Most control architectures for legged locomotion are either torque or position controlled. In this paper, we investigate their differences and performances. Aiming to choose the most appropriate scheme for the robot TALOS, we benchmark three control schemes: The first one optimizes joint velocities based on hierarchical quadratic programming; the second one optimizes joint accelerations based on weighted quadratic programming; and the last one optimizes joint torques, also based on weighted quadratic programming. We compare these controllers in terms of tracking error, energy consumption and computational time by using Gazebo simulations of the robot walking on flat horizontal ground, tilted platforms, and stairs. Remarkably, our torque control scheme allowed TALOS to walk forward at 0.6m/s, the highest walking velocity achieved so far in simulation.