This laboratory experiment investigates whether coordination on sunspots may arise among a large number of participants (80-90 subjects), and how this coordination may be influenced by the payoff and the sunspot structures. Our setup is a bank run game where withdrawing is the safe option, but waiting is the payoff-dominant strategy. Comparing behavior in small and large groups, we find major differences that equilibrium refinements fail to predict. Coordination on sunspots never happens in large groups, while it sometimes happens in small groups. Furthermore, coordination failures are systematic in large groups as soon as coordination on the Pareto-dominant equilibrium is risky enough. In contrast, small groups may still converge to the optimal equilibrium even when the safe option is relatively more attractive.