The programming model for using GPU's for scientific computing is actually pretty well thought out and takes into account the fact that you need to get the data where it needs to be fast. It also fits pretty nicely into the fact that applications are mainly x86 based in this area.jonathan wrote: Have you programmed a GeForce 8800?
These all sound like arguments for why you shouldn't bother building an XMP-64 in the first place.
Don't get too bothered by who would pay for it... (you could just to it free of charge). Instead think about whether it would be a useful, interesting demo that might attract attention to the technology, the company etc.
I can think of many interesting things to do on this that would be a pain in the ass on the GeForce. It's not all about peak GIPS...
You are are correct about it not being about peak GIPS ..... in this area its all about peak GFLOPS.