[unlinked_dll]

There are diminishing returns when evaluating recursive formulas using SIMD, as the cost to gather/scatter into the vectors, duplication of computation, and transfer overhead to the GPU cost more than doing it in a single core with vector instructions.

Real-time audio processing is probably the most common case where you see huge gains in SIMD on the CPU but very minimal gains on the GPU. Other examples are certain polynomial approximations and physics simulations.

Also string manipulation, parsing, evaluating a regex. The overhead of getting to the GPU outweighs keeping it on the CPU core.