Data-dependent cycle-accurate power modeling of RTL-level IPs using machine learning

In a chip design project, early design planning has a strong impact on the schedule and the cost of design. Power estimation is part of early design planning, and it greatly affects design decisions. Power modeling performed at a high level of abstraction is fast but inaccurate due to lack of circuit switching activity information. By contrast, power modeling performed at a low level of abstraction is more accurate as the synthesized circuit synthesis is known, but this simulation is typically slow. This report explores a power modeling approach performed at register transfer level (RTL). It exploits machine learning models in order to have a fast yet relatively accurate cycle-by-cycle power estimation. The approach is data-dependent, where cycle-specific models are trained based on the switching activity of signals obtained from RTL simulation and cycle-by-cycle power values obtained from a reference gate-level simulation of an existing RTL design. Therefore, if any changes are applied to the RTL design, re-training of models is required. The approach aims at obtaining fast yet accurate power predictions for new invocations of a given trained model using signal activity information collected during simulation of the unmodified RTL. At a low level, the complete visibility of signals in a design unintuitively might cause overtraining the model leading to inaccurate estimation. The suggested model employs automatic feature selection in each cycle. Based on the invocations used to train the cycle-by-cycle models, only signals that may switch during a given cycle will be selected as the features for their respective cycle-specific model. The method was tested on an 8-by-8 DCT design and the power estimates were within 6.5% of those from a commercial power analysis tool. This report also simulates and compares the approach of cycle-specific models to the approach of a single global model for all cycles and show that the cycle-specific approach is twice as accurate.