ToggleGNN: Uncertainty-Aware Vectorless FPGA Activity Estimation

Abstract

Fast, pin-level switching activity estimation is essential for early power, performance, and area exploration on Field-Programmable Gate Arrays. While traditional vectorless tools offer high throughput, they ignore signal dependencies in reconvergent logic and frequently overestimate activity. Replacing these tools with standard machine learning models can reduce this overestimation, but their unquantified uncertainty can introduce the risk of an overly optimistic power budget. To address this, we introduce ToggleGNN, a graph neural network that predicts both switching activity and its underlying uncertainty. By outputting conditional quantiles rather than single point estimates, ToggleGNN allows designers to safely tune power margins and treat controlled underprediction as a calibrated correction. The model features a dual-branch architecture that independently processes circuit structure and signal activity, paired with a dual-head hurdle objective aimed at the highly zero-inflated nature of toggle rates. Experiments show that ToggleGNN improves accuracy over standard baselines, delivers reliable uncertainty bounds, and accelerates power estimation at a fraction of traditional vendor runtimes.