So I decided to poke around with Cycles and check out how preview works, using Blender 2.71
While I have GeForce 670 GTX with 2Gb of VRAM, and have set up GPU rendering in Blender’s settings, the rendering was super slow 
Do I need to install CUDA Tool kit for Cycles to utilize my GPU ?
I don’t think so. Have you tried increasing the tilesize? as a general rule: rendering on CPU = low tile size (16x16…), rendering on GPU = high tile size (128x128, 256x256, even 512x512, depending on the hardware.
AFAIK it’s only necessary to install CUDA Toolkit if you’re card is not initially supported by blender and you have to compile a render kernel on the fly (this feature is available under linux, don’t know if this is even possible under windows).
You might need to set in the Performance tab in Render settings the tile parts to match your GPU - maybe set it to 200x200 or 256x256 instead of 64x64. The default works best with CPU. If your User Prefs shows your GPU then it should be just about switching in Render Settings to GPU, and setting your preview samples to taste. I have never had to install anything other than Blender to get it to work besides the Nvidia driver updates as they come along.
Aye, I’ll check that, thanks.
Btw, does AO bog down performance?
No, not in viewport at low sample range. Preview at 20-50 on GPU, but set render to higher like 100-200 for test renders. LordOdin has been experimenting with Branched Path tracing and he can get some really good times for final render with that. I am experimenting with baking out the AO and the cube mapped image textures for bases to paint with, kinda fun.