Data size reduction for clustering-based binning of ICs using principal component analysis (PCA)

Accurate binning of ICs using analog characteristics such as I_DDQ requires using data from a number of vectors. From this data, information needs to be extracted using a method that will yield sufficiently high resolution. Using a large volume of data can require significant computation time. If n analog measurements are made for each chip, the data has n dimensions. However the measured I_DDQ values for a chip can he highly correlated. We examine an approach based on principal component analysis (PCA) for reducing the data size while preserving almost all of the information. PCA transforms the data by extracting statistically independent components and arranging them in the order of relative significance. Using industrial I_DDQ data we found that often n-dimensional data can be reduced to a single dimension with no substantial change in the clusters identified