The computational cost of COAL for solving a DSE problem contains two parts, the simulation cost spent on obtaining the actual labels of design configurations, and the training cost spent on selecting unlabeled data and updating the regression trees. Figure 5 compares the simulation cost and training cost consumed by COAL on different benchmarks. We can see that the simulation cost always overwhelmingly dominates the training cost. In average, the simulations consume 90% of the computational cost. On benchmark mesa, the simulation cost even exceeds 97% of the entire computational cost. It is worth noting that for each benchmark (benefited from SimPoint [Sherwood et al. 2002]) we only simulate 100 million representative dynamic instructions, while in industry all dynamic instructions (more than 10 billion for most benchmarks) should be simulated. Under this circumstance, the training cost of COAL can be completely neglected in practice.