For which particular use case? There are GPU-accelerated applications that need little CPU performance, because pretty much the entire workload runs on the GPU, and there is only a small amount of communication needed with the CPU. Some molecular dynamics applications fall into this category, to give a concrete example.
Looking at the totality of GPU-accelerated applications (thousands by now, I think), one observes that while many get excellent acceleration from the use of GPUs in the parallel portion of the workload, they still contain a certain, sometimes significant, serial portion. As GPUs get faster much quicker than CPUs, that serial portion becomes more and more of a bottleneck for the overall application (-> Amdahl’s Law).
Therefore I usually recommend to CUDA programmers and users the use of CPUs with very high single-thread performance (specifically, a non-boost CPU frequency > 3.4 GHz) to address the serial portion of hybrid CPU+GPU codes, while going easy on the core count (4 or 6 cores with hyperthreading) for cost reasons.
There are other aspects, such as CPU/GPU communication which can be impacted by the number of available PCIe lanes provided by the CPU, but this is a less common issue.
Note that “occupancy” is a crude measure of GPU utilization and has nothing to do with the CPU (at least not in any way I can think off right now). It is a function of the GPU architecture and resource utilization of the code running on the GPU.