Reliable multicast protocols provide all-or-none delivery to participants. Traditionally, such protocols suffer from large buffering requirements, as receivers have to buffer messages, and buffer sizes grow with the number of participants. In this paper, we describe an optimization that allows such protocols to reduce the amount of buffering drastically at the cost of a very small probability that all-or-none delivery is violated. We analyze this probability, and simulate an optimized version of an epidemic multicast protocol to validate the effectiveness of the optimization. We find that the buffering requirements are sub-constant, that is, the requirements shrink with group size, while the probability of all-or-none violation can be set to very small values.