@Don and Igor, I think your conclusions were spot-on for the most part. This product is indeed not ready for market. Coming from the software arena, this is the hardware equivalent of what I call "code" instead of "product". Unlike Nvidia with the 580, AMD does not seem to have invested adequate time in developing their cooler solution. Unlike Nvidia with Cuda, AMD does not appear to have invested enough time (yet) with ISVs to develop applications leveraing their new hardware (yet). Good god the potential is there, but it just hasn't been productized yet. So "preview" is the perfect word to describe this. For the record, I own 3 discrete and 2 integrated ATi GPUs, and only 1 integrated NV GPU.
That said, I have a thought on why AMD shifted to the GCN architecture, and has now abandoned VLIW4:
With GPUs getting closer and closer to CPUs (indeed, they're now found on the same die for most mainstream systems), the potential for the CPU to leverage the GPU as a multi-vector FP processor is too big to ignore. So many of the emerging human interface technologies (touch, image processing, Kinect, speech rec, etc.) require lots of number crunching. So if you can hand that work off to a vector FPU sitting next to your ALU on your CPU's die, then having a low-frequency, modest-performance CPU doesn't matter as much. Suddenly, the processing capabilities of mobile devices become far greater, while the power consumption levels stay flat or even reduce.
So I think that AMD saw the computing landscape, realized - as Nvidia and others have - that mobile is the future, and figured that adding powerful graphics and number-crunching capability to a modest-performing mobile CPU core was going to be the best way to capitalize on the "fusion" concept and beat Intel in its own market. GCN is designed for exactly that; an entry-level GCN-based design might only have a few Compute Units, but that's all it would need to wipe the floor with any Intel-based GPU technology. I expect that we will see the x64 instruction set augmented soon to directly access the vector FP capabilities of on-die GPUs, just as we have seen such extensions as SSEx, etc. That - APUs with low-end CPU and GPU power - is what the GCN redesign was done for, not for high-end gaming GPUs. Achieving beastly high-end performance was just gravy. Of course, I also have to believe that AMD got tired of reading about supercomputers built with thousands of high-end Fermi cards, and would like to see their own in one of those headlines.
This technology represents some phenomenal work by AMD. Without committing to a full-on redesign, they re-factored just enough to fix some compute performance problems, while dramatically upping the ante in the gaming space. I look forward to the release, and will likely end up getting something at the 77x0 level.