SHF: Large: General-Purpose Approximate Computing Across the System Stack

点击翻译按钮获取中文摘要

Energy efficiency is a fundamental challenge facing the IT industry. Benefits go beyond reduced power demands in data centers and longer battery life in mobile devices. It is a fundamental enabler of future systems as we approach the limits of silicon device scaling. Therefore, providing a novel and holistic approach to energy efficiency in computer systems can have a transformative effect on IT and society. Many important applications---e.g., computer vision, novel user interfaces, signal processing, web search, augmented reality, and big-data analytics---can inherently tolerate some forms of inaccurate computation at various levels. With approximate computing, this fact can be exploited for fundamentally more efficient computing systems. This is a direct analog to Daniel Kahneman's model of how our brains work: they do cheap and quick reasoning (using System 1) in an approximate way, and when required, they do more expensive (and tiring) detailed thinking (using System 2). This research project will develop a analogous model for computer systems, from hardware to programming tools. Taking advantage of approximate computing requires significant innovation: programming models, tools for testing and debugging, and system support with quality guarantees. This project will develop a comprehensive solution across the system stack, from programming language to hardware. To demonstrate the potentials, prototypes of compelling applications amenable to approximate computing (e.g., computer vision) will be created. The project involves work on systems, programming languages, formal methods, and architecture, matching the inter-disciplinary expertise of the PI team. In addition to research papers, the project scope also includes releasing tools, benchmarks, and general infrastructure to the academic and industrial communities. The PIs have a history of inclusion of minorities and undergraduate students in their research efforts.