DSPs and other application-specific processors expose the cache hierarchy as a set of scratchpads. This works very well for them, but not for any CPU that is shared between applications, like a server.