Power and IR Drop
GRANNITE reports more than an 18.7-fold speedup with less than 5.5% error versus a commercial reference. IR drop is better served by physics-informed losses than by unconstrained prediction alone. Power estimation centers on estimating swi…
1 sources - 5 claims
GRANNITE reports more than an 18.7-fold speedup with less than 5.5% error versus a commercial reference. IR drop is better served by physics-informed losses than by unconstrained prediction alone. Power estimation centers on estimating switching activity at internal nets. GRANNITE models switching activity using learned directed propagation over fan-in with gate and structural information. PGNN aligns IR-drop prediction with the PDN linear system by embedding the residual norm into the loss.