I actually like the idea of large chips, anybody remember the Alpha? But conceptually can be different then practice.

There are a few practical issues I think they should deal with in a consumer version. One such issue would be the factor of die size at the fabrication level. If they could make each of those GPC units fit on a single small chip. Then you could MCM 4 on a package to make a 512 shader card, 3 for a 384 shader card, 2 for a 256 shader card, etc. Ultimately I feel that's where this will lead, IMO. AMD will do it with CF and NV will probably do it with SLI (or CUDA). Intel will do it with a version of x86 and as AMD will do it with their processors. And lots of software will have to be rewritten to take advantage of the era of massively multicore, so could we use some open standards please?