EAGER: Language and Architecture Design for Approximation at Different Granularities

Georgia Tech Research Corporation, Atlanta GA

Investigators

Abstract

The IT industry's economic ecosystem mostly relies on continuously delivering new services and devices by exploiting continuous performance and efficiency improvements in general-purpose processing. However, as we enter the dark silicon era, the benefits from transistor scaling are diminishing and the current paradigm of processor design significantly falls short of the traditional cadence of performance improvements. These shortcomings can drastically curtail the industry's ability to continuously deliver new capabilities, breaking the backbone of its economic ecosystem. Radical departures from conventional approaches are necessary to provide large efficiency and performance gains for a wide range of applications. One such departure is general-purpose approximate computing that accepts error in computation and relaxes the traditional abstraction of "near-perfect" accuracy. Approximate computing leverages the inherent error tolerance of a large body of emerging applications. These applications span a wide range of domains including vision, big data analytics, machine learning, sensor processing, cyber-physical systems, multimedia, and web search. For these diverse domains of applications, it is timely and crucial to provide architectural mechanisms and programming abstractions that enable trading quality of results for gains in both performance and efficiency. This project aims to provide both architectural mechanisms and programming language constructs that make approximate computing effectively applicable to a wide range of domains of applications. Energy efficiency is the IT industry's biggest challenge. To maintain the nation's economic leadership in the IT industry, it is vital to develop solutions such as ours that provide significant gains in efficiency and performance. Many of these techniques allow approximation to permeate across the boundaries of hardware and software. Thus, it is essential to educate a workforce that not only deeply understands hardware and software but also can innovate across the boundaries of the two. This project provides a foundation for such research and education by producing benchmarks, tools, and general infrastructure for approximate computing. These artifacts will be made publicly available and will be integrated into the curriculum.