Manufacturing semiconductors is hard, making them work right is even harder, and making them work right the first time you turn them on is damn near impossible.
This difficult situation is caused by two strongly competing trends: the astounding and ever-increasing complexity of modern chips, and the ever-greater time-to-market demands of the modern technology product cycle. You can take some time and make a highly complex product, or you can make a simple product quickly, but it is extremely hard to make a highly-complex product very quickly. Unfortunately, this is exactly the dilemma faced by today's semiconductor companies.
We can't do anything about these market trends, and profitable products will always be complex and hard to build. But, we can do something about the daunting task of figuring out what is wrong with a many-million-transistor circuit that simply refuses to operate properly. This is where our work on fault diagnosis and silicon debug comes in.
We are working on ways of quickly but accurately determining the root cause of failure on a defective chip, even when many independent errors contribute to a single bad result. We have developed algorithms that analyze the way a circuit fails, and then use information about the circuit itself to infer the most likely cause of failure. We consider many types of circuit failure, from logic errors to intermittent or timing failures, and are targeting a range of defect mechanisms from common fabrication errors to complex and exotic defect scenarios. We are building tools that will be compatible with industry-standard data formats and toolflows so that our work can have immediate industrial benefit.
The goal of our research is to develop sophisticated but efficient algorithms that can deal with the enormous variety of ways that a modern chip can fail. Good diagnosis and debug tools are necessary if the semiconductor industry is to pursue its (nearly) impossible task of producing of higher-quality products at less expense and an ever-shorter time-to-market.
