A Quantitative Evaluation and Proposition of Cache Policies in Mobile GPUs

Abstract: Caches today are an important part of mitigating the performance gap between processors and memory which has increased during the last 50 years. Cache policies are vital for the performance of caches and their mission is to select the data which should be contained in the cache and are therefore important for the performance of processors. Graphics processors are very memory bandwidth intensive and rely heavily on caches to increase performance. Although much research of cache policies has been done in the context of general processors, there exists little research in the context of graphics processors. In this thesis we perform a quantitative evaluation of a wide range of cache placement and replacement policies. Based on our evaluation we propose some new policies and we also show that the common least-recently used cache policy performs very well in the context of graphics processors and the work- loads they are subjected to.