That was confusing and badly written ("This was not the first time I get entrepreneurial ideas.") so I've read it several times and I still can't figure out what he's on about.

Is his idea that somebody might want to pay to use his three-CPU, inhomogeneous, linked-by-ethernet cluster of useless old machines?

Or does he want to expand it to a bigger inhomogeneous cluster of useless old machines, until he figures out that he's going to need a much bigger air conditioner (somewhere around the 16 CPU mark)?

Realistically the compute-density of anything but made-for-purpose kit at the moment makes clustering desktop PCs at any scale economically infeasible. We think in units of computational power per kilowatt (including cooling) and per tonne of floor loading nowadays.

The only play in the desktop space is hoovering up unused cycles on desktops that an organization happens to own anyway, primarily for interactive use.

Also, that FORTRAN is a strong presence in numerical computing is very, very common knowledge. Someone who learnt that yesterday is unlikely to have much experience with large-scale computation... Certainly not enough to pull off the democratization of supercomputing by himself.

hugh, you're right.

VCs claim to hate pitches where the founders don't know their market. I never understood why -- until i read this post. Not a single link to prior art, no explanation of why his solution will be better than anything else. no compelling problem to solve.

The author's blog would be better if the author focused on problem 2: "I require a lot of number crunching now and then. time is a premium." Tell us more! He might be onto something, but I can't figure it out.

The home built "cluster" is not something new. A lot of academics have been doing this for years, sometimes in their labs, sometimes at home. The small cluster is pretty much dead (as per Chris Dadginian). If I wanted to do a lot of number crunching, the way to go would be to get a dedicated 8-core machine and likely use some kind of accelerator (GPU or other). If I wanted more juice, it does get tricky. A serious number crunching cluster needs some fast interconnects, etc, which are not cheap.

The fact that he is not familiar with now much parallel code is written in Fortran is a little troubling. I wonder how much he knows about Infiniband, high performance storage, cooling, etc, the tricks that lead to real high teraflops --> sub-petaflop performance, cause if I am not getting that, then his offering has limited use.

This article was somewhat interesting, but I have to wonder about whether building your own cluster at home would be more cost-effective (considering paying for hardware replacement, electricity, etc.) that just using Amazon S3/EC2.

Also, if I was one of the author's prospective customers, I would be more likely to put my trust in Amazon than some guy with 3 computers in his basement.

Maybe he has a plan to come up with something like Engine Yard. Every great drawing starts with a dot.

2P3s and 1 P4 is not a supercomputer.

I bought a Quadcore PC with 4GB Ram for 600€ from Dell recently. A price tag that makes the time to glue your own rotting hardware together a very questionable investment.

If really interested in "Supercomputing", I suppose Cell processor (PS3) or GPUs are the way to go...

Also, energy consumption might be important. Probably the old CPUs don't really look so good in that respect.

I totally agree. He's trying to retrofit hardware that is ridiculously outdated for this purpose. And why bother when a few thousand dollars will get the average Joe a good size PS3 cluster, which has been proven to be effective for near-supercomputing applications?

I wonder if a GPU might be more suitable for his task. Even the 8800GTX is known to do single precision FFTs at more than 55 Gflops which is an order of magnitude more than even contemporary CPUs let alone P4.

One danger of the term "gigaflop" is how it is measured, and what you use to measure it. Also note that CPUs get a lot better when you start using SIMD code instead of scalar.

One classic example of the danger of the word "gigaflop" is that of the exhaustive motion search. If we define a single mathematical operation as a "flop" (technically an iop, since this is integer math), using Sequential Elimination, an optimized exhaustive search algorithm, an 8-core Core 2 system can crank out over 2.7 teraflop-equivalents of processing.

Sorry for replying a bit late. For FFTs, flops are measured in a standardized way. If you are doing an FFT of length N, the number of flops is counted as 5 N log N no matter how the actual FFT is computed. So, in the case of FFT, really you just specify a length N and measure the time.

For CPUs, the numbers using FFTW, one of the fastest FFT libraries that does take advantage of SIMD, the numbers usually do not exceed 5-6 gflops particularly for larger lengths.

OTOH, the above 55 gflops figure is also somewhat misleading since it does not include transfer time of data b/w RAM and GPU. Actual throughput is somewhere around 20gflops. On one particular project using FFT, I got around 15 gflops using GPU including transfer time while testing several FFT libraries, I never got above 3 gflops on a 2.4ghz quad-core using all four cores. The lenghts were big enough not to fit into cache thus reducing CPU performance considerably.

I've had the inclination to do this sort of thing, but the more hardware you have, the more likely, you'll have a component on some machine that will fail. As one person, you might end up spending all your time replacing or throwing away broken parts to keep the cluster up.

The other this is to pay for the electricity of a couple boxes running, regardless of how slow they are at a larger scale.

This sort of thing is good as a learning experience, imo, but if one's hoping to scale it up expect to run into lots of infrastructure problems one has to solve.

Supercomputers at home will be common soon. Intel is going the road of multi-core, I'm typing this on a machine with two 4 core CPUs. Soon we'll have not just 8 but 16 and then 64, etc, cores in a sub $2K machine you can buy from Dell or HP or who ever.

Well, for a few grand, you can get a TeraOPS development board with a 336 core chip from Ambric. The chip at the heart of it is only a couple hundred (quantity 1000).

Arguably supercomputers at home are common, and have been for years. How does your machine stack up against a Cray Y-MP?

(That Cray got 330 megaflops/core, Macs are peaking at around 1400 - each with 8).

The term "supercomputer" becomes meaningless unless its definition is periodically revised. IMO these days you need over 1 TFLOPS to quality.

You might also find this tutorial interesting

http://www.linux.com/articles/49654

If he just wants to make money from old computers he could install MythTV on it, add tuner cards, and sell a subscription to a remote MythTV box. People overseas might pay for it.

Or break the machines up and sell each part on ebay?

"who would want a sumpercomputer you might ask?"

I want a sumpercomputer!

Go for it.

P3s and P4s? Congratulations, you're killing the planet.