This paper proposes a simulation-based soft error estimation methodology for computer systems. Accumulating soft error rates (SERs) of all memories in a computer system results in pessimistic soft error estimation. This is because memory cells are used spatially and temporally and not all soft errors in them make the computer system faulty. Our soft-error estimation methodology considers the locations and the timings of soft errors occurring at every level of memory hierarchy and estimates the soft errors of the whole system using instruction-set simulation. Our experiment demonstrates that the reliability of computer systems depends on not only SERs of memories but also the behavior of software running on the systems.