The Case for Lifetime Reliability-Aware Microprocessors

TLDR

Ensuring long processor lifetimes by limiting wear‑out hard errors is critical, yet continuous scaling and higher temperatures make reliability targets harder to meet, and current worst‑case qualification methods are overly conservative. The paper argues that persisting with worst‑case reliability methods unnecessarily constrains performance and proposes lifetime reliability awareness at the microarchitectural design stage. The authors introduce RAMP, an architecture‑level model that dynamically tracks lifetime reliability, and DRM, a technique that lets processors adapt to changing application behavior to maintain reliability targets. DRM allows processors to be qualified at lower, more realistic operating points, and experiments with RAMP show cost savings or performance improvements, with dynamic voltage scaling as an effective response and dynamic thermal management remaining distinct from DRM.

Abstract

Ensuring long processor lifetimes by limiting failuresdue to wear-out related hard errors is a critical requirementfor all microprocessor manufacturers. We observethat continuous device scaling and increasing temperaturesare making lifetime reliability targets even harder to meet.However, current methodologies for qualifying lifetime reliabilityare overly conservative since they assume worst-caseoperating conditions. This paper makes the case thatthe continued use of such methodologies will significantlyand unnecessarily constrain performance. Instead, lifetimereliability awareness at the microarchitectural design stagecan mitigate this problem, by designing processors that dynamicallyadapt in response to the observed usage to meeta reliability target.We make two specific contributions. First, we describean architecture-level model and its implementation, calledRAMP, that can dynamically track lifetime reliability, respondingto changes in application behavior. RAMP isbased on state-of-the-art device models for different wear-outmechanisms. Second, we propose dynamic reliabilitymanagement (DRM) - a technique where the processorcan respond to changing application behavior to maintainits lifetime reliability target. In contrast to currentworst-case behavior based reliability qualification methodologies,DRM allows processors to be qualified for reliabilityat lower (but more likely) operating points than theworst case. Using RAMP, we show that this can save costand/or improve performance, that dynamic voltage scalingis an effective response technique for DRM, and that dynamicthermal management neither subsumes nor is sub-sumedby DRM.