X-Centric: A Survey on Compute-, Memory- And Application-Centric Computer Architectures

Big Data and machine learning constitute the multifaceted challenge of computer engineering in the past decade. The meaningful processing of vast amounts of unstructured data from a myriad of sensors and devices is a complicated endeavor already. Aggravated by the need to enter the extremely power- and resource-constrained pocket-size mobile domain, the computing as we know it is rapidly evolving. Data-centric in- and near-memory computing, as well as highly heterogeneous accelerator-equipped application-centric architectures, are on the rise to tackle the unsatisfiable demand for evermore compute performance and efficiency. To learn from these innovations, this paper surveys compute-, memory-, and application-centric architectures and related programming paradigms and analyzes prominent chances and challenges. The key insights from the particular domains are: 1) The high nominal processing performance of compute-centric systems is thwarted by massively decreasing data-to-task locality and increased data movement. Nevertheless, the commodity of shared-memory programming and the presence of widespread legacy applications keep this domain alive. 2) Memory-centric designs help to mitigate the data locality wall and significantly improve power and performance efficiency. However, a memory-centric programming paradigm is still missing. 3) Heterogeneity, customization, and established ecosystems (like for mobile devices) enable application-centric optimization under often tight thermal, power, and resource constraints. However, a holistic SoC-level design approach is required to utilize and program the diversity of processing units in different application domains efficiently. A one-size-fits-all architecture approach seems not in sight because of the wide diversity in domain-specific requirements and constraints. Therefore, established ecosystems, 3D-stacked logic-enhanced memory devices, and commoditized architecture-aware programming models seem fundamental for performant and programmable future-proof computer architectures.