We feel for [Jeff Geerling]. He spent a lot of effort building an AI cluster out of Raspberry PI boards and $3,000 later, he’s a bit regretful. As you can see in the video below, it is a neat build. As Jeff points out, it is relatively low power and dense. But dollar for dollar, it isn’t much of a supercomputer.
Of course, the most obvious thing is that there’s plenty of CPU, but no GPU. We can sympathize, too, with the fact that he had to strip it down twice and rebuild it for a total of three rebuilds. One time, he decided to homogenize the SSDs for each board. The second time was to affix the heatsinks. It is always something.
With ten “blades” — otherwise known as compute modules — the plucky little computer turned in about 325 gigaflops on tests. That sounds pretty good, but a Framework Desktop x4 manages 1,180 gigaflops. What’s more is that the Framework turned out cheaper per gigaflop, too. Each dollar bought about 110 megaflops for the Pis, but about 140 for the Framework.
So was it good for AI anyway? Predictably, no. While the Pi 5 does have an integrated GPU, llama can’t use the version of Vulkan for speedups. Even a cheap consumer PC can turn in better performance. The Framework without its GPU did about six or seven times better. With the GPU? Around 14X compared to the Pi cluster.
Should you build it? [Jeff] says no, unless you have a very special use case for it. However, we build plenty of things that aren’t super practical. If you have a use for the beast, let us know in the comments.
Even if your cluster isn’t as powerful as this one, you can still pretend it is a Cray. We wonder if ten Pi 5s can beat 1,060 Pi 3s?
most home computer users dont have a need for vector compute beyond perhaps rendering. besides if you did need light parallel compute you would find your gpu has you covered. been programming for years and never had a problem that i could just throw all the cores at. if you find you need a lot, its probibly cheaper just to rent time on a real supercomputer than cobble something together out of pis or gpus or whatnot.
Never found a case? I’ve certainly found it useful several times. xz will use all cores when compressing. Optimization problems and analyzing games are slightly esoteric, but I’ve written code for those several times. Image processing also can be done massively parallel, things like old school ray tracing or fractals.
Also compilation can often be done in parallel, compiling many files at once.
I think Jeff Geerling found that Pis weren’t cost effective. Renting cloud space can be cost-effective, but buying the right equipment means you can always call upon it, sometimes without noticing it, and it’s a sunk cost, not something you have to worry about or set up. No data leaves your system, and there’s zero chance of a huge unexpected cloud bill.
im talking more parallel capabilities beyond consumer accessible hardware.
Haven’t watched THIS video, but Jeff’s videos are often named with a little tongue in cheek. His variant of being over the top to feed the algorithm.
I would guess he knew EXACTLY that RPis are not good at AI – cluster or not.
This.
YT creators job is making content (bleh). The only reason for building “$3K pee cluster” was clickbait, and it worked.
Forget AI. For $3,000 I would turn this into a DIY planetary ground station mesh. Strap SDR dongles on the nodes, scatter antennas, and suddenly your Pi cluster is listening to weather sats, ham traffic, ADSB, even deep space signals. Not a failed AI box, a personal window into the universe
What an imaginative, positive and lovely thought. Thanks Zachary.
I was thinking about a dedicated area repeater traffic recorder, but I suspect that spinning rust would be better suited for the task than SSDs. Your idea definitely expounds on it with other with other cool ideas.
Appreciate that. Picture this cluster pulling in faint whispers from weather sats and old probes while also scanning today’s traffic. It would be part cosmic archive, part live stream. For $3000 you did not buy hardware, you bought a personal backstage pass to the universe.
Can you synchronise cheap SDR dongles?
If not, you might not get much gain from having multiple antennas. But if you can sample synchronously, you can do some really cool things.
Yes you can. You need to things. First, give each SDR a common clock. Add some RF switches so each SDR can either connect to its antenna or a common noise source. After powering the SDRs on, you switch to receive the noise source and use the cross correlation of the signals to determine what delays need to be applied to the signal from each SDR to make them coherent.
I spent some fun in grad school laying out this sort of system for some RTL-SDRs before I found a cache of Ettus Research SDRs that where better suited for the task. Now days, you can just buy a KrakenSDR off the shelf!
Aside from being less bleak, RF is also a place where having this many distinct nodes is potentially rewarded architecturally. Unless you’ve got a particularly cooperative signal or just loads of top shelf coax on hand you will probably have a better time if the receiver can be right next to the antenna and then whatever-IP back to you; rather than RF cabling all the way from the antenna to a receiver right next to you; for which purpose little compute modules with enough punch to do some SDR can be quite helpful.
indeed that seems a better application of rPIs… for 3k i’d rather buy a card suitable for ai…still could write an article about it….
For $600 you can pick up a 4070 Super with 365 AI TERA ops.
I don’t know what this person was expecting without a GPU
I guy in our local linux group put together a pi cluster to help solve GM’s east coast parts ordering division. There was horrible lag in the order system, their it staff kept throwing more hardware at the issue without actually looking into the problem. first issue was you would enter a part query and it would return 50mb of data. the majority wasnt needed. He cleaned that up and got better responce.. but added the pi cluster as the gpu’s could process more requests of the part numbers and get the right info to those making the request.
“IT staff kept throwing more hardware at the issue…”
Ah, a common solution that is self-serving.
– More hardware, more staff;
– More staff, bigger budget.
in reality just stick a few more boxes in the server rack you are already maintaining and it will have paid itself in no time, hardware is cheap
“IT staff kept throwing more hardware at the issue…”
And yet we are still running DOS software at work. No Joke!!
I’m a confused how Raspberry Pi was the correct solution to a business-critical web server at one of the largest companies on the planet.
At a company that size, one person should be responsible for writing the software and people with a totally different job title should be managing the hardware (in the form of a private cloud). Otherwise you get situations like this where a software engineer at the 19th largest company in America decides to run a critical system on some less-than-consumer-grade hardware hidden in a closet where it will sit (without security patches) until the default Raspberry Pi OS logging settings wear out the SSD a year or two later.
I’m confused by this. Even the latest Pi’s have pretty limited GPU’s.
IT at GM should have more than enough bare metal and VM licenses to be able to spin up 1000 rPi’s in about 10 minutes.
I’m not up on the latest for Pi clusters — it’s been a few years, , but it used to require lots of manual copy and config – once you got an image working, you still had a relatively slow flash, even with SSD, then take the Pi out of the cluster, change name, address, etc., test, then add back to the cluster, test some more. Not exactly Chef or whatever.
Maybe this was some garage or warehouse someone built a hobby solution for?
This story sounds wrong. Even a thin client from 5 generations ago (Like Dell Micro, Lenov Tiny, HP Mini) has much more compute than a Pi Cluster. Why they wouldn’t use a desktop already in the recycle pile, or more realistically a server they already had powered on, makes this story sound fake.
If there was more details, like they used a Pi to test the implementation before moving it to the server, or maybe the Pi was used for the parts department ordering screens it would make a little more sense.
At first read, this sounds like BS.
Then I recall that EDS (now HP Enterprise) was GM’s IT solutions vendor of choice for decades.
ANYTHING is possible, except a competently implemented solution.
If you ever wonder where Tata and Infosys got the business model of ‘Never deliver anything good or hire anyone technically competent’ it was EDS who did it first.
They were/are good at selling services to fortune 500s and government agencies.
Once the BJs are over, just a bottomless money pit of incompetent recent CS 2.0 students who don’t actually know how to code or admin a network.
Foreigners don’t laugh, I also ran into EDS at ETSA.
The south Australian power company.
At the time all ETSA servers were in Victoria, ETSA’s offices were running on one ISDN class line.
They were ready to setup a shadow network in the office, cat5 on walls.
But scared of rules cops, got paid to sit ass/drink tea waiting for network instead.
Apparently the fear was justified, building full of rats, only one floor of engineers.
Apparently some twit had contracted EDS to do all SA state IT.
Which included the power company.
Bet the BJ was great.
Last I heard Boeing, GM, GE, etc etc where heavily “investing” into SAP and Six Sigma Black Belts to address this exact need, company-wide inventory of parts that can be shared among its many divisions.
I am guessing, quarter century later, and billions spent on managers’ salaries didn’t not produce the results wanted and a group of non-SAP-trained contractors could accomplish the same without the hefty budgets.
First time I’ve heard “SAP will fix ALL the warehousing issues” was mid-1990s. Entire departments were corporate-restructured into “SAP EVERYTHING” and those who didn’t were let go. Late 1990s, early 2000s if I’d mention “SAP” the response would be “managers’ toy”.
Rewind to present and find AI in place of SAP. Oh, well, the more things change, the more they stay the same.
I’m from Belarus but all my family moved to US 4 years ago and I still find it disgusting how people there are allowed to waste money on things and projects that benefit no one and just waste available resources. In my country if someone tried to buy 10 R-Pis for a joke video they’d probably get asked some serious questions by tax authorities and militsya. For reference, when I found work after finishing school, I saved money for 2 years to buy my first e-bike. Here people buy them like it’s nothing. Why?
Because despite significant stressors, the US is still basically a democracy, unlike Belarus, which is basically a military dictatorship.
Because while the US has many problems and inequities, it is still ranked by independent NGOs in the top quartile in the world for human rights, whereas Belarus is ranked in the bottom quartile.
Because US culture, right or wrong, values individual hegemony and autonomy, especially with respect to personal financial behaviors, as one of its most important values and Belarus does not.
I could go on but you are comparing a small former bloc country which did not embrace either democracy or capitalism with a relatively large country that did.
It’s like asking why education or health care or gun control is so poor in the US relative to Japan or Germany or Sweden or Denmark.
Answer – it just is. If it was like those countries it would be a very very different US.
Agreed 100 per cent !!
Do you also got mad when a studio buys a car to crash it in a movie?
This is no where near as bad, Jeff made his video and now the pis will be retooled for other projects. They’re not going into the dumpster.
Because capitalism works. Jeff had 3000 dollars, and made a choice about how to spend that. Letting individuals control their own wealth and make their own choices has been economically successful. Besides the deeper economics, paying people to worry about what every middle class individual is buying costs a lot, and it discourages people from buying things, which drives the economy.
Glad I live here in USA. If I want to spend $3000 or more on electronics ‘for fun’…. That’s my business. Called freedom. Freedom to spend your disposable income any way you please. Looking around my home office I have 8 RPIs running ‘fun’ projects. A couple more Pico 2s running, and current working with a Adafruit Fruit Jam board…. Not counting desktops, servers, printers, networking hardware and such scattered around….
That said, Jeff can turn around and reuse the cluster for other uses than silly AI projects. It’s a hobby. Good stuff.
In your country the tax authority questions 10 R-pis because only oligarchs, drug dealers, and terrorists have that kind of money to spare.
The average monthly salary in Belarus as of May 2025 was approximately 2,718 BYN (Belarusian Rubles), which is around $925 USD
The U.S. average monthly salary is about $5,200.
You need to save for 2 years to buy an ebike.
We just need to skip McDs and Starbucks for a couple of weeks.
“The U.S. average monthly salary is about $5,200.”
In my area of the USA it’s more like $1,700
If you make $10 hr you are doing well.
When I started as a 911 dispatcher 26 years ago I made $7.85hr…I
had to work 2 jobs just to keep a roof over our heads.
Not that much better today. The agency I started at is still only ~$12.00hr !!!(Today)
Move or stop bitching.
Your choice.
Spoken like a true neoliberal capitalist.
The two holy tenets of neoliberalism:
1) Because markets !
2) Go die !
Username checks out. LOL ! Apparently the neoliberal capitalist world is so delicate that any criticism of it must be shouted down. Not exactly a defensible position…
WHERE????
Mississippi consistently ranks as the state with the lowest average monthly salary, with average annual salaries in 2024 around $47,570, which translates to approximately $3,964 per month.
Rallys burgers in New Orleans is paying $15/hr 27 hours a week is more than you claim is average in your area.
Walmart pays $19/hr to people who shop orders for delivery. 21 hours a week would exceed your claimed average.
So seriously, WHERE????
“How people there are allowed to waste money on things and projects that benefit no one and just waste available resources.”
Apart from politics and economy you make a 2 significant mistakes assuming there is no benefit and only waste.
Benefit here is knowledge and skills gained during making and running the project. For his 3000$ we have now learned it’s not worth it in most practical applications.
And since his resources are not unlimited he will probably reuse all the parts – there are comments here that already give ideas how (bonus: social points). Not much waste.
Ironically he gets the money from people watching the video. You watched the video, so you helped him make projects. I did not watch the video. So it is an odd situation. I personally feel that a better result would be obtained with a single CPU machine, and he does indeed have a 128-Core ARM desktop. I don’t know if he compared the two.
I assume he wanted to learn about building supercomputing clusters, the problems and inefficiencies of multiple systems working on a singular task. This platform has plenty of both.
Belarus GDP per Capita $8,316
US GDP per Capita $85,809
That’s why.
Beyond that…..
Belarus GDP $75.96 billion
Nearly 500X
US GDP $30.35 TRILLION
The CITY of Minneapolis, Minnesota has a GDP of $297 Billion almost 4X that of the NATION of Belarus, and its only our 15th wealthiest city by GDP.
There are only 5 STATES in the US with lower GDP than Belarus.
edit:
Belarus GDP $75.96 billion
At nearly 500X that
the US GDP is $30.35 TRILLION
GRR Edit of Edit:
I meant to say
Belarus GDP $75.96 billion
At nearly 400X that
the US GDP is $30.35 TRILLION
Seriously HAD get an edit button
I feel like I regret most Jeff Geerling Pi videos.
Outside of the ones that gave him recompile the kernel and upstreaming patches.
That’s quite harsh. He’s pretty entertaining.
Should I be surprised? The Raspberry Pi by itself isn’t very powerful, a whole bunch of them? Still not very powerful.
Are they super tiny and save a lot of space? Yup!
I have no idea how those different AI software platforms operate, but I’m curious if the ones built for ARM processors will work on those tiny Nvidia SBCs that come with their own GPUs? The guy spent $3k on that project, probably could have bought a couple of Nvidia Jetsons with that budget, depending on the model that is, the higher end ones are pricey af…
You can build a desktop of standard parts for half the price and include several options for AI. Either AMD iGPU with 64GB of GPU VRAM allocation, or a 16GB or better nVidia GPU, probably both.
Also the PS5 CPU blade servers exist, with 16GB shared RAM, (6-cores on the CPU and 20 of the 24 GPU slices are active compared to a PS5) I assume you can do AI on them. The price fluctuates between $50-100 for a single blade, but you could probably get 2 of the (12-blade?) boxes with cooling fans and power supplies for less than $3k.
I did not watch the video because the very idea was dumb, but I am assuming he bought the marked up 8GB models. Not sure if he could have bought the 1GB/2GB model and reflowed the RAM chips for significant savings. Some models are easier than others to do that.
I don’t know of any ‘AI’ that isn’t a RAM hog above all else – the architecture is always going to fall far below that in importance. So this cluster of Pi was never going to be good at these ‘AI’, and Jeff doubtless knew that, but got the chance to run models a Pi couldn’t hope to run at all thanks to clustering them.
These Pi clusters and racks can have practical uses, but that was never one of them – Tasks that can be broken up to many parallel jobs where the performance of the task is probably more limited by the bulk storage speed or network resources, or when you need uptime/scalability more than performance. And once you stop worrying about clustering them the reliability/security of dedicated separate hardware trumps any faux isolation provided by virtualisation and creates a system tolerant of multiple node failures. You can’t build a more compact and performant group of isolated computers than a CM5 in one of those blade concepts right now as far as I’m aware. You can have more performant nodes obviously, but packing 20 of them into a single 1U rack slot not a chance.
You can get a far more powerful machine that will happily run a reasoning LLM such as Qwen3-ST-The-Next-Generation-II-E32-v1-256k-ctx-6B.i1-Q4_K_M.gguf via llama.cpp for half of what he spent.
AMD AM5 Ryzen 7 8700G AI 8 Core CPU
Gigabyte AM5 MicroATX B650M Gaming WIFI DDR5 Motherboard
256GB DDR5 (4x64GB)
2TB Kingston NV3 PCIe 4.0 NVMe SSD
That is it, not even a dedicated GPU card.
Mac mini M4 is about 10x faster than the RPi5, so that could have been a more economical option.
So bored of Jeff Geerling’s bad takes on Rpis. Underpowered? What did he want? A mining computer. Loads of good use cases for that system, but $3k would have been better spent elsewhere anyway. Luckily he has hundreds of free Rpis to waste on projects.
These have value for the learning experience 🤠
pi clusters are great for learning how to make pi clusters and not much else. Its fun to build sure, but everything you could do with one can be done better faster cheaper with random laptop and a bunch of VMs.
I legit don’t understand the fascination with RPis. Yes, they have their use cases but some of this stuff is just… Pointless?
I made my father a photo picture frame with one which allowed him to easily put photos from his phone on the picture frame as a non-technical person. Also added a Pi to my Ender 3 pro printer to get it connected to the network, which was another good use case…. but people running virtual machines, servers, services, home assistant, et al off these things is just dumb. For what a Pi5 costs, I bought a 65-watt HP 800 mini G4 with an i7 processor, 32gb of ram, 1tb nvme.. a machine that is orders of magnitude more powerful than a Pi5.