Supercomputing of circuits simulation

• Authors: Y. Fukui;H. Yoshida;S. Higono
• Venue: Proceedings of the 1989 ACM/IEEE conference on Supercomputing
• Year: 1989

The Circuit Analysis is very important in the development of LSI. We have been conducting the speed-up of the circuit analysis program SPICE-GT. SPICE-GT is based on the SPICE 2G.6 program from the University of California Berkely. We achieved speed-up by the factor of 300 compared to the original SPICE. In this paper, we discussed the speed-up methods concerning mainly on a computer side of view.There are DC analysis, AC analysis and transient analysis in a circuit analysis, the transient analysis is the most resource consuming step. It can be divided into next three items : (1) computation of circuit matrix, (2) LU decomposition, (3) others (input/output, etc). Because these three items differ in character, we applied different tuning method for each item. Our tuning method are (1) vectorization, (2) multi-tasking(parallel processing), (3) code-generation.The computation of circuit matrix can be executed in parallel, therefore we applied the multi-tasking and the vectorization. The multi-tasking version is about two times faster than the uni-tasking version in elapse time. When we applied the vectorization, the vectorization version is about three times faster than the scalar version in CPU time. Our vectorization method is generalized for the programs that have complicated conditional jump.The computation of LU decomposition can not be executed in parallel, we applied the code-generation method (scalar version, vector version, block elimination version). The code generation is the speed-up method for the linear equations, it is about 100 times faster than FORTRAN version.In input/output, we applied usual tuning method, these methods all together resulted in the speed-up by the factor of about 11 in the circuit analysis.