[Athas]

Functional array programming on GPUs works pretty well, and is an immutable programming style. Your object granularity tends to be pretty big (entire arrays), so the low-level shared memory concerns that inevitably arise at some point in the software stack are easy to handle for the compiler and/or runtime. Basically, you have much fewer synchronisation points, so you can afford for them to be slow.

It's a more restricted programming style than common functional programming - in particular, recursion and advanced control flow is out - but that's the name of the game on GPUs, no matter what you do.

[pjmlp]

And there are at least Haskell and F# implementations for GPGPU, not sure about other languages.