I prefer per-core benchmarks. Things like that scale linearly, so it gives you the best idea about machine performance.