design configurations as shown in Table III, where the promising configurations found by its opponent (state-of-the-art DSE approach using ANN) are also presented. We can clearly see that ANN and COAL suggest different design configurations for these two benchmarks given the above design specification.
Table IV further presents the corresponding performance and power of these two configurations (namely, “ANN-conf ” and “COAL-conf ”). On both benchmarks, the actual performance (obtained by cycle accurate simulations) of COAL-confs is better than that of ANN-confs given the 100 watt power constraints. It is notable that on benchmark crafty, the predicted and actual performance/power of COAL-conf are very close while those of ANN-conf are quite different. Quantitatively, the relative error of the predicted IPC of ANN-conf over the actual IPC is 21.5%. Hence, we can conclude that COAL is more effective than the state-of-the-art ANN-based approach.