Fast statistical-based interconnect modeling using automatic differentiation

Automatic differentiation is a technique for computing derivatives accurately and efficiently with minimal human effort. We employed this technique to generate derivative information of FCAP2/3 programs that simulate the parasitic effects of interconnects. This derivative information is used in the statistical modeling of worst-case interconnect delays and on-chip crosstalks. The ADIC (Automatic Differentiation in C) tool generated new versions of FCAP2 and FCAP3 programs that compute both the original results and the derivative information. We report on the rise of automatic differentiation and divided difference approaches for computing derivatives for FCAP3 programs. The results show that ADIC-generated code computes derivatives more accurately, more robustly, and faster than the divided difference approach