In this paper we present high sample rate array architectures, stack filter-based architectures and sorting network-based architectures for computing recursive median filters. We first develop non-pipelined architectures with lower sample periods than existing architectures, and then reduce the sample period further by pipelining. Pipelining is achieved by developing new recursive algorithms which have additional delays in the feedback paths, and using the delays as pipeline latches. We show that for 2 levels of pipelining, the sample period of the array and stack filter-based architectures reduces by a factor of 2; the sample period does not however reduce proportionately for higher levels of pipelining because of the large implementation overhead. The sample period of sorting network-based architectures, on the other hand, can be reduced to any level for MSB-first implementations.