Teaching Chips To Recognize And Exploit Data Patterns For Maximum Efficiency: A Comprehensive Framework For Sparse Matrix Processing In Pulsating Array Architectures

Abstract

In today’s data obsessed world, we are working against an interesting paradox where we are seeing datasets grow exponentially while the numbers we are processing are largely zeros. Sparsity — the observation that many things are sparse, meaning they're mostly zeros — shows up everywhere from social network connections to scientific simulations, but our computing systems wastefully process these zeros as if they were meaningful data. In this paper, we introduce a new way to teach computer chips to automatically detect and take advantage of these patterns of sparsity, significantly boosting performance and energy efficiency. A complete framework is presented which incorporates intelligent data compression, adaptive task scheduling and specialized hardware design in pulsating array processors. We also present novel compression techniques that reduce memory usage up to 85% with no performance degradation and dynamic scheduling that schedules processing units busy over 90% of the time. Our approach is built on adaptive control systems that adapt to different data patterns and produce impressive 3.2× speedups and 67% energy reduction. We show that this framework could revolutionize energy efficient computing, for the next generation of data-intensive applications, through extensive testing across a wide range of applications from scientific computing to machine learning.