The idea of CPU mining is a noble concept but flawed. Once you get the source code out there, someone will privately figure out how to do it via GPU's faster then on CPU's.
Not necessarily.
There exist numerous tricks and treats that render certain types of hardware relatively inefficient.
It is certainly possible to write a crypto-PoW in a manner that is outright hostile to any modern, and for the immediately foreseeable future, possible, GPU.
Ok. It may be possible. But possible isn't going to happen in this case. This is simply not a single threaded type of work. So even if a GPU is bad at it, it still will be possible to break it up. Use lots of ram? Ok, so some (or even 3/4's) of the GPU processing units will be idle, but that will still be many times faster then a CPU. Floating point? SC miners will switch to nvidia vs ATI. The profit motive will break this thing.
Well, there has been some work (in adjacent fields) on types of cryptographic functions that are hostile to pretty much anything but CPUs and verily complex FPGA designs.
Methinks that making something run way worse on any modern GPU than on CPU is definitely possible, though whether CH will succeed in actually ensuring GPU-hostility is something that has yet to be seen.
Conway's Game of Life is an example. There is a lot of memory access and decision making. Although not a cryptographic hash, it is trivial to do
SHA256(SHA256(header) + CGoL(header))
where CGoL is a certain amount of generations of Conway's Game of Life.