[Beowulf] El Reg: AMD reveals potent parallel processing breakthrough

[Lux, Jim (337C)]

On 5/11/13 2:06 AM, "Vincent Diepeveen" <diep at xs4all.nl> wrote:

>
>On May 11, 2013, at 2:17 AM, Lux, Jim (337C) wrote:
>
>> I agree with you..
>> But proving that you can build a HPC using whatever, be it PS/3s or
>> Furbys or nVidia cards, if you can't find people to program it,
>> that's a problem.
>>
>> (As someone who proposed building a cluster with Arduinos, I
>> recognize that I am eating my own dogfood now...  But I DO happen
>> to have 5 Arduino Uno Ethernets here in my desk drawer, and I have
>> a 5 port hub/switch as well.  The switch claims 1 Gbps, so I
>> probably won't have to worry about saturating its bisection
>> bandwidth. Or constructing some sort of toroidal fabric.  Funny, I
>> don't see an implementation of MPIduino anywhere though.)
>
>That's because if you produce a product it's cheaper to have your own
>board designed and you put on it the chip that's fast enough for your
>software.



I don't think a Arduino cluster is something you would build to do actual
computation. Just imagine.. All those little 18MHz CPUs with their 16 bit
integer CPU just merrily spinning away.

You'd do it to fool with cluster interconnect topologies, simple
parallelism, experimenting with fault tolerance when a link disappears,
and stuff like that. Particularly in an educational setting, where you
could fairly inexpensively set up 20 or 30 people with a 15-20 node
cluster.

 There's an inherent satisfaction to doing this kind of thing with
separate little boxes and cables and stuff, rather than just spinning up
instances on Vms in your laptop.

It's kind of like chem lab. Sure, you can simulate the reactions (with
errors) on the computer, but I think you lose a lot by not actually
handling the glassware and chemicals.




>
>That's far cheaper than using existing combination of boards.
>To give one example for the robot i'm building.
>
>A kind of tankchassis with robotarm on it, i want to drive a few DC
>motors and a bunch of servo's. as well as robot vision with a camera.
>
>Clustering 40 Mhz cpu's (or clocked something low like that) is
>pretty stupid then if alternative is for example a board from odroid
>or so that has a quad core ARM a9 on it.
>
>That's $89 for the U2 with a quad core 1.7Ghz chip and $69 for quad
>core 1.4Ghz
>
>A lot cheaper and faster practically for vision software than any of
>your r-pi/arduino cluster ideas :)
>
>http://www.hardkernel.com/renewal_2011/products/prdt_info.php?
>g_code=G135341359084


As noted above, one big fast computer beats many small computers
computationally, almost all the time.
(same is true in many other situations.. E.g. High power radio
transmitters.. One big tube is almost always better than many small tubes)


Nobody would cluster Furby's, Arduinos, Teensys, PICs, rPis, etc. for
computational reasons. You'd do it because you're curious and it's cheap,
or because it happens to fit a peculiar use case (e.g. A large LED video
wall with distributed processing, so you don't have to run separate wires
to each LED)