Basically, you place two cameras pointing in the same direction a short distance apart, like a person's left and right eyes, and render the scene from each camera.
There's a lot of details to get right. You need nonstandard projection matrices for the stereo cameras to make the images mesh properly when viewed on a 3D TV or similar device; you have to decide how far into the scene you want the focal plane to be, and where the user's screen is relative to the scene (for instance, does anything in the scene "pop out" of the screen, or is it all behind the screen); there are also issues with comfort zones, where you can cause people to experience headaches or strain while viewing the image if you deviate outside the comfort zone.
Here is a good set of presentations that go into these issues in a lot more detail: http://developer.nvidia.com/siggraph-2011-stereoscopy-course. Once you know the key terms, I'm sure you can find more on the Web as well.