I want to use this computer to only crunch wit the gpu. How to I do that?

Use the Milkyway preferences on your main account page and set them to not use the CPU for the location (venue) you run your host on.

Also, deselect "MilkyWay@Home N-Body Simulation". It claims to be a GPU application but is a CPU only application that uses all the cores on your machine. At least that is how it behaves under linux.

Also, deselect "MilkyWay@Home N-Body Simulation". It claims to be a GPU application but is a CPU only application that uses all the cores on your machine. At least that is how it behaves under linux.

I am not positive but I think some units use more cpu's then other units.

Hallo Mikey!
All threads dedicated for crunching on the GPU do use neccessarely also the CPU.
The CPU do handel the datatransfer from and to the data-server, has to set up the programm on the GPU and handels the data transfer to and from the GPU and, at the end do arange the results for tranfer back to the data-server. Especially the last one at the end of the task takes allways some seconds, which you easily can observe in the BOINC Manager/Tasks. But this neccessary CPU-time is in this project very low. It varies at me between 4 and 57[sec], with a mean of 10.64 +/- 5.71 [sec] for my GTX550Ti and 16.55 +/- 8.90[sec] for the slower GT440. The relative CPU-usage (CPU-time / Run-Time) is 1.12 +/- 0,58[%] for the GTX550Ti and 0.757 +/- 0.388[%] for the slower GT440. ( These data where derived from 1073 crunched tasks on the GTX550Ti and 472 crunched tasks on the GT440 within the last 12 days.) All GPU-threads are running at priority level ´Normal´, without disturbing my normal use. The higher CPU-time for the slower GPU card results, I assume, from the smaller bandwith of the memory path. The GTX550Ti has 192 lines , whereas the GT440 has 128 lines only.
I hope, this will answer some of your questions.

Kind regards and happy crunching
Martin

Hallo Mikey!
All threads dedicated for crunching on the GPU do use neccessarely also the CPU.
The CPU do handel the datatransfer from and to the data-server, has to set up the programm on the GPU and handels the data transfer to and from the GPU and, at the end do arange the results for tranfer back to the data-server. Especially the last one at the end of the task takes allways some seconds, which you easily can observe in the BOINC Manager/Tasks. But this neccessary CPU-time is in this project very low. It varies at me between 4 and 57[sec], with a mean of 10.64 +/- 5.71 [sec] for my GTX550Ti and 16.55 +/- 8.90[sec] for the slower GT440. The relative CPU-usage (CPU-time / Run-Time) is 1.12 +/- 0,58[%] for the GTX550Ti and 0.757 +/- 0.388[%] for the slower GT440. ( These data where derived from 1073 crunched tasks on the GTX550Ti and 472 crunched tasks on the GT440 within the last 12 days.) All GPU-threads are running at priority level ´Normal´, without disturbing my normal use. The higher CPU-time for the slower GPU card results, I assume, from the smaller bandwith of the memory path. The GTX550Ti has 192 lines , whereas the GT440 has 128 lines only.
I hope, this will answer some of your questions.

Kind regards and happy crunching
Martin

I think he was talking about the cpu units that can expand to using several cpu's at the same time, the n body units I mean. Some cpu units can expand to using every cpu you have allocated to Boinc, while others only use some of the cpu's allocated. It is a new thing for here and cpu units, using mroe then one cpu to crunch the sme unit. Kind of like Moo does when using multiple gpu's to crunch one unit.

I use my Nvidia 560Ti's here when I crunch here and I too always leave a cpu open for them. When using my AMD gpu's though I do not. The Nvidia cards seem to require more of the cpu then the AMD cards do, I do not know though if that means more work is being done. It could be designed that way from the beginning to be different, better use of the card etc.