Considerations When Evaluating Microprocessor Platforms.

Motivated by recent papers comparing CPU and GPU performance, this paper explores the questions: Why do we compare microprocessors and by what means should we compare them? We distinguish two distinct perspectives from which to make comparisons: application developers and computer architecture researchers. We survey the distinct concerns of these groups, identifying essential information each group expects when interpreting comparisons. We believe the needs of both groups should be addressed separately, as the goals of application developers are quite different from those of computer architects. Reproducibility of results is widely acknowledged as the foundation of scientific investigation. Accordingly, it is imperative that platform comparisons supply enough detail for others to reproduce and contextualize results. As parallel processing continues to increase in importance, and parallel microprocessor architectures continue to proliferate, the importance of conducting and publishing reproducible microprocessor platform comparisons will also increase. We seek to add our voice to the discussion about how these comparisons should be conducted.