From reading, the latest gpugrid should use very little CPU resource. There are quite a few aficionados around to chip in their experience.

my experience with GPU use.

since I don't do GPUGrid, I can't tell about them. My main machine does both WCG and F@H (Stanford's Folding at home) and the CPU use of the F@H executable (FAH_core11.exe) is below 1%, lost in the background, everything else being used by WCG. F@H does not use boinc and the 2 applications work completely independently and in perfect harmony.

F@H runs smoothly, except when watching HD video (motion is a bit jerky) or 3D apps (ok, that means games), but the F@H application can be paused during that time or exited.

If you want to run GPU calculations on your machine, you should be cautious about the temperature of your GPU. Reaching above 80C (175F) might not be a big problem for a gaming machine, but should really be avoided for long term use.

I have a NVidia GTX260 card (that came with the fan set at only 40%) and use a fan control profile in the NVidia driver system tools (V6.05+) with the following setpoints : 30% @ 40C (104F), 67% at 67C(150F), 100% at 75C(167F). Riva Tuner also does the job. The slightly overclocked card (650/1100/1400Mhz) runs between 62C and 65C.

Before using your GPU card for calculations, you must make sure it's extremely stable. This is done by overclocking it a bit more than you intend for everyday use (I set mine to 680/1150/1450 for the test) and testing it with OCCT (V3.1+) for at least one hour with error detection. Do not attempt at stock fan speed ! During the whole process the temperature can be monitored in OCCT, or using GPU-Z. The tests are optimized for maximum power consumption, and your machine will be much cooler and less noisy while running the real application.

The same advice is good for the OCCT CPU test that should be performed before running WCG to make sure your machine is stable and will return accurate results.

Since your machine will have both the CPU (With WCG) and the GPU (With F@H or GPUGrid) heavily loaded, you want to make sure your power supply is correctly sized. Low end GPUs will be ok with a 450W power supply, higher end ones require at least a 550W supply, and more if you want multiple GPUs...

If you want to buy a new power supply, make sure it's power efficiency is certified '80Plus' or '80Plus Bronze' (Silver & gold are too expensive now), else, you'll just waste power and more money than the cost of the power supply. Buying a power supply rated twice the real consumption ensure it's running quietly and at it's maximum efficiency.

My E8400 (dual core OC'ed 3.8Ghz) machine uses 125W while idle, 190W while running WCG and 260W while running both WCG and F@H. OCCT power supply test (both CPU & GPU test simultaneously) peaks at 370W. Power is measured with a "Kill-A-Watt" meter.

Given the inefficiency of running F@H with the CPU client (average of 1.07 X86 GFLOPS/CPU) versus 200 X86 GFLOPS/GPU and 59 X86 GFLOPS/Playstation, it would be nice if people without GPU would use their machine for WCG while people from WCG with a GPU would run both. That way, both projects would benefit.

The F@H statistics don't mean GPUs run 200x faster than a CPU, as the CPU mix include older machines, and many of those machines are not run around the clock. Still, the contribution of GPU enabled machines is impressive.

Here are a few links :
F@H main page :
http://folding.stanford.edu/English/Main
F@H GPU client download (for both ATI & NVIDIA)
http://folding.stanford.edu/English/DownloadWinOther
F@H contribution of CPU and GPU clients :
http://fah-web.stanford.edu/cgi-bin/main.py?qtype=osstats
NVidia system tools V6.05
http://www.nvidia.com/object/nvidia_system_tools_6.05.html
OCCT CPU & GPU stability tester :
http://www.ocbase.com/perestroika/index.php?Download
CPU-Z
http://www.cpuid.com/cpuz.php
GPU-Z
http://www.techpowerup.com/gpuz/
The US 80Plus energy efficiency program
http://www.80plus.org/manu/psu/psu_join.aspx
A good, efficient & quiet power supply (Antec TP-650) :
http://www.newegg.com/Product/Product.aspx?Item=N82E16817371021

About double precision capable GPUS : They are nice to have and a wise choice when buying a new video card, but not required for either GPUGrid of F@H at this time.
ATI 3690, 3830, 3850, 3870 and 4830 are also able to perform double precision calculations.
http://developer.amd.com/gpu/ATIStreamSDK/pag...amSystemRequirements.aspx

Although the above Nvidia GPUs are double precision capable, they aren't as double precision capable as you might like. The GT200 architecture only has 1 double precision processor per Streaming Multiprocessor, which basically means that the double precision processing power of GT200 GPUs is roughly an eighth of the single precision crunching power. Still, it's better than nothing.

I'm not sure if something similar is true of the ATI/AMD GPUs, but it wouldn't surprise me.