I really dont have a high hope for the denver cores, the only good reason for denver cores+GPU is that you can built a cluster with only GPUs without the need of any x86.
But I dont believe any supercomputer in the future will build solely based on Maxwells, the most obvious reasons is the significant amount of software used in HPC were based on x86, it takes huge investment and effects to port some of them to a system with a new ISA, I dont belive Nvidia have the musle to do that.
For desktop application, the reason is the same, the x86 market is so huge that any shift to an different ISA is so costly to the degree no one would ever think of that, let along a relatively small weight like Nvidia.
So at least for the generation of Maxwell, it is safe to say the vast majority of them will remain at a PCIE slot.
And considering the existence of serious PCIE bandwidth constraints and very capable x86 host CPUs, I seriously doubt the point of having a local CPU on a GPU.
Most of the algorthims I wrote for hybrid computing systems will divide the workloads of tasks into two parts, one part is highly parallelized, and sending them to GPU, the other part is not that easy to be parallelized, and so I let CPU to handle.
With async data transfer/work, the mechanism works very well, utilize the system's computing resources better, and most importantly, let each computing devices handle only the task they are GOOD at. I seriously doubt a Denver ARM core could outperform the latest Xeon core on serialized tasks, not to mention the added die-size/silcon that would otherwise being more useful to GPU's main job: doing highly paralelized computations.
