Right, but a lot of those shaders will run fine on the CPU as opposed to the GPU. Especially with the current lineup of supported Intel Macs, there's pretty much no supported machines that don't have comprehensive vector computation pipelines on the CPU-side of things.
IMO the best balance would be maxing out the AVX pipeline but keeping the CPU at base clock, and then leave the rendering/framebuffer manipulation to the GPU. I suspect that running both the simulation and 5k/2k displays simultaneously is too much for the GPU to handle.
GPUs are multiple orders of magnitude faster than a CPU's max theoretical throughput. If a GPU is slowing down the solution is a faster GPU, the CPU won't be able to help out much at the same stuff.
You think running on a CPU would save power? CPUs are the least power efficient because they are the most general. You should put some numbers on 'running slower' and 'less powerful'. GPUs for the same power are going to have about 20 times more single precision floating point operations.
IMO the best balance would be maxing out the AVX pipeline but keeping the CPU at base clock, and then leave the rendering/framebuffer manipulation to the GPU. I suspect that running both the simulation and 5k/2k displays simultaneously is too much for the GPU to handle.