[jwilliams]

I'd be curious as to the underlying why here.

I imagine that CSS parsers aren't particularly designed with any kind of parallel operation, whereas grabbing images (& decoding them) is largely done in parallel (up to the number of max connections). So while you're parsing the CSS you can be getting the images, offsetting the connection cost.

In the CSS case, the device needs to get the (slightly larger) CSS, un-gzip it (with more complex tables), Base64 decode - and then decode the image as before. I wouldn't be surprised if this is a completely sequential activity with the rest of the CSS parsing.