This is a common problem with modern supercomputers, which are all clusters of some sort. We have a supercomputer at NASA Goddard, much smaller than Pleiades, that gets new "Scalable Units" every few years. The first parts of it are now long gone and what's there is a mix of several recent generations of Intel CPU and some Intel Phis. I'm told that some codes run better on the older CPUs due to caching and other architectural issues.