I'm secretely hoping Sweeney is right and "larrabee like" are the way to go. Actually my secrety wish (a geek wish not a reasonable wish) is that the chip would even pass on texture sampling as Sweeney describe in his paper. Texturing latencies put a hell of a pressure on the design I'm confident (based on thin air ... treu... lol) that one could be able to fit a bunch of vectors on one chip, delivers crunching power in TFLOPS matching nowadays processors, while being possibly more power efficient. I actually believe the thing could turn pretty efficient at crunching numbers, the real question is... graphic performances...
I state it has a joke but I also state it as we have no serious reference. intel gave references (and based on simulation not the real deal) on how GeoW would perform on Larrabee that one thing, a proper reference would be Render GeoW using clever tricks Larrabee allows not running Geow on top of a huge software layer on larrabee. End of the rant it won't happen