There's already a lot of knowledge and technology around using the GPU for high performance 3D rendering. Doing it in the browser (with WebGL) isn't that different. The breakthrough of WebRender is (mostly) to efficiently use the GPU for 2D path rendering.
It's actually not so much 2D path rendering as CSS. NVIDIA made the breakthrough in 2D path rendering with NV_path_rendering (which is a great paper, and I cite it in the talk). The way I see it, WebRender takes NVIDIA's work a step further by observing that you don't actually need 2D path rendering to render most CSS: you just need a good blitter, a good box clipper, and some specialized shaders.
Speaking of your specialized shaders: how do you do the rounded rectangle corners? There was a quick mention about it in the talk but no details. Some alpha blending tricks or coverage sample masking (gl_SampleMask)?
Dead simple: we use a shader to draw corners to an atlas and then alpha mask when we go to paint the page. Doing this allows us to avoid state changes when many border corners are involved.
