Next: SPICE
Up: Analysis of Results
Previous: VIS
Figure 2 shows the relative performance of various
machines for this application. We observe that increasing the circuit
size does not have too much effect on the performance ratio. This is
because the circuit is dynamically partitioned into stages and
computation is performed on each stage separately. These stages are
typically small and fit the first level cache of the machines. The
instructions performed are mostly floating point. Hence, L-1 cache
size and access time and the floating point unit are the major factors
in determining the performance. 21064_182, which has the smallest
L-1 cache access time, was found to be consistently the fastest
machine. The i586_120 and the Sparc20_100 were found to have
comparable performance due to their similar L-1 access time. R4400_60
and Sparc10_50 were found to be the slowest.
Amit Mehrotra
Tue May 6 11:41:31 PDT 1997