Resilience Proportionality—A Paradigm for Efficient and Reliable System Design

Abstract

Reliability, Availability, and Serviceability (RAS) are key considerations in hardware design, be it for mobile devices or high-end servers. However, provisioning RAS is often at odds with meeting performance and energy targets and increases the overall cost of design of the chip. As a result of this tension, chip design companies have to make difficult decisions about how much RAS they can incorporate into each product in their portfolio and even what customers and market segments they can realistically target. On the other hand, highly scaled silicon technology nodes are susceptible to a variety of reliability problems and emerging technologies such as die-stacking and non-volatile memory, while critical for meeting the demands of future computing needs, have significant reliability challenges of their own. RAS features can actually serve to reduce the deployment costs of these technologies (e.g., by increasing effective yield). Determining the tradeoff between design cost, deployment cost, and the RAS needs of a market is the critical issue to address when evaluating RAS features. In this article, we shed light on this struggle between driving greater efficiency, lowering costs, and meeting the RAS demands of various market segments from an industry perspective. We argue that ending this struggle requires having sufficient flexibility in the design to adapt to the needs of a wide range of applications and hardware configurations. We call such an approach “resilience proportionality” and believe that this approach should guide future architectural reliability research. Finally, we discuss how resilience proportionality can be achieved and certain challenges that need to be addressed.