That is pretty amazing, but did you check the power when you were using the GPU as well as the CPU? Granted, a CPU takes the most power of all the components, but modern GPUs are getting up there, too.
Nick makes an excellent point. I considered this and passed on intentionally saturating the GPU in my tests. The expectation of power efficiency vanishes when you slam the GPU. The only metric that users consider relevant at that point is frame rate. No review I've read of a PC gaming platform brags about how little power it consumes. If anything, it's the opposite!
In a commercial setting, the likelihood of encountering high, continuous GPU load is practically nil. But in a consumer setting, it'll be commonplace. I take your point and will act on it. I'm off to terrorize a GPU.
As an aside, Apple tells me that the CPU does most of the heavy lifting in Tiger x86's visuals. They said I'd be surprised to learn how little Tiger relies on GPU offloading for performance. I've got to dig into that, because I posited at the Intel transition announcement that the GPU would be used to compensate for Pentium-M's (at the time, I thought it would be Pentium 4) "see me after class" math scores. But then, Apple had Intel's peerless compilers, performance primitives and math libraries in its back pocket. I expect that these alone could cut trips to the GPU in half compared to gcc.