We present a detailed analysis of the performance of the semiempirical divide and conquer method as compared with standard semiempirical MO calculations. The influence of different subsetting schemes involving dual buffer regions on the magnitude of the errors in energies and computational cost of the calculations are discussed. In addition, the results of geometry optimizations on several protein systems (453 to 4088 atoms) driven by a quasi-Newton algorithm are also presented. These results indicate that the divide and conquer approach gives reliable energies and gradients and suggest that protein geometry optimization using semiempirical methods can be routinely feasible using current computational resources.