In other words, the comparison is between the CPU redrawing a gradient 60x a second, and, an image being moved around on screen. Not entirely surprising that the one that does less rendering is going to be faster.
I suspect (but can't confirm without the source code to the demo) that shouldRasterize would actually hurt performance in one of the two cases in the article. Specifically, I imagine that it would make the score for adding and removing views from screen even worse than it already is.
Yes, it renders into a bitmap, and reduces CPU usage, once you've paid the upfront rendering cost. But, it doesn't rasterize opacity, and, will use up a ton of memory (because it doesn't result in stretchable images).