4 Pis on a mini ITX board! The Turing Pi 2
Вставка
- Опубліковано 18 кві 2024
- The Turing Pi 2 puts 4 Raspberry Pi CM4s on a mini ITX board. But how do they perform? Can a little Pi supercomputer make the Top500 list? Find out all that and more!
Check out Patrick's video on @ServeTheHomeVideo - • Ultimate Cluster in a ...
Mentioned in this video (some links are affiliate links):
- Turing Pi 2: turingpi.com
- Open source repo (with HPL benchmark setup): github.com/geerlingguy/turing...
- BC1 mini ITX build platform: amzn.to/3xHoHXt
- Corsair SF600 PSU: amzn.to/32J9NEO
- Pico PSU: amzn.to/3CZ7mdL
- 8A 12V PSU: amzn.to/3o2lXka
- Nvidia Jetson Nano Dev Kit: amzn.to/3HWVBrM
- Compute Module 4 8GB Lite: www.raspberrypi.com/products/...
- Waveshare CM4 Heat Sink: amzn.to/3D8y1Vi
- SanDisk Extreme 32GB microSD card: amzn.to/3ljch3f
- Crucial MX500 2TB SSD: amzn.to/3llZ9tU
- Noctua NF-A14 5V PWM Fan: amzn.to/3I86DL8
- mini PCIe to M.2 A+E key adapter: amzn.to/3pameRF
Support me on Patreon: / geerlingguy
Sponsor me on GitHub: github.com/sponsors/geerlingguy
Merch: redshirtjeff.com
Additional footage from Pexels.com, Patrick Kennedy (ServeTheHome), and from InsideHPC's SC17 video: • An Affordable Supercom...
#RaspberryPi #TuringPi2 #Homelab
Contents:
00:00 - The challenge
01:13 - The build
02:44 - A standard ATX PSU?
04:18 - Installing Turing Pi 2
05:31 - Thermals and heatsinks
07:33 - Expansion slots
09:27 - Fan control
10:22 - Pi OS
10:53 - Booting the cluster
12:40 - Blinkenlights!
13:19 - 146,772 Compute Modules?
16:40 - 1.4 Terabits?!
20:57 - What's the point?
21:56 - Rackmount and outtakes - Наука та технологія
Super cool! Very jealous you have the Turing Pi 2! Super fun meeting up and doing this.
There’s something that makes me so happy when I find out UA-camrs I like also like each other (my wording is bad, but you get what i mean :) )
I'd like to know the comparison of performance per $ between those systems.
VIEW * NICE AWES0ME C00L KEEP_IT_UP ! INDEED ....📳💻💻💻💻🖥🖥🖥🖥🖥⌨⌨⌨⌨⌨🖱🖱🖱🖱💽🏆🥓🥓🥓🥓🥓🏎🏎🏎🏎🏎💨💨💨💨💨💨💨
Love the collaboration. Thanks for bringing this all together gents.
But is a test bed really a case? 😉
The fact you open-source everything you do says a lot about your love for computer science and the community community. Much love Jeff Alex from the UK
On god….
"The fact you open-source everything you do says a lot about your love for computer science and the community community." Is there a non-community community? o.o
@@chrisakaschulbus4903 yes... Join the Gun community lol its nothing but calling other people poor.
In my experience, open-sourcing random things is sometimes useful for other people to figure out how you did it, but it's not actually as awesome as people make it out to be. It's like, imagine if everyone put everything they did on UA-cam. Sure, all the good stuff would be there, but also mountains of useless junk to filter out.
OOOH! It comes with the ATX power connector! Yay!
Yes but everyone noticed in post already. Anyway yay now we can use a pc psu
Wow. Patric knocking it out of the park with compute power. It’s really cool to see server vs hobby grade cluster.
sadly Patric did not run any benchmark
eeeh, thats more like baby server vs hobby grade haha.
I still remember the times when there was no such concept as a "hobby grade cluster" ... before people figured out that clustering PCs could get you respectable supercomputing scores for much less $$$ than traditional supercomputers!
@@KaiHenningsen Must be along time ago we even had a guy living next to us in the late 1990s that showed me how a cluster of 4 Amigas could draw a 3d model quicker
Different people invest differently into their hobbies. But I think most hobbyists, overclockers, enthusiasts, powerusers, gamers, miners aren't going to put over $25K into a their computing box.
Ahh yes.
Noctua fan and the SF600 inside an ITX build are exactly what I think of when the Raspberry Pi comes to mind.
Actually no, my gaming rig comes to mind inside a Node202 but you make me want this board too despite having no use case for it.
You built a super computer training rig, where people can learn to run super computers without the huge cost of a super computer. Colleges and Uni's will be queuing up for them.
I've been waiting for this build for years, I am super happy and proud to finally see it complete!! Thank you!
Been looking fwd to this video literally for years, and it absolutely lived up to expectations. Great work Jeff!
Can't wait to see what else you get up to with this rig
Your build I could mostly get my head around, but Patrick just blows my mind. Nice one Jeff!
The message in 21:22 is very important, not only in terms of physical hardware, but also in terms of buying servers in the cloud. People with money tend to get the beefiest ones they could find, but they probably could have made the servers scale up/down depending on the usage. :D This generally applies to individual website owners, e.g. IMers though.
Love this channel. It always has something interesting and Jeff's approach to the technology is reminiscent of my early days at Rockwell International. Keep up the good work Jeff!
“M1 laptop can’t run Linux yet” nice put down line for Apple’s fast laptop! I am used to hearing Chris from Explaining Computers gentle wit and so I appreciated this line from you. Thanks for making me smile :-)
Love Chris' videos, he's a bit of a source of inspiration too ;)
It walks, for running it will be more dificult
This weeks update of M1 can easily do it.
@@szymex8341 yes it do. I am Manager of Software grp in Apple
@@szymex8341 do you know hindi as i am Indian and i know it well.
Jeff, I love it where you're super honest at 11.48 where you admit booting it already 😂
Great video (I'm 2/3 of the way through)
Wish we'd've seen the watts the other machine used during a benchmark. And would've liked to have the performance comparison too. Even though it wasn't the point. I'm still curious! Great video!
Also a gigaflops rating means nothing without how large of numbers its working with. FP16, FP32, and FP64 will be very different speeds.
Loved the geeky conversation between Patrick and you. Priceless.
I've been waiting for this thing since last year, jealous af, thanks for the demo/review!
It's a comparison like a dart to a space ship, yet both can fly. Thanks for the great video.
This was like comparing a schooner with a sloop, we can go sailing in either one of these. As a developer, your build is the better choice because anyone can test their project at home with a modest budget. Well done!
Idk man I can basically do exactly the same thing with far fewer parts in docker containers.
This is the first video of yours I've seen, but I adore your style and presentation. Subscribed!
This is really cool. I'd love to own and play around with both of these builds, but honestly I'd probably pick the Turing Pi 2 to keep long term (especially considering the direction electricity prices are heading here in Europe). Would be great for a little home cluster!
I think six nodes would be ideal for me though, but that might be a bit tight on a mini ITX board.
You can put seven on a mini-ITX Pine64 Clusterboard.
Energy efficient is future. Here electric company started 200MW wind farm last week and all of it goes to Google. And we import 10% of our electricity from Russia, scary enough, but add to that EUs highest price. oh no.
Low end GPUs are much more efficient, like on the order of 10-30x more gflops per watt.
This is awesome seeing two of the homeserver kings.
As a HAM, it might be something to look at for Disaster Comms (the LTE setup) for a heavy duty HAM Radio command center, something like Jason KM4ACK's Build A Pi setup. If you aren't a HAM, a lot of features can still be used...
there is a nice bunch of remote listening designs with PIs around.
Lets say, have 8 software decoding channels of RX in a single compact box as an Ethernet attached web GUI control.
Use each of the channels for decoding e.g. AM, FM, side-band modes, CW or the modern merely automatic digital modes.
Ah, and dont forget about a useable antenna, preferably at a location with low levels of any sort of noise - indeed most of it is man made.
regards, Alex, DG3MMF
Me: eats HAM
Disaster comms: *confused screaming*
LTE for disaster comms? What? If the cellphone network is still up, it's not exactly a disaster where you'd need to use ham radio, is it?
So nice of you to be sharing this content thank you!
Love this build platform, makes perfect for build a case around it... or adding some 19" rack mounts and rack mounting it...
Man that is some algorithm they got there, it's almost as if Jeff and UA-cam planned this and knew about this collaboration as I just stumbled upon Patrick's channel about 1.5-2 week(s) ago.
I love dreaming of the things I could do with STH's gear 😄
@@JeffGeerling Yeah for me though it's kind of like a dating an intimidatingly smart supermodel or owning a multi million dollar formula one racecar, I just wouldn't know what to do with them.
I think this is the future of computers; the computer modules and the mainboard, smaller daughter boards. In a way, this seems to make things return to the beginning of computing, like when we were all playing with Z80's and adding smaller boards, mixing boards, hybrid Frankenstein monster computers.
c128 lol
a z80 and a 6502 on the same computer, madness!
No way. Economy of scale is economy of scale. Swappable parts yes - multiple nodes no.
can we use AI to have a nas that monitors internet usage & pulls stuff so that can be used later
can network be smart enough to see anyone uploading to public near your area doesn't go to main server & later distributed?
sorry english is not my native language
@@alexc4924 I mean, running 4 or 10 Raspberry Pi 1s, just cause they're laying around... technically, even the big manufacturers, like AMD, are considering using Chiplet designs to help ease current design problems.
edit: come to think of it, we still do this with GPUs, Ethernet cards, and other PCIe based cards.
First time viewer & now subscriber. Really enjoyed the pace, production, and process. I would have loved a quick epilogue, cliff hanger, or blooper reel. XD Thanks for the great content!
yes! ive been excited for this all week!
I'd love to see how these could be used, and ways of managing them in a friendlier way.
I'll be exploring that in future videos!
Ansible and related toolings will do you a great favor.
You might rather lack of an application case for them. ;-)
Both clusters can run the exact same containers. Keep that in mind when comparing them.
I've been so excited to see the new board! The first turing pi board is what lead me to Jeff Geerling!
Jeff watching you GRIN like a super excited made me smile and made me wanna buy one of these even MORE lol
"SATA, not SATA" is something that belongs on a T-shirt
Pretty cool, but it's driving me a little crazy that you positioned the board with the i/o to the side. Doubt it's getting too warm, but the fan can't efficiently cool all boards facing that direction. Probably won't be an issue once you put it in a case.
Nice I'm so glad you are doing this with the pi
Great refresher Jeff!
Man. I’m just getting started here and both of these solutions sound like incredible solutions for clustering. I’ve got a few Jetson Nanos that I would love to put into a cluster board from Turing Pi. Cannot wait to see the final version and possibly pick one up. How did you get started in the IT industry with cluster computing? I’m getting started in the industry trying to learn different methods of balancing loads between different systems and running through GPU computation using python and C++ but I’m looking for things like your current project which seems like you will need to balance the nodes and hardware. How/what video of yours or someone else’s would recommend for learning this material from scratch?
On the PCIe, would be nice to have some kind of pcie switch onboard that lets you map peripherals to slots using the firmware, but this is a clever solution nonetheless
My dream have be updated, I love ITX boards. Thanks for the video
I like these outtakes way more than your other videos :)
Really interesting for home use, just wish there would be CM4s with more memory, I don’t need alot of compute power, but need more memory, this would be really sweet if the PIs had 4x16GB or even 4x32GB memory, for running k8s nodes.
Here in EU, power usage quickly get stupid expensive, this would fit well for that.
What workloads are you running that you need 64/128 gigs of RAM on a Pi? I suspect, even if those existed, you'd find the CPU and memory bandwidth on the Pi can't keep up at that point. You're probably better off looking outside the Pi family or even expanding cluster size (or multi-clustering) if your workloads can be broken down.
Here is a tip: used resedental solar panels are kinda cheap
4:56 I'd love to see a video testing various SODIMM riser boards/cables with the goal of seeing if we could get this thing down to a 1U form factor.
I think this 2U box can hold far more than a pair of 1U boxes which are full of risers and extenders.
I
f it's already greater compute density, higher power efficiency, and lower cost ... what would you gain by flattening it into a single blade?
@@pwnmeisterage I was mostly thinking about the remaining space in my rack than I was computer density.
YES JEFF YESS!! I haven't watched yet but I've been waiting for this one for a while!
You guys are one of the reasons why I can’t get any Pi component 😢. Good video thanks for sharing.
To be fair, the Bluefield2 cards also have a 1GBE interface each that's internally directly connected to the Arm side.
You node it's serious when it has an ATX power connector...
But seriously, this board ticks many boxes people have been asking for
Thank you for your work
Love the fact that you know your stuff.
Lucky SOB! I've been drooling over that V2 board for the last couple months as I want to use it for a Robot Operating System cluster for a robot. But between the V2 board not being out yet, and ANY RasPi right now being basically unobtanium, it'll be a while before any of us can build a setup like what you showed... #jealous #f-covid
Don't just build overpowered temporary setups. Show use cases. Show permanent builds that you (or someone else) will use for something real.
anything that runs on kubernetes can run here
@@tianlechen but it could also run on one normal server with 4 times the power
which is why clusters only make sense when you need more power than you can get from one server. Which is why small clusters are just pointless. The exception is "to learn how to program clusters" but if you are making videos about clusters and going to cluster conferences, that is not your use case.
@@alexc4924 Power consumption?
@@sceerane8662 no, compute power
Another admirable project!
this is sooo cool, thank you for makeing this video :)
So what would be real world use case for any of those systems?
What would one compute and store on any of those?
I am confused though on fan placement considering it can be screwed on any side in 2 positions each side
oh man... I can't wait for this to come out! I'm planning to use these as a kubernetes learning lab.
OMG I've been waiting for this for a long time Here we go!
Real German: "Relaxen und sich die Blinklichter anschauen" :D
I didn't know you could configure wifi and ssh from pi imager, that's really nice to know
Yeah, saves a lot of time :)
@@JeffGeerling and what is the 'Open Sesame' to access that option in app??
@@VaibhavKurde ctrl-shift-x or cmd-shift-x depending on your os/keyboard
Jeff, great build. Sorry to have missed you at SC21!
Just amazing - throughly enjoyed it !
So Jeff , what is this thing good for though? Does it make a really awesome nas?? I'm just confused on why would you want clustering??
That's a good question! Check out my video on the topic here: ua-cam.com/video/8zXG4ySy1m8/v-deo.html
Parallel computing with overset grids for computational fluid dynamics is a great reason for clusters. By dividing huge grids that won't fit on one node (memory wise) into much smaller subgrids each fitting on one of many nodes, very large CFD problems can be run. NASA has been doing this for decades but with much larger and much more expensive computers. This was how the flow field around the space shuttle was computed back in the '80s.
Jeff has a very good point here: m1 macbooks are more power efficient comparing to pi. I've seen people justifying pi's horrible desktop experience with saving power. lol that's not true at all. It would be something if apple would one day release something like a M1-compute module, but alas that day will never come.
There is no need to think about the hearsay from others, like the irrational justifications you pointed out (which made me laugh, BTW. Thanks:) )
Let's just stick to the truth and the facts: Raspberry Pi < $100 vs. MacBook M1 > $1000. There is no need to compare Apples and Pears ... (pun intended:P hehehe). I second your thought about "Maker-Stuff" from Apple. The company would have to publish circuit diagrams and treat its customers like people and not incompetent idiots from whom they have to protect their "important secrets" (as usual nowadays, "The right to repair", LoL). That only happens when hell freezes over:) Thanks for your comment, Zhen:)
Really nice book with a lot of informations, I appreciate it. I'm using it in my IT training at my company in Germany.
Both builds look good!
Very nice content, thanks for taking the time to inform us!
However, looking at the technology; I really wonder what the benefits would be. No doubt that board + 4 pi's and so fort, costs about the same as a ryzen 7 mainboard and a cpu with 8 cores.
I don't see the advantage. At all. The pi is limited in bus width (pci express lanes), not really expandable easily, and is non-standaard ARM hardware.
Also, that ryzen setup woud totally trash the pi performance wise. So I don't see any advantage or reason to want to do this. If you want to run a cluster so bad, you could also run VM's on a normal PC setup. And they would still easily outperform the pi.
21:11 it can be used as a cheaper diy network server that runs on pie instead of a more expensive pc. pushing low end hardware is fun for hobbyists and is why the pie exists in the first place. this could also be for those who don't want to have a cloud based solution. a cheap chinese mobile phone has more power, but you don't see that being used anywhere for these things.
@@VD-cc4hx thats exactly my point. Setup like this, it's NOT cheaper.
So... lets math this out. Turing Pi 2 Mobo, $200, $10x4 Pi CM4 Adapter boards, power supply $30-$50, Case $0-$200 CM4 boards x4 $35-$85 depending on where you get them and what version you get, we'll say another $100 - $150 on the low end for SD cards / or SSDs etc. Looking at a little cluster in a box for $500 to $700. Fairly expensive but considering the learning potential and the small size and flexibility, actually rather cheap. Totally worth it for all that you can learn on it, basically a homelab in a box.
you're spot on with the estimate!
For less than € 200 you can buy a chinese motherboard with 2 intel Xeon E5 processors = 24 core each and 16 Gb memory…
@@xaty1808 Sure, but that's not the aim here. This is more bare metal cluster, not a machine that can virtualize a cluster. It's not about the performance per $ spent. It's about the environment that you can build to learn and test clustered sytems. You could also easily buy 4 older small form factor business desktops for about $50 each and cluster those, and have the same rough environment. But it will take up far more space and draw more power than the pi cluster ever would. So it's really about where you want to spend your money.
Wow! Can't wait to see this in the rack with software running :D
Pretty cool project! Thanks for sharing! :)
Jeff, you've got that smirk on again that means you're up to no good .... ! ;) Hehe. EDIT: It's not a smirk, it's a "cunning smile".
I am amused that the ARM-powered Pi needs an ARM CPU as a controller. ARM really is flexible.
It's ARMs all the way down!
@@JeffGeerling Waiting for someone to call the 8 slot cluster board the Octopus.
@@JeffGeerling Yeah, since I'm pretty sure there are also ARM controller chips inside the microSD cards!
Enjoyable and educational, thank you 👍
solid content, thank you
Now i want to see you build a cluster with Jetson Nano and run some tensorflow training.
I accidentally read it as turning pi. Was gonna say mmmmm....turned pi...
Great video Jeff!
Thanks for the raspberry-pi imager pre-install tips, might be verry usefull.
Nice video 🙂
can it run crysis?
It's SATA and not SATA.
Go crazy with every cluster!
Great video! I love both channels!
Cool clusters, cool episode!
I want one so badly! Looking forward to see a series on the TouringPiv2
ha! I can't imagine Patrick building anything 'tiny' ;) great video!!!
I like that! and will be fallowing you for more great ideas ...currently I run all my amateur radio equipment from solar and my pi3 and pi zero. The pi3 runs my aprs and pi zero runs my satellite tracking.
i love to watch video like this.
seeing people commenting also make me feels like "wow you guys are awesome too"
And at the end of the video i realize that i understand nothing. lol
Jeff: mini itx board
Patrick: so anyway here’s my threadripper
Pity that even Jef likes to pretend that four RPI4 becomes a "superomputer".
This channel is more about all the improbable things one CAN use a Pi for, not what makes much sense.
If you want hardware with usable performance, buy used professional hardware from ebay. I got me a couple of 12 core HP DL-380 servers for about half the cost of four RPi 4. Oh, so they consume a couple of hundred watts each, but you can run real loads on them. Those 12-core servers, actually are "super computers", in that their CPU's DO share ram and resources. Then you section them up with ProxMox and play with Kub all you want. :-)
The "limited budget" argument isn't particularly valid I think, given the prices of used professional hardware.
I read the intro on STH first, then before I got into it, I came over here to watch the two vids. How did I know Patrick would go completely nutso on his build?
Really love the channel and it is great how well it is doing :) UA-cam’s thanks feature is great
Sold!! I'm so getting one 💚
Great Video Jeff !!!!
I can't wait for more videos on the Turing Pi V2
Nice video clip, keep it up, thank you :)
Nice little project Jeff. So the long and sort of this project, you got yourself a hi tech Xmas tree
amazing project.
Brilliant video! very informative,and to be honest I love how cute looking the Turing Pi 2 is, Patricks was a bit of a monster to be honest lol. Anyways, new subscriber here :)
Nice ❤ this is definitely a step in the right direction
LOL 17:18 you've just been BASIC'd by Patrick; this is pretty much that chad meme right here.
Hey Jeff, Love your work! Could you do a video doing some processing on your cluster using Dask?
Get well soon my man!
The SF600 Platinum was the first thing I spotted on your desk. Rule 1 of fight club is to NEVER skimp and save on the PSU and these PSUs are quality. Recently purchased the SF750 and am very pleased
very informative...thanks!