We can even go so far to imagine a realistic scenerio where this was demonstrated. "Here's a scene rendered using 14 CUs. Now we add another 4 CUs to the rendering. Notice the improvement? No, we don't much either. So how's about we use those CUs to do non-graphics stuff like this...(effect showing something amazing like, I dunno, fluid dynamic volumetric smoke interacting with the environment (which of course should count as graphics!)). Better? We think so too. Of course, if you want to spend those extra 4 CUs on graphics, that's fine, but I hope you feel an incentive to use them for something else with a more dramatic impact on the game."
I consider that explanation as uniting all the evidence and soundbites, so it's the one I'm most comfortable with.
That's what I was saying as well. Ppl are way too touchy here on this issue. Ha! I wonder if they will need to use CU's for processing game audio. Sounds like their audio DSP only does compression/decompression. Maybe use CU's for synthesizing, mixing, filtering, etc instead of using the CPU for that stuff?
Sadly they weren't at all clear. He used loose language with numerous connotations (what does "round" mean? Certainly not the geometric arrangement of the hardware, making it completely unclear) and that's why people are reading it different ways. Cerny provided an ink-blot of understanding, and people are seeing it in different ways, which is to be expected.
I think he simply means that there are nuances to fully exploiting it that will take time to for devs to get used to working with.