Enhancing Power Prediction in Digital VLSI Circuits Using Diffusion Models: Synthetic Data Generation and Performance Evaluation

Abstract

Accurate power forecasting plays a crucial role in optimizing the performance of digital VLSI circuits, particularly as design complexities continue to grow. This research delves into the use of diffusion models to create synthetic data aimed at improving the accuracy of power predictions in machine learning frameworks. Running simulations within the HSPICE environment and using advanced CMOS nodes yielded realistic datasets that were employed to train the proposed models. The synthetic data not only resembled real-world data closely but also effectively complemented limited datasets, leading to a significant improvement in power prediction performance metrics. This study underscores the potential of using data augmentation through diffusion models as an innovative strategy in VLSI design.