I'm skeptical of the 8x speedup for several reasons [...] yet the CPU is only running a single thread

That's why I'm not skeptical at all. A GPU program can operate over THOUSANDS many more data items in parallel than a single-threaded scalar CPU program can. Yet the speedup is not thousands, not even hundreds, but a mere 8.

Fits perfectly with

main one being that this particular problem does not fit the paradigm of problems that work well on the GPU