ML for Design QoR Prediction

Design quality of results (QoR) spans metrics of design process outcomes, such as power, performance, area, or runtime, at all stages of the design process. Prediction of design QoR enables efficient searching of the solution space for design, through pruning of unpromising design paths. Without forward-looking predictions, outcomes can only be anticipated constructively, i.e., by running tools. Additionally, a predictor of outcomes for a given stage of design optimization can inform the optimization objectives at earlier, higher-level stages. This chapter gives an overview of machine learning (ML)-based design QoR modeling and prediction: its scope and challenges, key methods, and example application contexts. Generic uses of ML in QoR prediction include shifting the cost vs. accuracy trade-off that is inherent to estimating performance or other design process outcomes, “seeing ahead” to enable early design space exploration and doomed-run filtering, and providing learning and autotuning within multistage design optimizations. At the same time, challenges for ML-based QoR prediction include limitations on data and infrastructure, noise in design tool or flow outcomes in regimes of interest, and difficulty in formulating useful and actionable predictions.