Recent work has argued that federated learning over wireless channels can be accelerated by a factor of $K$ (the numbers of users), by using computation over multiple-access channels to directly average the gradients. This implicitly presumes that timely channel state information is available at the transmitters, which may not be feasible for large $K$. This paper presents a simple scheduling algorithm that only uses channel state information at the receiver to activate a subset of the users for computation, and accelerates averaging by a factor of $K^{2/3}$.