Isn't the problem that that CUDA for FreeBSD does not exist in the first place?
Nvidia can't be bothered to port it.
What about using CUDA with bhyve with "PCI-E passthrough"? It should work near native speed. This was suggested here: https://news.ycombinator.com/item?id=10767172
This is best explained with an example. Say you want to install TensorFlow with CUDA 7.5 and cuDNN v4 support, which is necessary for training large neural nets.
Here's how you do it on, say, Ubuntu:
And here's a blog post with a long list of instructions, explanations, and caveats for doing the same thing on FreeBSD: Except that on FreeBSD apparently no one has posted instructions for getting CUDA or cuDNN to work with TensorFlow, so you're in for a world of "fun" trying to figure out how to do that on your own.