I don't think there is a rule written down anywhere, but hopefully this will help:
First, let's clear up some definitions. I think offscreen vs onscreen rendering is not the overriding concern most of the time, because offscreen rendering can be as fast as onscreen. The main issue is whether the rendering is done in hardware or software.
There is also very little practical difference between using layers and views. Views are just a thin wrapper around CALayer and they don't introduce a significant performance penalty most of the time. You can override the type of layer used by a view using the +layerClass method if you want to have a view backed by a CAShapeLayer or CATileLayer, etc.
Generally, on iOS, pixel effects and Quartz / Core Graphics drawing are not hardware accelerated, and most other things are.
The following things are not hardware accelerated, which means that they need to be done in software (offscreen):
Anything done in a drawRect. If your view has a drawRect, even an empty one, the drawing is not done in hardware, and there is a performance penalty.
Any layer with the shouldRasterize property set to YES.
Any layer with a mask or drop shadow.
Text (any kind, including UILabels, CATextLayers, Core Text, etc).
Any drawing you do yourself (either onscreen or offscreen) using a CGContext.
Most other things are hardware accelerated, so they are much faster. However, this may not mean what you think it does.
Any of the above types of drawing are slow compared to hardware accelerated drawing, however they don't necessarily slow down your app because they don't need to happen every frame. For example, drawing a drop shadow on a view is slow the first time, but after it is drawn it is cached, and is only redrawn if the view changes size or shape.
The same goes for rasterised views or views with a custom drawRect: the view typically isn't redrawn every frame, it is drawn once and then cached, so the performance after the view is first set up is no worse, unless the bounds change or you call setNeedsDisplay on it.
For good performance, the trick is to avoid using software drawing for views that change every frame. For example, if you need an animated vector shape you'll get better performance using CAShapeLayer or OpenGL than drawRect and Core Graphics. But if you draw a shape once and then don't need to change it, it won't make much difference.
Similarly, don't put a drop shadow on an animated view because it will slow down your frame rate. But a shadow on a view that doesn't change from frame to frame won't have much negative impact.
Another thing to watch out for is slowing down the view setup time. For example, suppose you have a page of text with drop shadows on all the text; this will take a very long time to draw initially since both the text and shadows all need to be rendered in software, but once drawn it will be fast. You will therefore want to set up this view in advance when your application loads, and keep a copy of it in memory so that the user doesn't have to wait ages for the view to display when it first appears on screen.
This is probably the reason for the apparent contradiction in the WWDC videos. For large, complex views that don't change every frame, drawing them once in software (after which they are cached and don't need to be redrawn) will yield better performance than having the hardware re-composite them every frame, even though it will be slower to draw the first time.
But for views that must be redrawn constantly, like table cells (the cells are recycled so they must be redrawn each time one cell scrolls offscreen and is re-used as it scrolls back onto the other side as a different row), software drawing may slow things down a lot.
