I placed a low bid on an auction for 25 Elitedesk 800 G1s on a government auction and unexpectedly won (ultimately paying less than $20 per computer)
In the long run I plan on selling 15 or so of them to friends and family for cheap, and I’ll probably have 4 with Proxmox, 3 for a lab cluster and 1 for the always-on home server and keep a few for spares and random desktops around the house where I could use one.
But while I have all 25 of them what crazy clustering software/configurations should I run? Any fun benchmarks I should know about that I could run for the lolz?
Edit to add:
Specs based on the auction listing and looking computer models:
- 4th gen i5s (probably i5-4560s or similar)
- 8GB of DDR3 RAM
- 256GB SSDs
- Windows 10 Pro (no mention of licenses, so that remains to be seen)
- Looks like 3 PCIe Slots (2 1x and 2 16x physically, presumably half-height)
Possible projects I plan on doing:
- Proxmox cluster
- Baremetal Kubernetes cluster
- Harvester HCI cluster (which has the benefit of also being a Rancher cluster)
- Automated Windows Image creation, deployment and testing
- Pentesting lab
- Multi-site enterprise network setup and maintenance
- Linpack benchmark then compare to previous TOP500 lists
I think the only answer is “Doom”
If OP actually does do this I recommend Odamex
Although he’d also need 25 monitors lol
Although he’d also need 25 monitors lol
Back to the government auctions then!
We’re you thinking like Doom Lan party, or some weird supercluster with the pure focus of running Doom?
But can they run Crysis?
Put a different operating system on each one, and make each a gateway to access the next. See who can make it through.
HungerGamesOS. I love this idea!!!
Do a giant Proxmox cluster. You are going to need one hell of a switch though
Senior year of Highschool, I put Unreal Tournament on the school server. If it were me, I’d recreate that experience, including our teacher looking around the class. That was almost 20 years ago, I hope everyone is doing alright.
I have a box with 10 old laptops that I keep around, just for that. Unreal tournament 2004, Insane, Brood Wars and all the Id classics. I don’t get to set it up a lot, but when I do it’s always a hit.
You could possibly run ai horde if they have enough ram or vram. You could run bare metal kubernetes or inside proxmox.
Distcc, maybe gluster. Run a docker swarm setup on pve or something.
Models like those are a little hard to exploit well because of limited network bandwidth between them. Other mini PC models that have a pcie slot are fun because you can jam high speed networking into them along with NVMe then do rapid fail over between machines with very little impact when one goes offline.
If you do want to bump your bandwidth per machine you might be able to repurpose the wlan m2 slot for a 2.5gbe port, but you’ll likely have to hang the module out the back through a serial port or something. Aquantia USB modules work well too, those can provide 5gbe fairly stably.
Edit: Oh, you’re talking about the larger desktop elitedesk g1, not the USFF tiny machines. Yeah, you can jam whatever hh cards into these you want - go wild.
From the listing photos these actually have half-height expansion slots! So GPU options are practically nonexistant, but networking and storage is blown wide open for options compared to the miniPCs that are more prevalent now.
Yeah, you’ll be fairly limited as far as GPU solutions go. I have a handful of hh AMD cards kicking around that were originally shipped in t740s and similar but they’re really only good for hardware transcoding or hanging extra monitors off the machine - it’s difficult to find a hh board with a useful amount of vram for ml/ai tasks.
I would personally attempt the Kubernetes cluster if I had that many physical machines!
25 machines at say 100W each is about 2.5KW. Can you even power them all at the same time at home without tripping circuit breakers? At your mentioned .12/KWH that is about 30 cents an hour, or over $200 to run them for a month, so that adds up too.
i5-4560S is 4597 passmark which isn’t that great. 25 of them is 115k at best, so about like a big Ryzen server that you can rent for the same $200 or so. I can think of various computation projects that could use that, but I don’t think I’d bother with a room full of crufty old PC’s if I was pursuing something like that.
UK here, we could run that from 1 plug.
Psh 1 plug aint shit. Every Pic I see from anyone who lives out in those ghettos of India, Central America or any spacific islands they also only rock 1 plug but theyre running the corner store, the liquor store, the hospital, their style of little school middle school and old school, 3 hair salons if Latin or 3 nail salons if spasific, Bollywood, every stadium from every country in the world cup, and always 1 dude trying to squeeze 1 more plug in cuz hes runing low bats. Idk why the American ghetto is so pussy. One time i seen a family that fuckin put covers over empty sockets?!? Come on dog thats like wearing a condom jerking off. NGL tho, I get super jelly seeing pictures from those countries thp with their thousands of power lines, phone lines, sidelines, cable lines, borderlines, internet lines… fuck I don’t know much about how my AOL works but those wizards must be streaming some Hella fast Tokyo banddrifts with all them wires.
those wizards must be streaming some Hella fast Tokyo banddrifts with all them wires.
That part is wrong for India, at least.
Here’s a random site with some stats
India, you can expect ~100Mb/s with FTTH and 50Mb/s otherwise. Reliability is even worse.Rest is right.
India has gigabit fiber
And Japan has a 300+ Tb/s connection. Your point?
My point is that the average Indian is not doing “Hella fast Tokyo banddrifts” (not sure what banddrift even means, but no).And yes, a 1Gb/s connection is theoretically available, but how many people are using the ~₹4000/month connection?
Considering how many people tend to just not have Broadband at home, relying just on mobile internet, we can see how things compare with others.
Also, to point to the tread starter, most of the “thousands of” cables that you see on poles in congested areas, are just abandoned cables from older installations which nobody cared to remove.
I’m not the same dude that was talking about banddrifts and congested poles.
Indian, btw.
Also ~100Mb/s is in no way the average speed in an Indian household. It’s usually lower. I also don’t see any specific mentions of india in your link up there to that random site.
Also ~100Mb/s is in no way the average speed in an Indian household.
You’re right. It’s not.
I also don’t see any specific mentions of india in your link up there to that random site.
I don’t see any either. Guess why. Because it only has the top 10, further emphasising the point that :
the average Indian is not doing “Hella fast Tokyo banddrifts”
That’s less than a kettle, in the UK at least.
Of course I wouldn’t want to be running that all the time, because electric ain’t cheap.
This is only about 21 amps. Most outlets in a home are 15amps but 20amps isn’t unheard of. From one outlet doubtful but yes one house would provide that much power easily if you split them up to three or 4 rooms on different breakers.
Now it would be fun to watch his electric meter spin like a saw blade … (yes I’m old … I remember meters that had spinning discs)
Just two 15A breakers is enough actually. Outlets are supposed to be able to sustain 80% power, so you should be able to pull 1.44kW from a singly puny Nema 5-15.
Well true but I was assuming the circuits had some things drawing a little power. Flipping on a device and tripping a breaker with 12 machines on it wouldn’t be ideal :)
I have done this before in my upstairs home lab. 3 beefy ESXi machines, some nas storage, and a basic 10gbe switch eats up a lot of a single 15amp circuit. And apparently turning on a TV pushes it over the edge. Luckily the UPS saved my but while a reset the breaker and shut some stuff off.
deleted by creator
I have a couple of these (only the G2 and G3 SFF) and they consume between 6-10w when not under load, and they max out at 35w (or 65w depending on CPU). I run proxmox with 64gb ram and they are surprisingly efficient.
I won’t be leaving all of them on for long at all. I’ve got a few basically unused 15A electrical circuits in the unfinished basement (can see the wires and visually trace the entire runs) I’ll probably only run all 25 long enough to run a linpack benchmark and maybe run some kind of AI model on the distributed compute then start getting rid of at least half of them
Jack into the local coffee shop
Do you have particularly cheap or free electricity?
12 cents per kilowatt-hour. I certainly don’t plan on leaving more than a couple on long term. I might get lucky with the weather and need the heating though :)
God damn. What are the specs on those? I gotta check out some government auctions.
4th gen intel i5s, 8GB of RAM and 256GB SSDs, so not terrible for a basic Windows desktop even today (except of course for the fact that no supported Windows desktop operating system will officially support these system come Q4 2025)
But don’t get your hopes up, when I’ve bid on auctions like this before the lots have gone for closer to $80 per computer, so I was genuinely surprised I could win with such a low bid. Also every state has entirely different auction setups. When I’ve looked into it in the past, some just dump everything to a third party auction, some only do an in-person auction annually at a central auction house, and some have a snazzy dedicated auction site. Oh and because its the US, states do it differently from the federal government. So it might take some research and digging around to find the most convenient option for wherever you are (which could just be making a friend in an IT department somewhere that will let you dumpster dive)
They’re actually decent. Congratulations!
Shitty k8s cluster/space heater?
Run 70b llama3 on one and have a 100% local, gpt4 level home assistant . Hook it up with coqui.Ai xttsv2 for mind baffling natural language speech (100% local too ) that can imitate anyone’s voice. Now, you got yourself Jarvis from Ironman.
Edit : thought they were some kind of beast machines with 192gb ram and stuff. They’re just regular middle-low tier pcs.
I tried doing that on my home server, but running it on the CPU is super slow, and the model won’t fit on the GPU. Not sure what I’m doing wrong
Sadly, can’t really help you much. I have a potato pc and the biggest model I ran on it was Microsoft phi-2 using the candle framework. I used to tinker with Llama.cpp on colab, but it seems they don’t handle llama3 yet. ollama says it does , but I’ve never tried it before. For the speed, It’s kinda expected for a 70b model to be really slow on the CPU. How much slow is too slow ? I don’t really know…
You can always try the 8b model. People says it’s really great and even replaced the 70b models they’ve been using.
Show as in I waited a few minutes and finally killed it when it didn’t seem like it was going anywhere. And this was with the 7b model…
It shouldn’t happen for a 8b model. Even on CPU, it’s supposed to be decently fast. There’s definitely something wrong here.
Hm… Alright, I’ll have to take another look at it. I kinda gave up, figuring my old server just didn’t have the specs for it
Specs? Try mistral with llama.ccp.
It has a Intel Xeon E3-1225 V2, 20gb of ram, and a Strix GTX 970 with 4gb of VRAM. I’ve actually tried Mistral 7b and Decapoda Llama 7b, running them in Python with Huggingface’s Transformers library (from local models)
These are 10 year old mid range machines. Llama 7b won’t even run well
The key is quantized models. A full model wouldn’t fit but a 4bit 8b llama3 would fit.
It would fit but it would be very slow
No. Quantization make it go faster. Not blazing fast, but decent.
I don’t understand why people want to use so many PC’s rather than just run multiple VM’s on a single server that has more cores.
Having multiple machines can protect against hardware failures.
If hardware fails, you have dono machines.
It’s good learning, both for provisioning and for the physical (cleaning, customising, wiring, networking with multiple nics), and for multi-node clusters.Virt is convenient, but doesn’t teach you everything
I’m not sure if running multiple single SSD machines would provide much redundancy over a server with multiple PSU’s and drives. Sure the CPU or mobo could fail but the downtime would be less hassle than 25 old PC’s.
Of course there is a learning experience in more hardware but 25 PC’s does seem slightly overkill. I can imagine 3-5 max.
I’m probably looking at this from a homelab point of view who just wants to run stuff though, not really as the hobby being “setting up the PC’s themselves”.
Learning.
“I don’t understand why you’d run so many VMs can you can just run it on bare metal”
It’s fun! This is a hobby. It doesn’t have to be practical.
Of course, but installing everything on multiple bare metal machines which take IP addresses, against just running it in VM’s which have IP addresses… It just takes a lot of extra power and doesn’t achieve much. Of course that can be said about any hobby, but I just want OP to know that there is no real reason to do this and I don’t understand so many people hyping it up.
Damn zuck meta is eating you up. Take a breather it’s just for fun. Bro doesn’t have to find the cure for cancer just to poke around on some new hardware
I already said in the original post I plan on sellong off and giving away ~15 of them, keeping a few as spares, and only actually leaving one on 24/7
bare metal machines which take IP addresses, against just running it in VM’s which have IP addresses
Both bare metal and VMs require IPs, it’s just about what networks you toss them on. Thanks to NAT IPs are free and there’s about 18 million of them to pick from in just the private IPv4 space
Big reason for bare metal for clustering is it takes the guess work out of virtual networking since there’s physical cables to trace. I don’t have to guess if a given virtual network has an L3 device that the virtual network helpfully added or is all L2, I can see the blinky lights for an estimate as to how much activity is going on on the network, and I can physically degrade a connection if I want to simulate an unreliable connection to a remote site. I can yank the power on a physical machine to simulate a power/host failure, you have to hope the virtual host actually yanks the virtual power and doesn’t do some pre shutdown stuff before killing the VM to protect you from yourself. Sure you can ultimately do all of this virtually, but having a few physical machines in the mix takes the guesswork out of it and makes your labbing more “real world”
I also want to invest the time and money into doing some real clustering technologies kinda close to right. Ever since I ran a ceph cluster in college on DDR2 era hardware over gigabit links I’ve been curious to see what level of investment is needed to make ceph perform reasonably, and how ceph compares to say glusterFS for example. I also want to setup an OpenShift cluster to play with and that calls for about 5 4-8 core 32GB RAM machines as a minimum (which happens to be the maximum hardware config of these machines). Similar with Harvester HCI
It just takes a lot of extra power and doesn’t achieve much
I just plan on running all of them just long enough to get some benchmark porn then starting to sell them off. Most won’t even be plugged in for more than a few hours before I sell them off
there is no real reason to do this and I don’t understand so many people hyping it up.
Because it’s fun? I got 25 computers for a bit more than the price of one (based on current eBay pricing). Why not do some stupid silly stuff while I have all of them? Why have an actual reason beyond “because I can!”
25 PC’s does seem slightly overkill. I can imagine 3-5 max.
25 computers is definitely overkill, but the auction wasn’t for 6 computers it was for 25 of them. And again, I seriously expected to be out of and the winning bid to be over a grand. I didn’t expect to get 25 computers for about the price of one. But now I have them so I’m gonna play with them
I see I was picturing a 25 pile stack of PC’s this makes a lot more sense thanks for the explanation.
If I had 25 surprise desktops I imagine I’d discover a long dormant need for a Beowulf cluster.
The thought did cross my mind to run Linpack and see where I fall on the Top500 (or the Top500 of 2000 for example for a more fair comparison haha)
- Slurm cluster
- MPI development
