Embedded Feature Selection in MCU Performance Screening

In safety-critical applications, microcontrollers must satisfy strict quality constraints in terms of maximum operating frequency (Fmax). Data from on-chip ring oscillators, the so-called Speed Monitors, can be used as features of Machine Learning models to predict Fmax. Increasing the number of ring oscillators on the chip can increase the information retrieved about the device’s speed. However this may also lead to overfitting, and a lack of generalization capabilities.This paper focuses on supervised feature selection in performance screening during the early phase of prototyping. The aim is to reduce the number of features while maintaining the accuracy of machine learning models. Two distinct approaches for obtaining feature rankings based on ring oscillators’ significance in predicting performance are compared: one based on Recursive Feature Elimination, and one on regularized linear models. Experiments showed that the chosen subset of features leads to simpler ML models that can achieve lower prediction error, reducing overfitting. This permits avoiding inserting the full set of sensors in the final product, saving money and physical space in the silicon.