CSR: Small: Accelerating Microprocessor Post-Silicon Diagnosis with Statistical Inference

Regents Of The University Of Michigan - Ann Arbor, Ann Arbor MI

Investigators

Abstract

Finding bugs in modern complex microprocessors is a critical and daunting task that must be deftly mastered to move designs from first silicon to launch. Escaped bugs may lead to the demise of a digital silicon company. At the same time, looking for design errors on first silicon prototypes lacks the observability, controllability and repeatability afforded by pre-silicon simulation frameworks. To complicate the process further, many of the bugs that manifest in real silicon are the result of complex asynchronous interactions and/or electrical anomalies that are often not easily or frequently repeatable. Because of these challenges, debugging these "fleeting bugs" in early silicon is a black art that can significantly impact design schedules if the debugging process does not proceed smoothly. This project investigates solutions to support the efficient diagnosis of these most challenging post-silicon validation bugs, those that manifest only occasionally. These bugs may be functional, timing or electrical errors, or also missed manufacturing defects. The approach entails placing lightweight instrumentation on-chip to collect data during a prototype's test executions. The data is then analyzed offline using statistical inference algorithms to quickly point verification engineers to offending components. The research explores a range of ideas to find the most promising instrumentation and algorithms for analysis. The results of this research effort allow semiconductor companies to shorten their time to market while delivering high quality products, with low incidence of escaped bugs; in turn, the work benefits society in that it unlocks further scaling and growth for the electronics industry.