That city graphic is really precise.
(edit: it looks like the 64 bit math is used only for the map projection part of the shader... so it's not quite as ridiculous as it might seem, my haswell system is limited by CPU, not by vertex shader performance)
You got some of it, pretty is part of the point here. You can do nicer looking things with WebGL than with an HTML canvas.
High performance and interactivity is another side of what to look for here. Spend some time rotating and zooming these examples to see the kinds of interactions you don't get in other data visualizations. You can (potentially) visualize a lot more data at interactive rates with a WebGL view than you can with other techniques.
The laying/compositing tech is a third piece of what is interesting here, allowing different data slices to be rendered and included separately. This is useful as it's common to put everything in single WebGL view for ease of implementation initially, but can get complicated and may not scale as well as this layering approach.
Generally it seems to me that whenever something involves WebGL, performance and UX goes out the window.
Hardly interesting from a data viz perspective, but I guess good on them for trying to push the tech envelope?
