PowerNet: Transferable Dynamic IR Drop Estimation via Maximum Convolutional Neural Network

Abstract

IR drop is a fundamental constraint required by almost all chip designs. However, its evaluation usually takes a long time that hinders mitigation techniques for fixing its violations. In this work, we develop a fast dynamic IR drop estimation technique, named PowerNet, based on a convolutional neural network (CNN). It can handle both vector-based and vectorless IR analyses. Moreover, the proposed CNN model is general and transferable to different designs. This is in contrast to most existing machine learning (ML) approaches, where a model is applicable only to a specific design. Experimental results show that PowerNet outperforms the latest ML method by 9% in accuracy for the challenging case of vectorless IR drop and achieves a 30× speedup compared to an accurate IR drop commercial tool. Further, a mitigation tool guided by PowerNet reduces IR drop hotspots by 26% and 31% on two industrial designs, respectively, with very limited modification on their power grids.

Dynamic IR-drop ECO optimization by cell movement with current waveform staggering and machine learning guidance

“…The cumulative current of each branch in the PG network is the direct cause of dynamic IR-drop. However, state-of-the-art dynamic IR-drop prediction works [6,7] use only the current/power of individual cells as features for training without considering how cumulative current flows through the PG network. Therefore, it is hard for their models to learn the relationship between the raw data and the IR-drop.…”

“…When cell is placed at a specific row, the aligned PG rails are determined accordingly. Constraint (7) ensures that when two single row-height cells come from the same double rowheight cell, their power pins should be aligned to the same rail. We also set the maximum allowable vertical movement to avoid timing degradation and thus can remove variables of zero value from Constraints (2) to (7).…”

“…Constraint (7) ensures that when two single row-height cells come from the same double rowheight cell, their power pins should be aligned to the same rail. We also set the maximum allowable vertical movement to avoid timing degradation and thus can remove variables of zero value from Constraints (2) to (7). Totally, we have ( ) binary variables and ( ( + )) constraints in the ILP for an ECO region with 2 rows and standard cells.…”

“…Compared with time-consuming IR-drop analysis (a chip-level analysis usually takes over 12 hours for a multi-milliongate design), a fast and accurate IR-drop prediction model shortens the turnaround time between IR-drop optimization and IR-drop analysis. Therefore, as summarized in Table 1, state-of-the-art IR-drop prediction works rely on machine learning instead of approximation by analytical formulae [5][6][7]. Ho and Kahng in [5] propose an XG-Boost model to predict static IR-drop for incremental modification of placement and PG network.…”

“…Because the worst voltage droop and the worst ground bounce usually do not occur at the same time, the sum leads to over-pessimistic predictions. Xie et al in [7] devise a CNN-based model to predict dynamic IR-drop. However, they do not extract PG network features and can predict only the IR-drop for each bin instead of each cell.…”