Using Computational Biology Methods to Improve Post-silicon Microprocessor Testing

Abstract

Hardware testing is an expensive process at different stages of hardware design and manufacturing. It includes pre-silicon, post-silicon and production testing. Testing is expensive both in terms of manpower and in computing resources, and it directly affects the hardware profitability and the time to market. This problem is especially acute for Systems on Chip (SoC) where both manpower and timing constraints are very tight. Therefore it is important to reduce the total number of tests without sacrificing testing quality.

To learn the behavior of a large test set smart algorithms are needed. In addition, visualization techniques can provide a bird’s-eye view of the total test coverage data.

Our goal is to optimize post-silicon hardware test suites based on coverage metrics and to provide test coverage visualization. We utilize ideas and methods developed in machine learning and bioinformatics, and develop new biology-inspired methods to analyze and visualize post-silicon data. In a different effort, we are exploring combinatorial methods of covering and domination for the same problem.

Mathematically, the results of post-silicon tests can be presented as a matrix whose rows correspond to the tests performed on the chip and columns correspond to certain events of interest occurring during the test’s runs. The matrix values are the number of times the event occurred in the test. Such a matrix can then be used to define a similarity measure between tests and analyze their relations.

In computational biology (bioinformatics), advanced methods were developed to handle gene expression microarray data [1], which has a similar structure. The result of a set of microarray experiments is a gene expression values matrix where rows are genes and columns are conditions. A rich spectrum of methods was developed for analysis of such data [2], and we adapt them for the post-silicon analysis. For example, clustering techniques divide the tests into similarity groups, identifying subsets of tests that cover similar events. The identified groups can then be analyzed by the hardware validation engineers in order to identify coverage holes and to improve the test suite quality. In addition, similar test groups can be investigated for enrichment of certain chip properties as done for gene groups with biological properties. This can give further insight on the chip’s operation and the tests scope. Gene expression software tools that combine advanced analysis and visualization can assist in visual comprehension of the post silicon validation process. We are using for this task the Expander tool developed in Prof. Shamir’s group [3-5].

We describe initial results obtained by applying computational biology methods to post-Si test suite optimization and visualization. Though we experimented only with post-silicon test data, most of the developed methods should be applicable with appropriate modifications also to pre-silicon, production, and even to software testing.