PreRoutSGAT: Pre-Routing Timing Prediction Based on Graph Neural Network with Global Attention

Pre-routing Static Timing Analysis (STA) is pivotal in evaluating the timing functionality during digital design verification. Its accuracy and speed in assessing various node features within the circuit, including arrival time, slew, and slack, as well as edge features such as net delay and cell delay, have a profound impact on the entire digital design process. An unsatisfactory STA can lead to recurrent iterations of the entire placement and routing process, or even a redesign of RTL code. Leveraging the congruity between circuit timing paths and graph structures, we propose our Graph Neural Network (GNN) model, PreRoutSGAT (Pre-Routing Simple Global Attention), for prerouting timing prediction. We utilize a simple and effective residual structure to alleviate the prediction accuracy degradation brought by deep GNN, and greatly minimize the GPU memory usage required for training with a linear computational complexity attention convolution layer. Based on PreRoutSGAT, our approach encompasses three models that are built on the same concept, but differ slightly in structure. These disparate models are trained on 14 circuit timing graphs and evaluated on 7 graphs. Experimental results demonstrate that the proposed models, in comparison to the previous SOTA model TimingGCN, achieve superior results in predicting multiple dimensions. Additionally, the proposed models achieve better performance and the average memory usage of the GPU is reduced by about 77.4% without increasing the time overhead.