My final year thesis was on implementing some algorithms with Accelerate, and one of the things I noted was that on a 2009 Macbook Pro (256 megabyte integrated Nvidia GPU), a single threaded C program runs faster than using Accelerate, even when all it does is multiplying each element of an array by two. The performance discrepancy is even greater for more complicated problems. So, before you jump in to use this and expect better performance on embarrassingly parallel problems, make sure your Nvida GPU is not integrated and has lots of memory.

Of course this new package is different because it uses both CPU/GPU...

I also found Accelerate programs hard to debug. You cannot use "trace" to print out stuff during computation because that is a CPU instruction.