Actually, I would consider a multi GPU farm as some sort of super computer.

I believe the typical high end GPU does around 2.5 tera FLOPS.

UC Berkeley actually has an effort which solicits PCs and higher to do distributed computation using what we would consider to be a multi GPU PC using their own ""mining"" software.

The effort is called BOINC and supplies computational horsepower to universities and startups to work on protein folding, black holes etc etc.