Early stages in the development of the inversion-capped atmospheric convective boundary layer (CBL) were investigated by means of large eddy simulation (LES) in conjunction with data from a wind tunnel model of the CBL. The LES experiments were conducted for two CBL types. The first investigated flow was the spatially evolving quasi-stationary CBL reproduced in the wind tunnel. The second was the traditional nonsteady, horizontally quasi-homogeneous CBL in a periodic domain. Thus, a pure case of spatial CBL evolution was analyzed versus a pure case of CBL temporal development. A constant value of heat flux at the flat bottom surface was prescribed in both cases. The employed LES used either the turbulence kinetic energy (Deardorff) or Smagorinsky subgrid closures. The LES runs for the spatially evolving CBL were conducted only with Deardorff closure. Numerical experiments for both boundary layer types, using Deardorff closure, showed that in the absence of, or with very small, initial temperature disturbances, drastic heat buildup occurs for some time (over some distance) in the near-surface region of the flow. With no disturbances, the resolved-scale convection was either initiated with considerable delay (in the case of a spatially evolving CBL) or not initiated at all (in the nonsteady CBL). In the case with very small initial temperature perturbations, the heat buildup was followed by explosive development of convection, resulting in large velocities that rapidly mixed heat vertically, overcompensating for the specified heat flux at the surface. In the nonsteady CBL case, this resulted, for a time, in excessive mixing so that thermal gradients in the surface layer became insufficient to maintain the intensity of convection. A subsequent decrease in convection occurred, followed by a secondary heat buildup at the surface, indicating a cyclic nature to the initiation of convection. The intensity of this cycle quickly decreased as a steady convective state was reached. For cases with a rather large initial temperature perturbation (on the order of 5K), rapid convection immediately ensued, causing a similar, but shallower and less severe, convective over-mixing of the near-surface temperature field. These cases tended to achieve a steady state earlier in time than those with a very small temperature perturbation. In between, cases with moderate temperature perturbations (on the order of 1K) quickly achieved a steady convective state without cyclic bursts in convection. The LES turned out to be sensitive to the nature of initial perturbations. When Deardorff closure was used, vertical velocity perturbations provided a rather rough model startup (followed by recovery), but simulations with temperature perturbations ran smoothly. When Smagorinsky closure was applied, initial temperature perturbations forced the crash of the LES. However initialization with velocity perturbations at start time resulted in a smooth CBL development.