October 26, 2010 - Andrew Chien

Title: 10x10: A New Paradigm for Computer Architecture (Meeting the Challenges of the New Technology Scaling Landscape)

Abstract:
Two decades transistor scaling in density, speed, and energy (a.k.a. Moore’s Law)  have enabled microprocessor architects to deliver 1000-fold performance improvement.  This dramatic improvement has enabled computing as we know it today – tiny, powerful, inexpensive, and therefore ubiquitous.  Current projections suggest future scaling in density, but only decreasing improvements in transistor speed and energy.  In this era of energy-constrained performance, the computing industry is engaged in a rapid, broad-based shift to increasing parallelism (multicore) from the largest data centers to small mobile devices.

In the new technology scaling landscape, more narrowly specialized designs (heterogeneity) are increasingly attractive, but computer architects have lacked a paradigm to deal with it systematically.  We believe it is time to move beyond the general purpose architecture paradigm and 90/10 optimization which has served us well for 25 years, and replace it with a new paradigm, “10x10”, which divides workloads into clusters, enabling systematic exploitation of specialization in the architecture, implementation, and software.  We call this new paradigm “10x10” because it divides the workloads and optimizes for 10 different 10% cases, not a monolithic 90/10.  The 10x10 approach can enable 10-fold or more improvement in energy efficiency and performance compared to conventional general-purpose approaches.  In addition, 10x10 has the potential to bring discipline to increasing heterogeneity in computing systems.  We will also outline a few critical challenges for future computing systems in the new technology scaling landscape, including software and applications.