Data Learning Techniques for Functional/System Fmax Prediction

Abstract

In this talk, we will present a data learning methodology for building a Fmax predictor based on structural test measurements. Given Fmax and structural test measurements on a set of sample chips, we will show that correlation between the two frequency variations can be greatly improved if “noisy” samples are removed. We develop a method to identify such noisy samples. We explain the data learning methodology and study various learning techniques using data collected on a recent high-performance microprocessor design.