My first brush with "supercomputers" was when I read an article about Beowulf computing, and was amazed that some people at CalTech were able to put together a supercomputer from donations of old computers.

Initially, they used it to calculate 10 different weather variables per square mile across the whole of the US in real time. I believe they also used it to compute the gravitational effects of a hundred million stars on each other.

This first supercomputer had a speed of about 2 Gigaflops, a flop being a single floating point operation, so for instance, adding two decimals is a flop. To multiply you would use bit shifting and adding, so to multiply by 17 you would bitshift four places (equivalent to multiplying by 16, we're talking binary here) and then add once, for a total of two flops.

Nowadays, an ordinary cpu will do 3 Gflops on its own. However most of a CPU's work is figuring out where in memory everything is stored, which limits their ability to do high end computing.

GPUs, the chips used to do the graphics on ordinary computers are not hampered by this, so their full power can be used to do computations, and they have become far more powerful than their CPU brethren as a result.

With modern PCI-express motherboards it is possible to connect four ofthese together, with the CPU handling disk i/o, and allocating thevarious parts of a computing problem between the GPUs and because they are connected directly to the motherboard, there isn't the same delay as that experienced by the Beowulf cluster, which did most of its message passing through Ethernet cables.

Just recently a group of students at the University of Antwerp in Belgium used exactly this configuration to put together a supercomputer called the FASTRA using off the shelf graphics cards (4 NVidia 3800x CHECK THIS) to do the grunt work and although they were very good with providing the specification sheet of the computer and even put up a video on YouTube to inform people of what they were doing (high end tomography from PET scans), there were precious few details on how they put it together, what problems they came up against along the way, or how they got all the various programs working together.

Since then, and we are only talking less than six months ago when I read the article, ATI have brought out a graphics card that is clocked at just under a teraflop (a million million operations per second), NVidia have brought out CUDA 1.2 (a program to help them all work together), and DirectX11 (a graphics rendering engine) has been brought out, which has functions aimed at easing number crunching on this type of setup. No doubt, by the time I get to actually doing this, NVidia will have brought out an equivalent or better card, and with optimisations, I believe 3-4 Teraflops from this one box, is possible. To put that in context, the fastest supercomputer in the world was clocked at 1 Petaflop.

So what I'd like to do is get the components, make a video of putting it together, and document and blog about lessons learnt in getting it to a point where I can measure its speed, and any optimisations done on the way. I have a fast synchronous connection, so it would be ideally suited to using it for Folding@Home, or any other use where something this powerful would be useful. So if you think this would be interesting for you, just click the Paypal link to the bottom right. Thank you.