Neural Combinatorial Solvers Efficiency Threshold Defined

A recent study published on arXiv presents the Amortized Efficiency Threshold (AET), a new metric designed to evaluate neural and heuristic combinatorial optimization solvers. The AET indicates the deployment volume at which a neural solver matches the total energy or carbon cost of a heuristic baseline, given a constraint on solution quality. The authors contend that the prevalent criticism labeling neural solvers as less energy-efficient than CPU metaheuristics overlooks the significance of deployment volume. While training a neural network incurs a substantial fixed GPU energy cost, metaheuristics utilize minimal CPU energy per instance. When the neural network outperforms, the cumulative energy ratio between the two solvers approaches a constant below one, suggesting that neural solvers can be more efficient with adequate deployment. The paper establishes a framework for equitable comparison and asserts that the leap from 'training is costly' to 'net-inefficient' is erroneous.