Global and Local Attention-based Inception U-Net for Static IR Drop Prediction

TL;DR

A global and local attention-based Inception U-Net for static IR drop prediction that incorporates the Transformer, CBAM, and Inception architectures to enhance its feature capture capability at different scales and improve the accuracy of predicted IR drop.

Abstract

Static IR drop analysis is a fundamental and critical task in chip design since the IR drop will significantly affect the design's functionality, performance, and reliability. However, the process of IR drop analysis can be time-consuming, potentially taking several hours. Therefore, a fast and accurate IR drop prediction is paramount for reducing the overall time invested in chip design. In this paper, we propose a global and local attention-based Inception U-Net for static IR drop prediction. Our U-Net incorporates the Transformer, CBAM, and Inception architectures to enhance its feature capture capability at different scales and improve the accuracy of predicted IR drop. Moreover, we propose 4 new features, which enhance our model with richer information. Finally, to balance the sampling probabilities across different regions in one design, we propose a series of novel data spatial adjustment techniques, with each batch randomly selecting one of them during training. Experimental results demonstrate that our proposed algorithm can achieve the best results among the winning teams of the ICCAD 2023 contest and the state-of-the-art algorithms.

Figures (8)

• Figure 1: Visualization of seven used features and a ground-truth of IR drop map for testcase17. (a)(b)(c) are proposed in the literature DBLP:conf/aspdac/ChhabriaAPPJS21. (d)(e)(f)(g) are our proposed new features.
• Figure 2: Our proposed model architecture, where $number \times H \times W$ signifies $channels \times height \times width$. During the encoding phase, the input features undergo four down-sampling operations with a factor of 2. Hence, $H_2 = H_1/2$, $H_3 = H_2/2$, and so forth. Conversely, during the decoding phase, four up-sampling operations are conducted with a factor of 2. We replace the majority of convolutions in the U-Net with Inception modules. To achieve a broader perceptual field, at the beginning of the decoder, we employ a global attention block, primarily utilizing the Transformer architecture. In the subsequent stages, we integrate the local attention blocks, whose pillar is CBAM. The schematics of these two attention blocks are respectively depicted in Figure \ref{['fig:global_atten']} and Figure \ref{['fig:local_atten']}.
• Figure 3: Three different Inception modules.
• Figure 4: Global attention block. (a) is the overall flow of the global attention block. (b) is the diagram of the Transformer block, where $\oplus$ represents addition.
• Figure 5: Local attention block, where $\oplus$ represents addition and $\otimes$ represents multiplication. For detailed architectural specifics, please refer to DBLP:conf/eccv/WooPLK18.