The steam-assisted gravity drainage (SAGD) process is currently the widely used one among the in-situ recovery methods to produce bitumen from Alberta oil sands in Western Canada. A thermal process requires very small grid size to provide the better description in the reservoir simulation model than the coarse grid; however the simulation runtime will take longer. The relationship between the number of grids and runtime is not linear but exponential. It is important to design the proper grid size giving reasonable results with shorter runtime. In this study, the optimal grid system design has been investigated through numerical simulation sensitivity studies for the SAGD process. A 1×25×1 m (i, j, k direction; j is wellbore direction) grid size is accepted as a standard size for the SAGD simulation. Grid size sensitivity study has been conducted to determine the maximum grid size that shows a closer result to the 1×1 m case and also the impact of grid size in j-direction as well as the i/k ratio in SAGD simulation. The simulation results shown an i/k ratio is more important than a grid size itself. Based on the CSOR and oil production as well as steam chamber shape, the 2×2 m grid case is closer to the 1×1m case results than 3×1 m case. For the grid size optimization in the wellbore direction (j direction), the grid size of over 25 m cases are no big difference in both production and steam chamber shape for a homogeneous model, and there is no steam chamber propagation to the j-direction. The maximum grid size in j-direction to see the wellbore end effect is 10 m in this study. Considering the reservoir heterogeneity, if shale barriers exist, the impact of grid size in j direction is more important and the smaller grid size is required for the proper numerical simulations to describe steam chamber development in field scale SAGD project. For j-direction grid design, a hybrid type of grid system, a fine grid size in wellbore end zone and regular grid size in wellbore zone, may help to see the steam chamber development with saving a degree of simulation runtime.