clusters of rpi4 with clusterHAT holding 4 pi zero ?

[phi6ndepa]

i just want an affordable supercomputer to run mathematica and learn and do science/math/geometry/physics/coding

and i have wondered how many raspberry pi zero i can put together but i don’t know if there is a limit;

but also, i don't have a-lot of money; and my dreams of putting 4 raspberry pi 4 in parallel were shattered when i realized that one isn’t faster than an iphone; and 4 could maybe match or fall short ? maybe surpass by a little ?

But, not a true number cruncher; and it would be like having mathematica on a laptop (i’m hoping to get it for my desktop eventually, but raspberry pi has other capabilities seemingly)


So i was just wondering if i could stack the 4 rpi0 on a rpi4 and then link those type of things with each other; semi-affordabily

or bare bones; if i can find and buy enough bare bones rpi0 and link a ton of them together and just buy more every month, until i could calculate physics stuff when i am educated enough to model black holes or etc (like the ps4 supercomputer, but i won’t have that kind of money)

so please let me know
(and if anyone can answer my troubleshooting post (not the one i asked for deletion), in my user posts, in my profile; that would help too)

thanks

[Daniel Gessel]

I’d recommend focusing on the education first, learning how to write simulations of black holes and other stuff you’re interested in. A pile of pi zeros won’t be able to work well together and your education will get you access to the computing power you need.

In the meantime, you could start to learn how to write simulations of all sorts on a Pi 4. Learn to use all 4 cores and the gpu. There are enough interesting ideas there to keep busy.

[W. H. Heydt]

A cluster of Pis isn't in any way going to be a "supercomputer". What it would be is a way to learn about parallel processing. Do that with the 4 cores on a single Pi and only then try extending it across multiple Pis. Look up and learn about "Beowulf Clusters" (the way it's usually done with Pis) to understand what it takes, how to use one, and what the pitfalls are.

It is unlikely that Mathematica would benefit from a cluster of Pis. Programs have to be specially written to take advantage of cluster computing...plus it's not trivial to do. On top of that, only certain types of problems can be partitioned into separate execute streams so as to run parts in parallel. It's big subject.

On the plus side, if you learn to program for a cluster and get good at it, you could make a lucrative career with that skill.

[fruitoftheloom]

i just want an affordable supercomputer to run mathematica and learn and do science/math/geometry/physics/coding

and i have wondered how many raspberry pi zero i can put together but i don’t know if there is a limit;

but also, i don't have a-lot of money; and my dreams of putting 4 raspberry pi 4 in parallel were shattered when i realized that one isn’t faster than an iphone; and 4 could maybe match or fall short ? maybe surpass by a little ?

But, not a true number cruncher; and it would be like having mathematica on a laptop (i’m hoping to get it for my desktop eventually, but raspberry pi has other capabilities seemingly)


So i was just wondering if i could stack the 4 rpi0 on a rpi4 and then link those type of things with each other; semi-affordabily

or bare bones; if i can find and buy enough bare bones rpi0 and link a ton of them together and just buy more every month, until i could calculate physics stuff when i am educated enough to model black holes or etc (like the ps4 supercomputer, but i won’t have that kind of money)

so please let me know
(and if anyone can answer my troubleshooting post (not the one i asked for deletion), in my user posts, in my profile; that would help too)

thanks

Linking Raspberry Pis to make a "supercomputer" is a perpetual myth.

You would not try to make a couple of Laptops act as as a "supercomputer".

The Raspberry Pi is as much a computer as a PC or Laptop, just in a smaller "package"

https://www.pidramble.com/

[alphanumeric]

Have a look see at this, https://shop.pimoroni.com/products/cluster-hat

[phi6ndepa]

cool, do you think that the cluster HAT would be connectable with wires to offset it from the board ?

i found a really nice heatsink with a fan; and although it is rather pricy at about $20 , it is still really the best heatsink i have seen for the pi; and it’s reviews of performance are incredible
https://www.seeedstudio.com/ICE-Tower-C ... -4097.html

it uses two GPIO pins, which i want to know about if they can kind of splice and still output power; like if i could get a GPIO band that comes off the pi as well as male/female connectors for the other parts to connect the cluster HAT and also then plug the ice tower fan into the GPIO pins by splicing a connection or soldering;

i don’t really know how this works yet; but basically i just am wondering about a workaround for getting the ice tower heatsink and the cluster HAT on one raspberry pi 4



also, is it possible that i could eventually learn to code mathematica to run on a parallel machine ? i don’t know, if there is a root coding, source code thing or some blocking license thing that keeps it from being used and tinkered with differently.

and also, not to naysay or anything ; why wouldn’t parallel laptops/desktops be a good approach to a supercomputer; i’m not naysaying what you wrote; i just desire a little better understanding of how they work, without tearing my hair out on google searches and endless articles.
i understand that, one) ps4 processor clusters were usable for supercomputers and that now, a similar and better one from the same company is used (article about a guy who modeled/models black holes)
and two) i was reading that some supercomputer uses one type of processor in parallel i think which is then fed to another cluster of different types of processors, so it’s a specific task type of thing; and then idk if those themselves are clustered again; but in general that a lot of pi 0 could play a role (with a lot of them maybe, unless they just never cut it for efficiency per dollar bought or run) in some stage of a machine like that;
it is not maybe something i will get to for a while; but the long road ahead is built well by always questioning it, and i do good with concepts, so i can get an idea of where i am mistaken and what to rethink about and stuff, just keeping it on my mind; so i can ask better questions, and that coupled with learning basics over time; so like a present and future approach, help me learn really well, like maybe a torrent or something, a little ahead, but also, what’s in front of me; and keep me from being bored and overwhelmed too; so like a mental dreamscape to go to and wonder about, before coming back to just having to deal with this basic stuff; so it helps mentally keep me focused by giving me, say, a vacation; but one that’s actually another mode of thought work or something.
so i know a lot of people don’t like to spoon feed; but i’ll definitely look up this stuff; it just really helps when some things are able to be asked about


thanks

[fruitoftheloom]

cool, do you think that the cluster HAT would be connectable with wires to offset it from the board ?

i found a really nice heatsink with a fan; and although it is rather pricy at about $20 , it is still really the best heatsink i have seen for the pi; and it’s reviews of performance are incredible
https://www.seeedstudio.com/ICE-Tower-C ... -4097.html

it uses two GPIO pins, which i want to know about if they can kind of splice and still output power; like if i could get a GPIO band that comes off the pi as well as male/female connectors for the other parts to connect the cluster HAT and also then plug the ice tower fan into the GPIO pins by splicing a connection or soldering;

i don’t really know how this works yet; but basically i just am wondering about a workaround for getting the ice tower heatsink and the cluster HAT on one raspberry pi 4



also, is it possible that i could eventually learn to code mathematica to run on a parallel machine ?

and also; why wouldn’t parallel laptops/desktops be a good approach to a supercomputer; i’m not naysaying what you wrote; i just desire a little better understanding of how they work, without tearing my hair out on google searches and endless articles.
i understand that, one) ps4 processor clusters were usable for supercomputers and that now, a similar and better one from the same company is used (article about a guy who modeled/models black holes)
and two) i was reading that some supercomputer uses one type of processor in parallel i think which is then fed to another cluster of different types of processors, so it’s a specific task type of thing; and then idk if those themselves are clustered again; but in general that a lot of pi 0 could play a role (with a lot of them maybe, unless they just never cut it for efficiency per dollar bought or run) in some stage of a machine like that;
it is not maybe something i will get to for a while; but the long road ahead is buit well by always questioning it, and i do good with concepts, so i can get an idea of where i am mistaken and what to rethink about and stuff, just keeping it on my mind; so i can ask better questions, and that coupled with learning basics over time; so like a present and future approach, help me learn really well, like maybe a torrent or something, a little ahead, but also, what’s in front of me; and keep me from being bored and overwhelmed too; so like a mental dreamscape to go to and wonder about, before just having to deal with this basic stuff


thanks

https://en.wikipedia.org/wiki/Parallel_computing


I posted a link to PiDramble above, it is in my opinion a good tutorial how clustering works at a software level.

https://support.wolfram.com/28109


If the Raspberry Pi 4B is not in a closed case it will not require cooling, it has built in thermal management:

https://www.seeedstudio.com/blog/2019/1 ... than-ever/


Also this forum is run on an entirely voluntary basis, so the depth of support may be limited.

[phi6ndepa]

thanks;

i was wondering though about overclocking it though; and i want to use that awesome fan anyway; i was just wondering; it can stack on the power pins of the GPIO along with other taps on the same, correct ? it will draw adaptively what it can alongside the senseHAT right ? i can’t get my head that far into it right now; so i’m not seeing the flaw in my thought.

also; one thing i was wondering is if a “super” computer could still be made that uses alot of pi 0, which merely out processes a good desktop; but far far cheaper; i was just wondering if at least; a clever use of low cost and efficient parts could out compete a high end or at least mid end desktop in Giga Flops.
but not be a true supercomputer.
?
i’m also reading this now https://www.raspberrypi.org/blog/raspbe ... me-of-age/

[fruitoftheloom]

thanks;

i was wondering though about overclocking it though; and i want to use that awesome fan anyway; i was just wondering; it can stack on the power pins of the GPIO along with other taps on the same, correct ? it will draw adaptively what it can alongside the senseHAT right ? i can’t get my head that far into it right now; so i’m not seeing the flaw in my thought.

also; one thing i was wondering is if a “super” computer could still be made that uses alot of pi 0, which merely out processes a good desktop; but far far cheaper; i was just wondering if at least; a clever use of low cost and efficient parts could out compete a high end or at least mid end desktop in Giga Flops.
but not be a true supercomputer.
?
i’m also reading this now https://www.raspberrypi.org/blog/raspbe ... me-of-age/

Considering you can buy a new i3 Tower PC for under a £100.00..

...and only the Zero WH can be bought in quantity at £15.00 each.

Therefore your assumptions are askew.

[phi6ndepa]

i got this from the one “ It’s important to avoid using micro SD cards and to connect the nodes using wired networks”
that was on the raspberrypi arcticle on the los alamos national laboratory supercomputer simulator testbed

https://www.raspberrypi.org/blog/pi-3-b ... e-awesome/


and also, from the article you gave me from wolfram mathematica;

“ The benefits of running Mathematica on a cluster are twofold: The number of available CPU cores, even on a single node, is usually more than on a desktop computer. Also, the speed of each CPU core is individually faster than those in a desktop computer.”

and importantly

“ Though there is usually a benefit in CPU speed over using a desktop computer, the front end via an interactive session is slower than running a local version of the front end. This is because the front end runs on the cluster while the interface is forwarded to the remote user’s computer.”

this isn’t intended to mean about raspberry pi , but still; if say four pi0 are on a pi4 and four of those link to another pi4 that would be the interactive session ?; while if that central pi4 was linked to a pi2 even, as a remote controller; the central pi4 would be freed up to run as another node ?

is is possible to outcompute a desktop with a strategy like this, before coming over cost ? a laptop at least ?

[phi6ndepa]

may be, but it won’t run mathematica for free, or help me learn parallel computing..


can’t the pi0 standard be had for $5 or 5pounds ? even if not in bulk, but built up over time ?

[fruitoftheloom]

may be, but it won’t run mathematica for free, or help me learn parallel computing..


can’t the pi0 standard be had for $5 or 5pounds ? even if not in bulk, but built up over time ?

https://projects.raspberrypi.org/en/pro ... -an-octapi

[phi6ndepa]

crap; where did their post go about the pico cluster ?
https://www.picocluster.com/collections/raspberry-pi

i had to add something to my post, which they quoted and responded to with that link; but now it glitched or something and erased their post

[phi6ndepa]

is that maybe good for keeping a bunch of tabs open in a browser ?
if the wifi can’t cut it; maybe a ethernet hub would be substituted ?

[phi6ndepa]

is it possible to use network booting of the raspbian system onto a virtual box or some free analog; on a few desktops as nodes that run back into a raspberry pi 4 and then lead to another one that controls them ? would the raspberry pi be the choke point ? or as long as it has enough ram, would be fine; or that the desktops were lead back into a bunch of pi0 so the ram adds up ok and that could feed back to the pi4 or two in parallel which lead back to another one as the controller ?

also does parallel computing remove the use of overclocking on things like the rpi4 if they controll less powerful nodes like the pi0, because they have to match job load or something or can they all get proportional loads ?

[alphanumeric]

The cluster hat could be mounted on this, https://shop.pimoroni.com/products/mini ... 448026055
You could likely also pick +5V and ground off for the fan on it. I wouldn't solder in the third 40 pin header, the one that is going to under the middle of the hat, just solder in the two you need to make it work. It shouldn't touch anything but why spend all the time soldering it if it isn't going to be used. There is a fully assembled (already all soldered) version too, if you don't like doing all the soldering.
https://shop.pimoroni.com/products/min ... 448025991

[Timescale]

Why do I get the feeling that the OP is ignoring most posts and randomly asking stuff without first trying to find an answer by searching?

RPi's, especially the Pi Zero will not get you a supercomputer. It might be a fun exercise and you could learn a lot, but on the parallel computing side, you'd learn nothing more or less than simply using a multi-core platform like a single RPi 4 or Intel platform. I don't know how far CUDA is on the RPi, but I'm sure that GPU features on the RPi are not comparable to that of a basic NVidia PC both in functionality and performance.

In the end, building a pi cluster may be a fun experiment, but it is not going to get you the best performance for your buck and a lot of time will be spend on not learning what you seem to want to learn.

Go for second hand gaming laptops, save a buck and some time and you'll be up and running with 8 cores and 2 GPUs within no time.

[phi6ndepa]

thanks; i do want to get good at soldering; but so far, my tools and attempts have been fails; so now, i have a new iron; and hope to use it for LED strips (with assist from a hot air resolderer, which turned out to not cut it on its own and is most definitely cooking components in the process; and also, i got a heavy duty giant soldering iron so there was no way anything was going to cool down to fast like my older crap ones that never melted anything and broke; but the tip is so wide and it throws so much heat, it was also singing components on its own or when trying to just touch the solder quick when heat gunning;
so now i have a good decent iron, or at least i really hope hope so; but i don’t want to bungle it where it counts. but soon i hope i will be a proficient solderer

[fruitoftheloom]

Why do I get the feeling that the OP is ignoring most posts and randomly asking stuff without first trying to find an answer by searching?

Quite so, just another throw 10 RPis together and it will be superduperspeedy, and just ignore any advice because we as volunteers should spend our time doing the OPs research, cheerio....

[Daniel Gessel]

A modern $500 GPU hits 10’s of teraflops. The Pi4 is less than 100 Gflops. So the GPU is at least 100 times faster by this metric (which the OP mentioned at least once).

You get at least 10 times more for your money by building around GPUs, which is how modern supercomputers seem to be getting built...

[phi6ndepa]

i don’t know much about jetson modules;

https://www.seeedstudio.com/NVIDIA-Jets ... -4414.html

that ~$300 one lists 1.33 TFLOPs of AI power at 7.5W / 15W
i hope they don’t raise the price after they release it.

but otherwise,


i just scoured around for the most efficient CPU for energy, and the only two that were near the top and available sort of cheaply, seem to be, in this order

Atom x7-Z8700 CPU
voyo v3

Intel Core i5-7Y57 CPU
YXKCJ Dell Latitude 7285 2-in-1 Motherboard Core i5-7Y57 1.2GHz 8GB 0YXKCJ


i had previously done this when looking for a GPU, when i needed one that had open GL 2.0(?) or higher to run the space simulator, space engine, on an old, now toasted, dell inspiron

that GPU had the best efficiency that i could find and i believe it is the

Invidia GeForce GT 640, code number GK208-400-A1 ,
the one with PCIe 2.0 ×8 and BUS width = 64bit,
803.3 GFLOPs at 49W TDP
Open GL 4.6 capable
and was around $55

i know they’re always coming out with more powerful things; and efficient things, ie: the Intel Core i5-10210Y
but the prices are high; and wattage is high; or the price is still high;

so i guess the future will see that chip hopefully go down in price; and by that time i guess they’ll have all game changing stuff,
so i guess if i train on small pi clusters i can get ready;
but also, maybe if i get some good low power stuff together, it will be “timeless” for a while; where the low watts allows battery usage and low cost to run off power or use mobile, or in power outages.

but now i’m looking at the next tier of less efficient chips to see if i can find the $30 dollar range

[Daniel Gessel]

I’m thinking of GPUs like: https://en.wikipedia.org/wiki/List_of_A ... 000_series

However, this is far off topic for both the original post and the RPi boards generally.

I believe the Pi4’s 4 cores with SIMD provide a theoretical 24 Gflops and the GPU another 32. Each Pi0 cpu gives you, I think, only 1 Gflop of performance, but the GPUs provide theoretically 24 Gflops each.

So a Pi4 with fully populated Pi0 cluster hat gives you about 156 Gflops to try to get out of the beast.

Not a ton of performance by today’s standards, but you could learn a lot trying to get it up to speed.

[phi6ndepa]

oh; wow, yeah those are much more powerful and efficient; i must have stayed away from calculations because of the price and wattage.

ha ha yeah, i see there are some threads to read on that; and as well as that; that jetson is a real beast; if “ai power” terraflops are equivalent, just flops meaning something; because they blow nvidias own high end gpu out of the water for the wattage use;

i imagine running one with the raspberry pi would really be a good project.
it looks like i’ve got to save some money.