Just seen an article in Linux Format about the $25 Raspberry Pi

Wiki

In the opinion of the forum would these be any use for a Boinc or Folding "farm"?

The cost/performance ratio of the rasberry pi seems very low; I'd guess the ARM CPU produces about 1 GFLOP (the Eee PC with a 900 MHz CPU produces 1.3 GFLOP, so the ARM is probably comparable), and thus to get the same performance from rasberry pi as a quad-core CPU (40 GFLOPS) would take $1000, while the quad core would probably be $200-$250 at most.

It'd be much better to just go with conventional computer equipment, unless the price of the rasberry goes way down.

The rasberry GPU looks like it gets about 1 GFLOP/dollar of single-precision arithmetic, while most modern GPUs get 2 to 5 times that ratio. A 9500 GT for example costs roughy $40 and produces close to 90 GFLOPs of single-precision arthimetic. GTX 570s, 580s and 590s are closer to 3 - 4 GFLOP/dollar.

Looking at the sums on this thread from a flops/watt perspective, it's not so bad.

A RaspberryPi takes only a couple of watts; 40 of them would draw 80W and be eqivalent (in terms of flops) to a quad core PC.

Not too shoddy.

Ian

In order to run a project on the RP (Raspberry Pi) you will have to have a version of their science application which has been compiled for the CPU on the RP. Yes, it does run Linux but the Linux it runs has been compiled for ARM. So must be the science app compiled for ARM architecture. That means you are limited either to projects that are open source in which case you may have to compile the science app yourself or to the few projects that already compile for ARM (if there are any projects in that category, I don't think there are)

The other big concern is RAM. 256 MB isn't a lot of RAM and you have to fit the OS plus BOINC plus the science app into that or be content to run some of it out of virtual ram which can reduce performance dramatically. Not saying it can't be done, just saying it's something to think about before buying.

The $25 RP doesn't have ethernet, you need the $35 model for that and neither includes a case according to the article from the RP group. That article may be dated however. Anyway, add another $5 for a case. And you need to supply your own cables and power supply which also push the cost up.

The article from the RP people say the CPU performance is roughly equal to a 300MHz P2. I think they meant P3 but that's a minor detail. The point is the performance on the CPU alone will be very poor by BOINC standards.

Adding the compute power of the GPU into the equation, the article says the compute power is roughly equivalent to an Xbox 1. I don't know how usable that power is though, in other word can any BOINC projects use the GPU on the RP? Do they have a science app compiled for that GPU? I doubt they do and I doubt they ever will.

Now mass storage. It has to have a disk of some sort. A disk for each RP in your farm could add up to a lot of money. You could use 1 disk networked to all of them over ethernet but the card has only 1 ethernet port so you're going to at least need another ethernet port per RP if you go for that scheme.

Now consider heat related problems. Using an RP for the applications the RP designer intend to be run on it is one thing. Using it for crunching is quite another. Suppose you buy 40 of them. Where are you gonna put them all? The tendency would be to stack them up in one big pile but that would drastically reduce the heat dissipation of the boxes in the center of the pile to the point where they'll run too hot and die an early death or shutdown and refuse to run until you cool them better. Not saying it can't work, just saying one should think about it before purchasing.

Now power supply. How are you going to plug in 40 power supplies? That will require a custom power panel. If you can't build that yourself you'll have to pay somebody to do it. Or you can buy 1 big power supply capable of powering all your RPs but then you need to do custom wiring or pay somebody to do it for you.

If there is 1 law in computing it is that greater computing power and lower cost per unit of performance is achieved through VLSI (Very Large Scale Integration). 40 RPs would be an example of dis-integration. I don't think it can work the way the OP would like it to.