You did not define what efficient means.

For example, a SIMD implementation will outperform a multithreaded implementation on a CPU without support for hardware threads.

My nonsense answer just happens to be what HPC is moving into, and when HPC wants to scale it uses MPI coupled GPUs, not plain old threads.