Exploiting Short-Lived Values for Low-Overhead Transient Fault Recovery

CMOS downscaling trends, manifested in the use of smaller transistor feature sizes and lower supply voltages, make microprocessors more and more vulnerable to transient errors with each new technology generation. One architectural approach to detecting such errors and recovering from them is to execute two copies of the same program and then compare the results. In this paper, we propose a technique to dramatically reduce the performance and energy overhead of result verification by exploiting the observation that a large percentage of generated result values in a datapath are short-lived. Our scheme avoids verification of short-lived values without impeding transient fault detection and recovery capabilities.