Dynamically Adaptive Approximate Computing Framework
In recent years, smart devices such as head mount displays and smart speakers enriches people’s lives. Programs running on these devices have the property that the quality required changes from moment to moment depending on users’ cognitive ability and inputs to the program. We think that by quickly following the required quality and performing approximate computes that satisfy the maximum acceptable error, we can process these programs with high efficiency. Therefore, we investigate processors and frameworks that can dynamically control the aggressiveness of approximation. With dynamic control of the aggressiveness of the approximation, we aim to achieve even higher efficiency, whereas conventional methods statically define the approximation accuracy.