They use your GPU + 1 CPU per WU. What you're seeing is normal.
Most people also running CPU WUs have changed it so that 1 CPU is available to the GPU WUs.

You're welcome.

I think that some sub-routines are better suited to being processed by a CPU. The other side of this is that not all routines is suitable/ideal for processing by a GPU. Thus, the need for a CPU thread from an otherwise all-work-exclusively-for-the-GPU.

As far as reserving a CPU thread is concerned, I see BOINC as treating a GPU-WU's CPU-part just like any other CPU thread. A cruncher need not pre-reserve a CPU-thread for HCC-GPU-WU. Here's what I observed happens: If a HCC-GPU-WU does not have a CPU-thread, it pauses one CPU thread to free that thread for the GPU's use. The paused CPU-thread will soon get its thread from the earliest CPU-based WU to complete. At this point, the reservation is now complete: N-1 CPU-threads reserved for CPU-based WUs while 1 thread reserved for the CPU-part of the HCC-GPU-WU. If any CPU-based CPU-thread is paused, another CPU-based WU will get the CPU-thread; and if a GPU-based CPU-thread is paused, another GPU-based WU will get that CPU-thread.
;

Collatz Conjecture is really simple math that can be easily accelerated entirely on a primitive graphics card with minimal CPU interaction, while a project like this is much more complicated. I doubt there are much changes they could do to lessen the number of cycles required on the CPU end of things, but I'm interested in why you say one full core is used all the time, on my FX-8120 the cores are only fully utilized when doing the start/end points, aka the CPU phases, and during the GPU phase there is minimal calculations happening (20% per core).

In short, you will need a HSA GPU with a HSA app before CPU load can be reduced.

Know one thing for sure, that with Hyperthreading [Intel!] all unused cycles go to use towards other CPU tasks when the load of the physical cores exceed 100%, meaning if a GPU task is not using all CPU cycles, other tasks benefit. I've tested this last week. AMD has no Hyperthreading, so suppose that an idle thread is an idle thread, so that would be the the difference.

EDIT: BTW, think programs like Process Lasso can tell which BOINC task/process is using which core [the affinity]. Tools like that can force for multiple GPU tasks to share the same CPU threads, as can it be used to force the full CPU tasks to use cores that you don't want the GPU tasks to run on. Else, think BOINC is automated to spread to just the next available idle thread for a task, and for the rest, is oblivious of outside hands forcing 1 or more onto a single CPU core. Been there done that [had 3 HPF2 running off 1 thread], a FAAH on a second thread and 2 cores idle on the quad [Q6600 has no HT].