Electrophysiological simulations can help to identify atrial arrhythmic mechanisms, although their spatial resolution poses a challenge between simulation accuracy and performance. Grid convergence can provide a resolution that balances robust results and unnecessary computation. Endocardial and epicardial left atrial surfaces were segmented from the CT scan of a patient referred for AF ablation, and four volumetric meshes were generated with different tetrahedral element resolution. Coronary sinus pacing was simulated with the Koivumaki atrial model, and simulations were characterized by averaged and/or summarized local activation time (LAT), action potential duration (APD) and amplitude (APA), to evaluate convergence to continuum resolution (zero-spacing) by Richardson extrapolation. Atrial meshes resulted in average edge sizes of 0.66, 0.52, 0.36 and 0.31 mm, from the coarser to the finest. Depolarization time converged from 160 ms (coarser) to 144 mm (zero-spacing extrapolation), showing a relative error of 11% on the coarser mesh, and < 2% for the 3 finest meshes. Average LAT converged from 98.01 to 87.19 ms, again showing errors >10% for the coarser mesh, and relative errors <1% on the rest. Average APD and APA showed fast convergence, with less than 1% of relative error in any case. Resolution of at least 0.5 mm is necessary to obtain relative errors less than 1% with respect to the ideal case (zero-spacing) in tetrahedral-based atrial simulations. Mesh convergence in other atrial rhythms, such as fibrillation or tachycardia, should be evaluated to identify the optimal simulation resolution.