Here's an interesting paper on harnessing the parallelism of a graphics processing unit in radial velocity searches for extrasolar planets:
http://arxiv.org/abs/0812.2976
They use the latest generation of NVIDIA GPUs, which support double precision arithmetic. However, single precision floating point arithmetic on these GPUs is still much faster, and so the focus of the paper seems to be on mixed precision techniques for evaluating Kepler's equation. Double precision is only used where absolutely necessary.
--Chris