Pappireddy, Madhukarreddy (2016-08). PRITEXT: Processor Reliability Improvement Through Exercise Technique. Master's Thesis.
With continuous improvements in CMOS technology, transistor sizes are shrinking aggressively every year. Unfortunately, such deep submicron process technologies are severely degraded by several wearout mechanisms which lead to prolonged operational stress and failure. Negative Bias Temperature Instability (NBTI) is a prominent failure mechanism which degrades the reliability of current semiconductor devices. Improving reliability of processors is necessary for ensuring long operational lifetime which obviates the necessity of mitigating the physical wearout mechanisms. NBTI severely degrades the performance of PMOS transistors in a circuit, when negatively biased, by increasing the threshold voltage leading to critical timing failures over operational lifetime. A lack of activity among the PMOS transistors for long duration leads to a steady increase in threshold voltage Vth. Interestingly, NBTI stress can be recovered by removing the negative bias using appropriate input vectors. Exercising the dormant critical components in the Processor has been proved to reduce the NBTI stress. We use a novel methodology to generate a minimal set of deterministic input vectors which we show to be effective in reducing the NBTI wearout in a superscalar processor core. We then propose and evaluate a new technique PRITEXT, which uses these input vectors in exercise mode to effectively reduce the NBTI stress and improve the operational lifetime of superscalar processors. PRITEXT, which uses Input Vector Control, leads to a 4.5x lifetime improvement of superscalar processor on average with a maximum lifetime improvement of 12.7x.