Approximate computing has evolved into an effective method for improving digital applications’ performance and energy utilization by tolerating a slight loss in precision. By exploiting specialized hardware accelerators built with approximation circuits, the power vs. performance trade-off may be resolved across various high-computational applications. Among these are image processing, machine learning (ML) and DSP. The goal is to analyze a suitable configuration for an approximation set-up that minimizes resources while respecting the error tolerances of the systems because these applications are frequently used in systems that demand a high level of fault tolerance, such as self-driving vehicles. This study reviews current researches on approx. computing and hardware accelerators for various applications, and evaluates the claimed performance gains of many different approaches. At last, the most pressing problems and prospective avenues for utilizing it in the field of deep learning have also been outlined.