I found this interesting development by Nvidia that uses a small add-in card to transform your desktop computer into a 500 gigaflop supercomputer for $1500…:eek:
The add-in card is the C870…
http://gizmodo.com/gadgets/spectacular-techno_porn/nvidias-tesla-hardware-supercomputers-fo-reals-270650.php
http://www.nvidia.com/object/tesla_computing_solutions.html
No, just the card is $1500. The server is $12,000 and a desktop version is $7,500.
and as I read it, that’s all you need
grimey is right.
The server is a tasty collection of 4 cards, and the desktop 2.
Just a note, this will only work for apps which can take advantage of having hundreds of small processors. Also you can only get single precision.
Does anyone know if blender is single or double precision?
It was more about the “super computer” thingy, that made me say that 
One such card doesn’t turn your system into a super computer. And further, they say that four of those server blades would give one enough processing power to qualify as a top 100 super computer, but at which end? 
It definitely looks awesome, but a super computer is far more than one or a bunch of Tesla cards / servers.
There’s no such thing as a single or double precision program (you can use both). But as far as I know, Blender uses mostly single precision data types.
But lots would have to be rewritten to make heavy use of multithreading, like said before. A dual or quad core is much more useful for the current versions of Blender.
You’re clearly the expert
No, I’m not. I’m just playing around a lot with the source code of Blender
(excuse me if I’m coming across as a “know-it-all”, it’s not meant like that just trying to be serious)
Many people do not seem to understand that The tesla platfrom does not run x86 code, and the C870 is NOT a CISC processor, and Cuda is a horrible language to write a full-blown program on. It is ok for certain very repetitive and highly parallelized calculations like monte-carlo simulations, but there are many things this $1500 beast can’t do, at least without the developer re-writing the entire code to take advantage of the highly parallel nature of the GPU.
for comparison: a quad core CPU (penryn) currently scores 100GFLOPS.