Nicol Bolas explains why you can't expect pixel-accurate reproductions, but that doesn't mean you can't test with screenshots.
Don't compare the screenshots byte-for-byte. Instead, use a perceptual diff to quantify how different the renders are. You can then set a threshold in your test and have the test fail if the renders are too different. A harsh threshold will let you filter out the rounding differences between platforms, and the slight differences around object silhouettes, but still detect objects in the wrong place or inconsistent texturing. A more forgiving threshold will allow bigger differences in texturing (or even with different shader versions or extensions, if you have different rendering paths for different hardware), and only complain if you have gross differences such as missing lights or a completely different scene being rendered.
The pdiff tool is commonly used for this task, and I've set up test systems based on it before.