Data Science Day @ Columbia University

The quest for energy efficiency affects all types of computer systems, from data centers to embedded devices. As the amount of sensor-generated data continues to grow, their processing by analytics applications is increasingly pushed to the  edge of the cloud. The demand for efficient embedded big-data processing translates in the emergence of heterogeneous architectures, which combine accelerators with general-purpose processors. Heterogeneity, however, exacerbates design and programming complexity. We propose a system-level design methodology that is scalable and application-driven. It relies on a flexible architecture balancing heterogeneity and regularity to ease the integration of accelerators and processors. It promotes intellectual property reuse and enables quick design-space exploration at both the component and application levels. Furthermore, it addresses the challenges of running legacy software and managing energy in accelerator-enhanced architectures. We demonstrate our methodology with a computer-vision application integrated with a real-time monitoring mechanism implemented on a high-end FPGA prototyping system.