Abstract: In order to accurately estimate execution cycles of programs running on ARM architectures as soon as possible, a simulation-based non-linear estimator is proposed. The estimator consists of two cascade modules: profiling program function module and program execution time prediction module. The module of profiling program function can be directly implemented by Sim-profile, an instruction-accurate simulator. According to the non-linear behavior of the program execution time in advanced processors and the dynamic instruction counts during program executions, the program execution time prediction module is implemented by an artificial neural network (ANN). However, besides the problem of local minimization, ANN is not suited to solve small-sample set regression. It depends on the priori knowledge of designers, which could determine the topology of the model and finally impact its performance. In order to conquer the limitations of ANN, a non-linear method based on the least squares support vector machine (LS-SVM) is further proposed to map the number of executed instructions into execution cycle counts. Experimental results show that simulation-based non-linear estimators implemented with the two non-linear methods, especially the LS-SVM based method, can achieve higher precision of estimating program execution cycles at lower simulation cost.