The main objective of computer chip testing has and continues to be the separation of good chips from bad ones (i.e., ones that do not meet the desired operational characteristics). Test is now however being expanded as a value-added endeavor. In this project, we are data-mining test data in order to continuously monitor chip quality. We propose to use diagnosis-extracted models of chip failures along with a new technique for estimating chip quality. Both are incorporated in an on-line, quality-monitoring methodology that ensures a desired level of quality by changing the actual tests applied to a computer chip to better match the characteristics of currently-failing chips.
This approach to quality is dynamic in nature and is a radical change from the typical approach. Without exception, each chip manufacturer (Intel, IBM, etc.) assumes that any type of defect can occur anywhere within their chip which means that each manufactured instance has to be thoroughly tested at considerable expense. This is akin to prescribing drugs for all possible diseases/ailments for every patient without performing one diagnostic examination. Opposed to the traditional approach, this proposed work instead diagnoses chips that have failed in the past to determine what ?diseases? (i.e., defects) actually are occurring within the fabricated ICs. The ?prescriptions? (i.e., the tests applied to the chips) can therefore be changed and/or minimized to match the diseases found instead of over-testing as is done now, leading to improved chip quality at minimal cost.
