TBB targets shared memory architectures. It works even on single-core single-processor machines, though for performance benefit there should be two or more hardware threads (i.e. full-fledged cores or at least hyper-threads) available.

TBB proivdes a high-level, system-independent interface; in this sense, it's "portable". Inside, TBB containsmachine-, OS-, and compiler-specific low level primitives which should be provided for (ported to) a platform of interest.

The TBB scheduler is a work-stealing scheduler. It's build on different principle than OpenMP schedulers you listed. The scheduler provides natural support for load balancing and nested and recursive parallelism. TBB provides high-level algorithm templates such as parallel_for to support "flat" loop parallelism; internally, it recursively divides the flat iteration space.

The exploitation of concurrency can not be pre-implemented in TBB; it's something you should do in your application with help of TBB (or other means). For this purpose, TBB provides algorithms to express both data and task decomposition, and if these are not enough, one can implement a custom algorithm using the low-level interface to the task scheduler provided by TBB.