Figure 1 shows the relative performance for various architectures. In the first stage, it performs simulated annealing which has floating point computation in every iteration. In the second stage, it does an iterative place and route in which there is symbolic computation, but experiments with TimberWolfMC has shown that very few iterations are needed in practice. So the computation is dominated by floating point operations. i586_120, 21064_182 and Sparc20_100 have running times in the same range with i586_120 having a slight edge. Sparc10_50 performs better than R4400_60. This again leads us to conjecture that the Sparc machines have a comparatively better floating point unit.