CIRCA-GPUs: Increasing instruction reuse through inexact computing in GP-GPUs

"Inexact computing" provides an opportunity for exploiting application characteristics to improve energy efficiency. This is particularly true of error-tolerant image processing applications that are often implemented using large scale parallel processing resources in general-purpose graphics processing units or GP-GPUs. GP-GPUs provide fine-grained parallel architecture by integrating a large number of relatively small and tightlycoupled processing cores. We propose CIRCA-GPUs, a method to jointly exploit fine-grained parallelism and inexact computing in GP-GPUs by reusing results of instructions across different cores and over time. This spatiotemporal instruction reuse with systematic approximation of computing results increases the rate of reuse to improve energy efficiency.