Abstract
Dynamic IR-drop, a voltage drop reflecting current (I) through resistance (R), has become a significant operational constraint in the power delivery network (PDN) of an integrated circuit. A transient matrix solver can take upwards of 6-12 hours to verify one engineering change order (ECO). Separately, a frame-based neural network may miss voltage deficits that build across consecutive clock cycles. To address these limitations, this a prediction method feeds per-cycle power tensors through a shared spatial encoder. From there, a convolutional long short-term memory (ConvLSTM) temporal core carries voltage history from cycle to cycle. A shared spatial decoder then reconstructs a transient voltage map for each cycle, and a temporal max-pooling stage extracts a worst-case peak IR-drop map. An asymmetric quantile loss weights under-prediction more heavily than over-prediction during training. The method can reach a root-mean-square error (RMSE) near 6.70 mV in about 29 seconds and can serve successive ECOs without retraining.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Pendyala, Prateek, "Multi-Cycle Dynamic IR-Drop Prediction for Power Delivery Networks Using a Shared-Weight Encoder-Decoder with Convolutional Recurrence", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/11699