Contemporary Greenland Ice Sheet (GrIS) mass loss is increasingly dominated by decreasing surface mass balance (SMB), rather than iceberg discharge. The most realistic gridded products of GrIS surface mass balance (SMB) are produced by regional climate models at fine spatial resolutions (5-25 km) with sophisticated snow models, forced by either atmospheric reanalyses (to reconstruct past SMB) or earth system models (ESMs, to produce future GriS SMB projections). Another, more self-consistent approach is to simulate GrIS climate and SMB inside the earth system model itself, thereby allowing for global-to-regional feedbacks on simulated SMB. Here, we use the variable-resolution Community Earth System Model (VR-CESM) and explore the impact of spatial resolution on the GrIS SMB. Three atmosphere-only VR-CESM simulations are carried out at increasing resolution over the GrIS: 1° (∼111 km), 0.5° (∼55 km), and 0.25° (∼28 km). We find that simulated SMB in the accumulation area is greatly improved in the higher-resolution simulations, due to better resolved topography and orographic precipitation. In the ablation zone, high-resolution experiments do not necessarily perform better than those at lower-resolution. Overall, our work demonstrates for the first time that regional grid refinement inside an ESM is a viable option for ice sheet SMB studies, providing a self-consistent way of dynamically downscaling coarse resolution models, while maintaining feedbacks with the rest of the global climate system.