In direct time-of-flight (D-TOF) light detection and ranging (LIDAR), accuracy and full-scale range (FSR) are the main performance parameters to consider. Particularly, in single-photon avalanche diodes (SPAD) based systems, the photon-counting statistics plays a fundamental role in determining the LIDAR performance. Also, the intrinsic performance ultimately depends on the system parameters and constraints, which are set by the application. However, the best-achievable performance directly depends on the selected depth estimation method and is not necessarily equal to intrinsic performance. We evaluate a D-TOF LIDAR system, in the particular context of smartphone applications, in terms of parameter trade-offs and estimation efficiency. First, we develop a simulation model by combining radiometry and photon-counting statistics. Next, we perform a trade-off analysis to study dependencies between system parameters and application constraints, as well as non-linearities caused by the detection method. Further, we derive an analytical model to calculate the Cramér-Rao lower bound (CRLB) of the LIDAR system, which analytically accounts for the shot noise. Finally, we evaluate a depth estimation method based on artificial intelligence (AI) and compare its performance to the CRLB. We demonstrate that the AI-based estimator fully compensates the non-linearity in depth estimation, which varies depending on application conditions such as target reflectivity.