Victim cache
A cache (or cache policy) that catches data lines just evicted ("victimized") from a higher cache level, instead of discarding them immediately. If the CPU soon needs the evicted line again it's still around one level down, avoiding an expensive memory access.
Mechanic: when a line must leave L2 (to make room for new data) it ends up in the "victim cache" / is handled by a victim policy in L3 instead of disappearing entirely. If the core needs it again shortly (common due to locality) it's found there — a hit instead of a long RAM fetch. AMD has long used a "victim cache" design for its L3: the L3 is filled with lines evicted from L2 (which makes it exclusive — it doesn't duplicate what's already in L2, but holds precisely what L2 no longer has room for). Win: better utilization of the total cache capacity and fewer misses on data that "just left". Originally a small separate buffer concept (Norman Jouppi), today a policy in the shared L3. Related to inclusive cache and cache miss.