Simultaneous Conjugate Gradient and iAFF-UNet for Accurate IR Drop Calculation

IR drop analysis has become a computationally challenging problem with the shrinking of advanced process nodes. Solving the IR drop problem is time-consuming and an accurate and fast IR drop calculator is crucial for shortening the design cycle. In this work, we introduce an innovative IR drop calculation framework based on the conjugate gradient method and iAFFUNet network. iAFFUNet incorporates the UNet structure with the iterative attention feature fusion (iAFF) blocks to refine conventional approaches of feature concatenation and fusion. iAFF blocks employ multi-scale channel attention modules to enhance feature representation. Furthermore, we leverage intermediate results from the conjugate gradient method as augmented features and utilize graph attention networks for initial value calculation, thereby expediting the iteration process. Alternatively, the matrix operation process can be further accelerated using GPU optimization. During the training phase, we adopt a transfer learning strategy by fine-tuning limited real circuit datasets based on a pre-trained model obtained from training with a substantial amount of synthetic circuit datasets. Experimental results on the ICCAD 2023 contest real hidden testcases under the Nangate 45nm process node show that our model achieves an average improvement of 48.7% and 53.9 % in MAE compared to the contest's champion and the second place, respectively. Additionally, our model achieves a 39.8% reduction in CPU runtime compared to the champion of the contest.