A supercomputer for £100

[a.lone.wolf]

Would it be possible to link a couple of boards up together and use them as a supercomputer, sorry if I am completely wrong and it is imposable but I couldn't stop my self asking?

[jamesh]

This has been discussed a bit on the fourms, but the Arm at only 700Mhz doesnt give supercomputer levels of performance, even if you use a lot of them.

What a cluster of these devices would do is give people a cheaper platform for trying out various schemes for distributed computing.

That said, if we could get a way of harnessing the GPU performance then you are talking some serious horsepower. They can give about 25GFlops. Which is a awful lot considering the low power consumption. I'm slowly looking in to it.

[pvgb]

Depends what you call a supercomputer ! ( http://www.top500.org/ )
Most of the big machines are now composed of a massive number of processors. Each of the individual processors are probably more computationally able than a RPi.
Four boards together probably have less computational "grunt" than a current PC. ( I am assuming the use of boards with Ethernet )

However, having four boards will allow you to try out the techniques for cluster computing, fit neatly into a small enclosure ( or possibly even a tower case with a PC in it already - probably want a host node anyway ) with a small outlay and good power performance.
The software should be easy enough as well - one option is a Beowulf cluster.

[reiuyi]

Imagine a single standard midi-tower computer case with a 20A 5v supply, a good switch and about 25 raspberry pis. You'd have your very own cluster-computer to teach advanced programming on multi-computer level. It'd still be cheaper than a regular computer, yet it has 25 CPUs and 25 GPUs !

[Piw32]

Ziilabs (Creative) will do it with their ZMS-40 52 Gflops quad core A9 chip  (OpenCL compatible).There was a rendering of stacked boards on their website.

http://www.ziilabs.com/news/re.....10504.aspx

[DavidS]

Just thinking out loud (on topic):

Would not it be possible to use a Raspberry Pi as a terminal and four or 5 home made boards with 16 ARMs each (the chips are only about $5USD each) for the brain.  Then you could use a later generation ARM, and there is no need for anything other than the PCBs some pin headers, a handful of SMT Capacitors and Resistors, some careful design, and the SoCs (because the Pi provides the IO).

[Sylvain]

DeliciousRaspberryCake said:


Imagine a single standard midi-tower computer case with a 20A 5v supply, a good switch and about 25 raspberry pis. You'd have your very own cluster-computer to teach advanced programming on multi-computer level. It'd still be cheaper than a regular computer, yet it has 25 CPUs and 25 GPUs !



25 Rpi model B (for network) costs 875 $. Sorry but it is much more than a "regular" (linux) computer.

For this price, you could have a 6/8 cores (AMD) computer with tons of RAM, a powerfull cuda GPU and all you need around that (case, PS ...). More interestingly, you could also have something like 4 dual core entry level systems (H61 + Celeron G530 + 2 Go of RAM).

For teaching distibuted computing, it can be interesting, for real applications it's a none sens, in terms of cumputing power and also most of the time in terms of power consumption !

Sylvain.

PS : For James, you are speaking of 25 GFlops for the GPU, is it comparable with for exemple the numbers given by NVidia for their graphics cards.

i.e. could we say that a 35$ GT520 card (155 GFlops according to Nvidia) is theorically 6 times more "powerfull" than the GPU of the Rpi ?

[jamesh]

Comparing a desktop graphics card with a mobile PGU is a bit unfair TBH. The desktop devices have no real requirement for power saving, so can be quite profligate with the GFlops vs Watts ratio!

The Videocore on the SoC uses < 1watt I think at full power, would be interested to know what the NVidia devices pulls.

But yes, that $35 cards is theoretically 6x more power, but does cost 6x as much and uses a lot more power than 6 Raspis.

[Sylvain]

Thanks for your quick reply James.

I knew my comparison was unfair in many ways but I also think this kind of numbers are interesting to know.

Sylvain.

[reiuyi]

Sylvain said:


25 Rpi model B (for network) costs 875 $. Sorry but it is much more than a "regular" (linux) computer.

For this price, you could have a 6/8 cores (AMD) computer with tons of RAM, a powerfull cuda GPU and all you need around that (case, PS …). More interestingly, you could also have something like 4 dual core entry level systems (H61 + Celeron G530 + 2 Go of RAM).

For teaching distibuted computing, it can be interesting, for real applications it's a none sens, in terms of computing power and also most of the time in terms of power consumption !


If you're strictly speaking about costs, Raspberry pi isn't really high on the GFlops/euro scale. It is a device made for teaching, so it doesn't need to be high on any scale apart from the pricetag. Also, I'm quite convinced 4 desktop computers have so much electricity costs; it'll end up being more expensive than a whole cluster of raspberry pis at the end of the year. I'd expect raspi to live for at least a decade before dying (high MTBF, at least 100kH), while regular desktop computers are only rated for 3 years.

OP did not specify what for his plans were. If it is strictly for teaching grid computing; I'd use €1000 to build a 25~40-computer-cluster with raspis. If OP was planning on folding proteins,  predicting weather patterns or whatever else may require extremely high GFlops; I'd just use the €1000 to buy an i7 or two because its GFlops/euro is enormous.

The mod mentioned that they designed a system with a very low energy-demand in mind (for whatever reason they did). Comparing power-saving systems to radiator-PCs is pure nonsense. A traditional supercomputer can heat up an entire cityblock, while a raspberry "bramble" uses less energy than my desklight

[gimp]

I agree with DeliciousRaspberryCake. This would be an excellent approach to teaching distributed computing on a small scale with small costs, but it would not do well at all to use any current SoC for a supercomputer. For that, we tend to stick with high-end x86 chips, or Itaniums, Sparcs, or Powers, depending on who's making it and what it's being used for.

[pvgb]

My understanding is that one of the primary objectives of the RPi is to get people ( especially youngsters ) back into the creation rather than consumption of software.

One of the skills needed for development of the next wave of systems is that of being able to program for the network. Many constructs already act across a number of machines ( web server connecting to database server for example )

How much resource would you need to be able to do things like this ? Even using cheap beige boxes you have to have physical space and power connectors.

The idea of being able to have even a couple of RPi boards for each learner to get away from the shared memory paradigm and start learning how to harness multiple machines is very attractive – and extends the basic objective.

[jamesh]

The teaching of distributed computer is something I think the Raspi should be pretty good at. There is another thread going on at the moment about it, search for stacked PI, or Brambles.

The reason this SoC is so low power is that it was designed for mobile devices running on batteries. Using tit means an overall low power device by default.

[Mushroom_Lord]

I am very much into the idea of somehow linking two Pi's together and using them as one computer. I suppose I will look into how it is done with other current computers....

(My Graphics card was considered budget when It was released 8 and a half years ago...)

I know there aren't going to be games, but it would make things sexy smooth

Also, for the RAM upgrade, as 256mb is definitely adequate for the machine to run, but the idea of 512 gives me the thought of even more flexibility. (now I'm sounding like one of those people with 24gb of RAM) - Assuming it would harness the additional RAM, that is.

[Digital-Wangateur]

RPi HA is a great idea for several reasons as apposed to just super computing … think  high-speed redundant storage, networking, data processing and acquisition etc. If this becomes a project count me in!!!!!

[mental2k]

Mushroom_Lord said:


I am very much into the idea of somehow linking two Pi's together and using them as one computer. I suppose I will look into how it is done with other current computers....

(My Graphics card was considered budget when It was released 8 and a half years ago...)

I know there aren't going to be games, but it would make things sexy smooth

Also, for the RAM upgrade, as 256mb is definitely adequate for the machine to run, but the idea of 512 gives me the thought of even more flexibility. (now I'm sounding like one of those people with 24gb of RAM) - Assuming it would harness the additional RAM, that is.


No games?  Arse to that feature=player_embedded

As a programmable embedded device 256mb is huge! The original xbox had 64mb and the 360 has only 512!  I have a feeling you'll find the Ras-Pi more than sexy smooth enough.  (Or be completely disappointed its not an i7 desktop)

[spritrig]

I think doing some distributed computing with Raspberry ? is a great idea!  Really, it is all about playing and learning.  Think about how much it would cost to cluster 5 PC"s together.  What kid can afford that?  However, if you get a classroom of kids, each one with a Raspbery pi, all the kids can learn to cluster.
The reality is that distributed computing could work quite well for certain applications, if the network is not the bottle neck.  It would work well for applications where the processor is the bottle neck rather than the networking.  That would be applications where a lot of computation is needed to produce the data that gets shared on the network. 
I'd love to have a bunch to cluster and try out.  It would be a great model for learning how to cluster cloud instances without having to pay bucks an hour rent.  If Rpi is really $25, a cluster of five would have a break even cost to renting 5 small amazon EC2 instances for ~125 hours.  The R?"s would last much longer than 125 hours!  
I think R? is a great alternative to Amazon EC2 cloud service.  EC2 offers economic instance scalability through rental.  R? can offer economic instance scalability by purchase.  

[n3tw0rk5]

I'm looking to treat the RasPi as blades and build a small chassis, possibly out of a shuttle case. (i'm kicking myself as i recently binned 2 of these as i thought i'd never use them)

[plugwash]

The thing is it's going to take a heck of a lot of pis to give you computing power equivilent to say one i5-2500. If we assume the performance per clock is the same (and I suspect the i5 has a higher performance per clock) then the i5 would be equivilent to 17 pis

Clustering pis may be useful for learning about clusters but it isn't likely to be economical as a way of getting processing done.

[spritrig]

£100 for a super computer isn't realistic unless you use an antique performance measure.

There may be applications that a cluster of Rpis may actually perform better than a single i5 computer.  Maybe you have a web data mining application that could benefit from 20 ethernet ports.  Let's say you want to experiment with such a configuration for such an application.  20 Rpi may be more economical than an i5.

[shaurz]

To make a supercomputer you need a really fast, low-latency interconnect like Infiniband.

[jamesh]

plugwash said:


The thing is it's going to take a heck of a lot of pis to give you computing power equivilent to say one i5-2500. If we assume the performance per clock is the same (and I suspect the i5 has a higher performance per clock) then the i5 would be equivilent to 17 pis

Clustering pis may be useful for learning about clusters but it isn't likely to be economical as a way of getting processing done.


And I think that is the point of the Raspi - it's for teaching, not hardcore work.

[Svartalf]

JamesH said:


This has been discussed a bit on the fourms, but the Arm at only 700Mhz doesnt give supercomputer levels of performance, even if you use a lot of them.

What a cluster of these devices would do is give people a cheaper platform for trying out various schemes for distributed computing.

That said, if we could get a way of harnessing the GPU performance then you are talking some serious horsepower. They can give about 25GFlops. Which is a awful lot considering the low power consumption. I"m slowly looking in to it.


Well, we should be able to do an MPI type cluster with each of the tasks distributed to the nodes having GPU compute kernels written for OpenGL ES 2.x.  Without OpenCL, it'll be slightly limiting, but you should be able to eke out at least a handful of tasks where you could see something impressive with a Bramble.  Without having some sort of migration scheme for the GPU side of things, you're not going to see much performance with a Shared System Image design.

[mateli]

Digital-Wangateur said:


RPi HA is a great idea for several reasons as apposed to just super computing … think  high-speed redundant storage, networking, data processing and acquisition etc. If this becomes a project count me in!!!!!



Amen to that. If you look at the purchase price then rpi is more expensive in price/performance then most other solutions.

But if you look at energy consumption then price/performance is very low, and in the long run computers energy consumption cost A LOT more then their purchase price.

In many cases performance is not even important. Often you cluster to get high availability and then price/node is important, not price/performance.

As a side-note i have been talking to Tilera about using Tile in servers, which would be a superior solution for. Tilera is a massive multicore CPU with up to 100 cores and is far more impressive to me then opencl-gpu:s.

If ordering 10 000 units or more we could get reasonable prices for TILEencore, their current server platform.

I think it would be an excellent combination to the rapspi. A minimalistic client that can be coupled with a server that can serve an entire school with raw power. It would turn the rapsi into a very powerful workstation to a fraction of the cost, due to the economy of scale. As power is located in a single server the cost/node becomes extreamly low.

[diegod]

Mushroom_Lord said:


I am very much into the idea of somehow linking two Pi's together and using them as one computer. I suppose I will look into how it is done with other current computers....

(My Graphics card was considered budget when It was released 8 and a half years ago...)

I know there aren't going to be games, but it would make things sexy smooth

Also, for the RAM upgrade, as 256mb is definitely adequate for the machine to run, but the idea of 512 gives me the thought of even more flexibility. (now I'm sounding like one of those people with 24gb of RAM) - Assuming it would harness the additional RAM, that is.



This is more of the idea I have.

I have no idea of how to do it, but I would love to make a cheap tablet but with 2 RPis, the space will be there in 10.1 in display.

Now, from what I gather from the thread is that the way to link them would be with the ethernet, that is bot bad, does anybody have an idea of how it could be done (the linking them and having them work as one computer) the screen and all that is not so hard to figure out, much less storage... but the linking them is the hardest part. Any ideas???