RAAS: Runtime Adaptive Approximation System

Software‐level approximation techniques, such as loop perforation and functional replacement, can improve energy and performance in error‐tolerant applications, but typically require extensive programmer intervention even when compiler support is available. Design‐time and compile‐time approximation approaches also make static decisions, limiting the ability to adapt approximation levels dynamically in response to runtime conditions or inputs. We present RAAS, a Runtime Adaptive Approximation System that identifies and applies approximation opportunities dynamically within a Just‐In‐Time compiler. The programmer marks approximation regions using simple annotations, and provides either built‐in or custom quality metrics along with acceptable error bounds. RAAS explores the approximation space at runtime, evaluating candidate transformations for performance and accuracy, and converges to configurations that maximize speedup while respecting the allowed error threshold. Evaluation on application kernels from domains such as simulation and image processing shows speedups of up to with quality degradation below 10%, outperforming static compilation approaches even when accounting for runtime adaptation overhead.