A Quantitative Performance Evaluation of Fast on-Chip Memories of GPUs

Modern Graphics Processing Units (GPUs) have evolved to high performance general purpose processors, forming an alternative to CPUs. However, programming them effectively has proven to be a challenge, not only due to the mandatory requirement of extracting massive fine grained parallelism but also due to its susceptible performance on memory traffic. Apart from regular memory caches, GPUs feature other types of fast memories as well, for instance scratchpads, texture caches, etc. In order to gain more insight to the efficient usage of these memory types some quantitative performance measures could be beneficial. In this paper we describe a set of micro-benchmarks which aim to provide effective bandwidth performance measurements of the on-chip special memories of GPUs. We compare the peak measurements of different memory types and the use of different data type sizes. In addition, we validate the peak measurements on real world problems as provided by the polybench-gpu benchmark suite. We compare the profiling bandwidth of on-chip memories with the peak measurements as captured with the proposed micro-benchmarks. The source code of the micro-benchmark suite is publicly available.