requirements for low power and high performance in VLSI designs have led to increased interest approximate computation. Approximate hardware modules can achieve improved energy efficiency compared accurate modules. While a number of previous works proposed arithmetic, these focus on solitary arithmetic operations. To utilize the benefit modules, CAD tools should be able quickly accurately estimate output quality composed designs. A work [10] proposes an interval-based approach evaluating certain However, their uses sampled error distributions store characterization data hardware, its accuracy is limited by intervals used during characterization. In this work, we propose estimation that based lookup table technique characterizes statistical properties hardwares regression-based composing statistics formulate quality. These two techniques improve speed several metrics over set multiply-accumulator testcases. Compared modeling [10], our estimating 3.75× more comparable runtime testcases achieves 8.4× reduction composition flow. We also demonstrate applicable general