Bounding Power Conversion Loss Point Estimates via Supervised Learning Based Bayesian Regression

Design of computing systems needs prediction of voltage regulator power losses for unseen systems based on a single measurement. Single measurement serves as a point estimate and fails to address variations like use of different controllers, drivers, layout, board stack-up, aging, etc. This paper presents existing concepts in regression, their limitations, and proposes use of supervised learning-based Bayesian ridge regression. This approach achieves both parameter shrinkage and distribution computation. Use of parameter distribution predicts the measurement variation from the mean to bound the power loss for unseen data. In addition, to show superiority of the proposed approach, power loss prediction is assessed in higher residency low current state of microprocessors and compared against other advanced regression methods. Finally, model order tradeoffs are addressed by factoring parameter count and modeling efficacy.