Fast Prediction of Dynamic IR-Drop Using RNN

What is it about?

IR-drop is voltage drop occurred by current flow in circuit power delivery network (PDN). Large IR-drop can lead to malfunction of circuit, thus IR-drop violation has to be checked. However, IR-drop calculation through time (so called dynamic IR-drop analysis) requires large runtime. We adopted recurrent U-net for fast dynamic IR-drop prediction.

Featured Image

Photo by Jason Leung on Unsplash

Photo by Jason Leung on Unsplash

Why is it important?

While decoupling capacitors are widely used to reduce IR-drop, previous machine learning based IR-drop prediction methods did not consider them. However, due to recurrent layer in our machine learning model, our method can consider time-varying effect of capacitors in PDN, so that dynamic IR-drop can be accurately predicted. As a result, we could achieve 16 times faster runtime than commercial IR-drop analysis tool with about 12% error.