A Standard Deviation Incorporated Attention Module Enhanced U-Net for Routability Prediction
Yinjie Chen, Di Zhang, Hua Chen · 2024
With the remarkable increase in the complexity of Very Large-Scale Integration (VLSI) designs, the integration of artificial intelligence (AI) into electronic design automation (EDA) has become indispensable, particularly in the realm of back-end design tasks like Routing Congestion (RC) prediction and Design Rule Checking (DRC) hotspot map prediction. In this paper, we introduce an innovative variant of the U-Net model, enhanced by an attention module, to accurately predict routability. The attention module incorporates standard deviation operation to extract more comprehensive distribution information from the features, enabling mitigate the adverse effects of data imbalance. Our proposed model outperforms existing models, particularly with rate reductions of 1.5% (RC) and 4.6% (DRC) in Avg-NRMSE when compared to the prior model ibUnet, as evidenced by experimental results on the CircuitNet dataset. Furthermore, it outperforms the ibUnet on model size which exhibits a significant improvement of over 50% rate reduction.