Models for soft errors in low-level caches