A Novel Quality Trade-offs Method For Approximate Acceleration By Iterative Training

Approximate computing is a promising design approach to achieve high energy-efficiency,by leveraging the inherent capability of error tolerance in the application level.Quality control is the key in approximate computing so that we can save the energy while keep the outcome satisfying the user's requirement.For better tradeoff of the energy-saving and the quality,previous works propose a hybrid architecture composed of a well trained classifier(error predictor)and a high energy-efficient accelerator.However,this hybrid architecture,relying on one-pass training process,has not been fully explored.Therefore,we propose a novel optimization framework,which advocates an iteratively training process to coordinate the training of the classifier and the accelerator with a judicious selection of training data.A dynamic threshold tuning algorithm is integrated into the iterative training process to maximize the invocation of the accelerator while satisfying the quality requirement.At last,we propose an efficient algorithm to explore the topologies of the accelerator and the classifier comprehensively.Experimental results shows significant improvement on the quality and the energy-efficiency compared to the conventional one-pass training method.