I didn't know about it! From a cursory look (all I can find is presentation slides?), it doesn't sound like you'll be able to take any off-the-shelf Haskell subroutines and run them in parallel on the GPU; rather, it's an embedded DSL for constructing shader programs. So I imagine you'd have to learn both CUDA and Obsidian to use it.

But it would be awesome if someone came along and proved me wrong.