The problem of extracting Instruction Level Parallelism at levels of 10 instructions per clock and higher is considered. Two different architectures which use speculation on memory accesses to achieve this level of performance are reviewed. It is pointed out that while this form of speculation gives high potential parallelism it is necessary to retain execution state so that incorrect speculation can be detected and subsequently squashed. Simulation results show that the space to store such state is a critical resource in obtaining good speedup. To make good use of the space it is essential that state be stored efficiently and that it be retired as soon as possible. A number of techniques for extracting the best usage from the available state storage are introduced.