Comprehensive Modeling Methodologies for NoC Router Estimation

Networks-on-Chip (NoCs) are increasingly used in many-core architectures. ORION2.0 [18] is a widely adopted NoC power and area estimation tool that is based on circuit-level templates, which use specific logic structures to model implementation of different router components. ORION2.0 estimation models can have large errors (up to 185%) versus actual implementation, often due to a mismatch between the actual router RTL and the templates assumed, as well as the effects of optimization tools in modern design flows. In this work, we propose comprehensive parametric and non-parametric modeling methodologies that fundamentally differ from logic template based approaches in that the estimation models are derived from actual physical implementation data. Specifically, we propose a new parametric modeling methodology as well as improvements to previous work on non-parametric modeling. Our work on parametric modeling proposes (1) ORION NEW models, developed using a new methodology that does not assume any logic templates for router components, (2) the use of parametric regression to fit the new models to post-layout power and area values, and (3) modeling extensions that enable more detailed flitlevel power estimation when integrated with simulation tools such as GARNET. The ORION NEW methodology analyzes netlists of NoC routers that have been placed and routed by commercial tools, and then performs explicit modeling of control and data paths followed by regression analysis to create highly accurate NoC models. Our work on non-parametric modeling expands on our previous work [17] and demonstrates the use of four widely used nonparametric modeling techniques: Radial Basis Functions, Kriging, Multivariate Adaptive Regression Splines, and Support Vector Machine Regression. The estimation models are also derived from actual physical implementation data. We observe that nonparametric modeling techniques have low overhead and complexity of modeling, and are highly accurate. When compared with actual implementations, our modeling methodologies achieve average estimation errors of no more than 9.8% across microarchitecture, implementation, and operational parameters as well as multiple router RTL generators. These methodologies are being implemented in an ORION3.0 distribution [19], [66].