Separated by birth: Hidden differences between seemingly-identical smartphone CPUs

Smartphone consumers, app developers, and even mobile systems researchers operate under the assumption that performance differences between identical smartphones should be small. Consumers pick a model to purchase and don't consider that the specific device they leave the store with may vary quite dramatically from the identical models it sat next to on the shelf. App rating systems typically collect the model from reviewers, but not more detailed information-again, assuming that all instances of a particular model perform similarly. Even mobile systems researchers will conduct studies using small numbers of devices that fail to account or control for inherent differences between identical phones. Unfortunately seemingly-identical smartphones can in fact have very different performance characteristics. Note that we are not referring to differences in battery or Flash performance caused over time by wear. Inherent differences would separate two brand-new phones still in the original packaging. Our experiments show up to 20% performance and energy consumption differences between otherwise identical devices. These differences result from process variation in the manufacture of smartphone CPUs, which causes some CPUs to perform much more poorly than others. This paper explains the causes of this variation, measures its impacts, and discusses implications for smartphone researchers, software developers, and consumers.