I think he is comparing apples to oranges here. Using HTMl as the "rendering device" works on a different (higher) level of abstraction. The additional layer of abstraction in HTML has already solved the complexity of doing efficient partial redraws based on changes to a subset of a scene graph. This isn't something that the canvas API was meant to handle. If one were to use a scene graph (or display list) abstraction on top of canvas, it would shield the application developer from this complexity just as well as the HTML does now. However, the HTML abstraction, although more general, is implemented in native code while a canvas based scene-graph would need to go through java-script, so it is at a disadvantage as to how far it can be optimized. Basically, what he needs is SVG. :-)