We're certainly planning to investigate it, particularly with the advent of vector units that reduce the latency of data transfers, such as AMD's new Opterons and Intel's Knight's Landing. The challenge here is that while several of the stages (e.g. CSS selector matching) can be trivially sped up on a GPU, the CPU/GPU latency cost is going to be close to the original CPU evaluation time.

There are some very sequential and unfortunately common "corner cases" in layout (e.g., uncleared floats) that have lead us to currently prefer the higher clock-speed CPU for parallelizing phases. Even if we find a great way to work around floats, it's likely there will still be a lot of CPU/GPU chatter, which makes it difficult to use today's GPU cards when you're trying to keep total page load well under 200ms and incremental reflows < 15ms.