foldy wrote:For nvidia GPUs you need one CPU core per GPU to feed them. So if you have 3 nvidia GPUs like kiore and a 4 core CPU then only one CPU core is left for CPU folding.
If he now would setup a CPU slot with 4 cores then it would be in conflict with 3 GPUs which means it slows them down.
I'm not sure exactly why, but the effect is worse than that though. If I run a CPU:4 slot on an 8 thread i7 with 2 GPU, I get an overall reduction in points. There must still be some contention and possibly increased latency to CPU and memory resources. Also, with more cores active, CPU boost speed goes down, and an extra 50 W or so dissipated in the chassis will increase the thermal throttling of the GPUs. Less than a 1% reduction in GPU performance (2x 980) will wipe out the PPD from a 4 thread CPU slot.That's all before you start considering the efficiency of things where the CPU comes out pretty poorly at the moment.
I'm sure there is still important science being run by CPUs, but in my opinion, that is probably best left to dedicated CPU machine, such as majority running with just integrated graphics or the multi-cpu servers and workstations where their CPU resources can really make a difference. If you're running GPUs, I'd concentrate on the getting the most out of them, rather than complicating things with a CPU slot.