....and some more info. If you are into online gaming, such as World of Tanks, you MUST set the GPU preferences to make sure no processing whilst active on that, it totally screws the game. You can leave the CPU to run all the time, but GPU has to be suspended whilst gaming.
Normal browsing, word processing etc is unaffected so GPU can be left running. Setting the prefs for GPU to stop whilst active is overkill, it only needs to be suspended for 'certain' apps, so it is a little more bothersome to control, but worth it!

1. It usually at least 6 time faster than CPU.
2. You need to have a graphics card that will run GPU tasks.

There are projects that use considerably more than .05%. POEM uses upwards of 90% of 1 CPU core.

I think the ATI/AMD GPU project applications that use tiny amounts of CPU are CAL or OpenCL wrapper type using an efficient IL coded core and are not native OpenCL type. My limited understanding of this is that native OpenCL uses just in time compiling and is somewhat inefficient compared to code optimised at the IL level. Even with applications that can be made to run in a highly parallel fashion that suits GPU processing, native OpenCL applications will tend to use increasing amounts of CPU as GPU load percentage increases.

Even the most efficient low CPU usage GPU applications can be slowed noticeably if all CPU cores are fully loaded with high intensity CPU tasks such as Einstein. To expect to be able to run a reasonably high GPU load native OpenCL GPU application without allocating CPU resources to support it is just a wish for the future at this stage in OpenCL development.

This may initially be an unpleasant concept for those new to GPU processing and a change they do not appreciate for those used to low CPU usage GPU applications. Better to think of it with a positive mindset rather than a negative one. Rather than being disappointed if a CPU core is used that could be used for other CPU applications, instead rejoice that one CPU core plus a GPU is doing the same work that 6 or more CPU cores can do. HD 7970 will theoretically do more work than 11 CPU cores if it is not overly throttled. Keep in mind the large increase in the amount of scientific research performed that GPU processing can deliver. After all that what's it's all about eventually, doing more scientific research more efficiently.

GPU applications are often a compromise between stability/usability for those who do not wish to concern themselves with tuning the application and performance for those who are keen to maximise their contribution and use electricity and their donated resources most efficiently.

Those new to GPU processing are sometimes surprised and disappointed with the compromises involved. They think GPU applications should run at maximum efficiency with little or no interference with other CPU or video applications like many project CPU applications do. However the graphics card has to draw the screen at the same time and also interact in varying amounts with the CPU, so expecting no compromises when running BOINC GPU applications is possibly similar to expecting a car to perform and handle normally when it is towing a heavy trailer.

Highly efficient GPU applications are not without problems anyway. On single GPU boxes that are not dedicated BOINC computers, a very efficient GPU application will cause screen lag. In the worst cases this makes the computer unusable and the GPU application needs to be slowed down either with adjustable project preferences or via <cmdline> parameters in an app_info.xml file.

Also overheating of voltage regulators is quite possible especially in hot climates or with poorly ventilated enclosures. If you run GPUs flat out with a high load GPU application, VRM temp monitoring is recommended if you don't wish to risk frying your card.

On top of all this, the efficiency comes at a price because high efficiency GPU applications draw a lot of power. It's good to be able to do lots more science with GPUs but not so good when a large electricity bill arrives. I've had to cut down from three GPUs to one, now that the tariff here is AU$0.32 per kWh. For keen full time GPU crunchers in high tariff countries/areas it is often false economy to buy or continue to use an older generation GPU that is less power efficient than the latest models. Not forgetting, less power, less heat; less heat, less bedroom sauna.

i see in the device profile that there is a gpu selection option to check if you want to crunch with the gpu.
Sek, from reading your posts, i gather we shouldnt have to adjust the cores to crunch since the wu will share a core and gpu. is this correct?
For those using ATI cards, you can use catalyst Control Center to adjust the card settings. If your running with overdrive on, you can lower your settings to cool down your card some.