GPU for computing

Really surprised while reading an email this morning, moving computation task on GPU instead of using CPU. Last year I was really surprised knowing graphic cards memory usage for virtual memory, I imagined how fast it would be. And today they moved further, move computational task on parallel GPU.
You can see there are four graphics cards installed, with 2 GPU each, so it has 8 GPU. Quite interesting technique, considering GPU has more processor register than CPU, since many years ago there are GPU with 128 bit processors, meanwhile CPU has just moved to 64 bit. GPU is not a multi purpose processor, so it is easier to design a sophisticated one.
Thanks for PCIexpress technology that makes multiple graphic cards installation possible. In AGP age it was truly impossible since AGP is only allowed single card installed.
Eventhough they said it is an affordable supercomputer, it is quite expensive, $6000, but in the next two years eventually it will be affordable. Would it be a big threat for sgi? No, I don't think so, SGI still provide far more reliable devices. Anyhow this Fastra computer was build upon consumer class devices, but maybe it is suitable for particular works that do not require strict reliability feature.

Further information: http://fastra.ua.ac.be
Picture is courtesy of the owner.