Detecting VLIW Hard Errors Cost-Effectively through a Software-Based Approach

Abstract

Research indicates that as technology scales, hard errors such as wear-out errors are increasingly becoming a critical challenge for microprocessor design. While hard errors in memory structures can be efficiently detected by error correction code, detecting hard errors for functional units cost-effectively is a challenging problem. In this paper, we propose to exploit the idle cycles of the under-utilized VLIW functional units to run test instructions for detecting wear-out errors without increasing the hardware cost or significantly impacting performance. We also explore the design space of this software-based approach to balance the error detection latency and the performance for VLIW architectures. Our experimental results indicate that such a software-based approach can effectively detect hard errors with minimum impact on performance for VLIW processors, which is particularly useful for reliable embedded applications with cost constraints.