This is the story of how I bought enterprise-grade AI hardware designed for liquid-cooled server racks that was converted to air cooling, and then back again, survived multiple near-disasters (including GPUs reporting temperatures of 16 million degrees), and ended up with a desktop that can run 235B parameter models at home. It’s a tale of questionable decisions, creative problem-solving, and what happens when you try to turn datacenter equipment into a daily driver.
# Tell the driver to completely ignore the NVLINK and it should allow the GPUs to initialise independently over PCIe !!!! This took a week of work to find, thanks Reddit!
I needed this info, thanks for putting it up. Can this really be an issue for every data center?
True, which is why I said “might”. Even in the US. I only have to call ahead if I want smaller bills - $20 and $100 they usually have plenty of unless it’s a tiny branch.
These are on a custom board from Nvidia, so its not possible to separate them. I think the seller usually gets H100's and them into a custom case, with a PCIE adapter to the server GPUs.
This thing too unwieldy to make into a desktop (you can see how much effort it took), and was in pretty bad condition. I think he just wanted to get rid of it without having to deal with returns. I took a bet on it, and was lucky it paid out.
> why didn't he just fit the two H100s into a better desktop box?
I expect because they were no longer in the sort of condition to sell as new machines? They were clearly well used and selling "as seen" is the lowest reputational risk associated with offload
There also weren't H100s available to scavenge. GH200 puts the Grace CPU and H100 GPU on a big module with a custom form factor and connectors, so the only viable route for using those GPUs was to keep all the electronics together and build a suitable case and cooling system around them. There wasn't any way to adapt any of this for use in an ordinary EATX case or with a different CPU, because the GPUs weren't PCIe add-in cards.
At that pricing I honestly thought they fell off a truck. Even well used H100 go for more than that entire system. In the US an RTX A6000 Ada is already close in price.
We build these desktops from Nvidia servers we buy from reputable manufacturers like Pegatron, Gigabyte, Asrock Rack, and many more.
H100 PCI and GH200 are two very different things. The advantages of Grace Hopper are much higher connections speeds, bandwidth and lower power consumption.
I've had a bit of practice, but I don't have the right gear for this level of soldering. It took maybe an hour to solder in 2 components, after many failed attempts. Persistence beats intelligence?
I recently had a similar experience, although not this size.
Pre-story:
For 3 years I wanted to build a rack-gaming-server, so I can play with my son in our small apartment where we don't have enough space for a gaming computer (wife also doesn't allow it). I have a stable IPsec connection to my parents house, where I have a powerfull PV plant (90kWp) and a rack server, for my freelance job.
Fast forward to 2 months ago, I see a Supermicro SYS-7049GP-TRT for 1400€ on Ebay. It looks clean, sold by some IT reuse-warehouse. No desription, just 3 photos and the case label. I ask the seller whether he knows whats in it and he says he didn't check. The case alone comes new at 3k here in Germany. I buy it.
It arrives. 64GB ECC memory, 2x Xeon silver, 1x 500GB SSD, 5x GBit LAN Cards. Dual 2200 Watt PowerSupply. I remove the airshroud, and: A Nvidia V100S 32GB emerges. I sell the card on ebay for 1600€ and buy 2x Xeon 6254 CPUs (100€ each) to replace the 2x Silver ones that are in it. Last week, I bought two Blackwell RTX 4000 Pro for 1100€ each. Enough for gaming with my son! (and I can do some fun with LLMs and home assistant/smart home..)
The case fits 4x dual-size GPUs, so I could fit 4x RTX 6000 in it (384GB VRAM). At a price of 3k, this would come at 12k (still too much for me.. but let's check back in a couple of years..).
Buying used enterprise gear is fun. I had so many good experiences and this stuff is just rock solid.
Love how a €7.5k 20 kilogram server is placed on a €5 particleboard table. I have owned several LACKs but would never put anything valuable on it. IKEA rates them at 25 kilogram maximum load.
Oh no, thats not right. 20 Kg was in the original server case. With the Aluminium frames, and glass panel, its more like 40 Kg now... Shit, maybe I should take it off the Lack table...
LACK tables specifically are well proven to be quite sturdy actually. They happen to be just the right width for servers / network devices, and so people have used them for that purpose for ages. Search for "LACK rack", or see e.g. https://wiki.eth0.nl/index.php/LackRack. 20kg is nothing; I've personally put >100kg on top.
They're a bit less usable that way now. The legs are basically completely hollow these days so you're not actually able to bear much weight on the screws so the only option is stacking the items so the weight is born by whatever surface is below the "rack" at which point you could just as easily call stacking the equipment an air rack (or an iLackaRack maybe /s).
While this is undoubtably still an excellent deal, the comparison to the new price of H100 is a bit misleading, since today you can buy a new, legit RTX 6000 Pro for about $7-8k, and get similar performance the first two of the models tested at least. As a bonus those can fit in a regular workstation or server, and you can buy multiple. This thing is not worth $80k in the same way that any old enterprise equipment is not worth nearly as much as its price when it was new.
Fair points, but the deal is still great because of the nuances of the RAM/VRAM.
The Blackwells are superior on paper, but there's some "Nvidia Math" involved: When they report performance in press announcements, they don't usually mention the precision. Yes, the Blackwells are more than double the speed of the Hopper H100's, but thats comparing FP8 to FP4 (the H100's can't do native FP4). Yes, thats great for certain workloads, but not the majority.
What's more interesting is the VRAM speed. The 6000 Pro has 96 GB of GPU memory and 1.8 TB/s bandwidth, the H100 haas the same amount, but with HBM3 at 4.9 TB/s. That 2.5X increase is very influential in the overall performance of the system.
Lastly, if it works, the NVLink-C2C does 900 GB/s of bandwidth between the cards, so about 5x what a pair of 6000 Pros could do over PCIE5. Big LLMs need well over the 96 GB on a single card, so this becomes the bottleneck.
The perf delta is smaller than I thought it'd be given the memory bandwidth difference. I guess likely comes from the Blackwell having native MXFP4, since GPT-OSS-120b has MXFP4 MOE layers.
The NVLink is definitely a strong point, I missed that detail. For LLM inference specifically it matters fairly little iirc, but for training it might.
you do realize he has 2 H100s, you would need to buy 2 RTX 6000 Pro for $15-$16k plus the hardware. The ram that came with that hardware is worth more than $7000 now.
I think he is still correct in saying that the gear OP bought is worth much less now and further deteriorating fast. See my comment above here https://news.ycombinator.com/item?id=46227813.
GPUs have such a short liefspan these days that it is really important to compare new vs. used.
Is it? The used data center P40s I bought for $150 2 years ago went back up to $450 a few months ago, I sold one for $400. I just checked and price is down to $200, so I'm still profitable. I bought MI50s for $90 less than a year ago, they are now going for $200. What deterioration? OPs gear was far less and is no longer deprecating. It will probably hold this value for the next 4 years.
Serious question: does this thing actually make games run really great? Or are they so optimized for AI/ML workloads that they either don’t work or run normal video games poorly?
Also:
> I arrived at a farmhouse in a small forest…
Were you not worried you were going to get murdered?
It was fun when the seller told me to come and look in the back of his dirty white van, because "the servers are in here". This was before I had seen the workshop etc.
I believe these gpus dont have direct hdmi/DisplayPort outputs, so at the very least its tricky to even run a game on them, I guess you need to run the game in a VM or so?
Copying between GPUs is a thing, that's how integrated/discrete GPU switching works. So if the drivers provide full vulkan support then rendering on the nvidia and copying to another GPU with outputs could work.
And it's an ARM CPU, so to run most games you need emulation (Wine+FEX), but Valve has been polishing that for their steamframe... so maybe?
People have gotten games to run on a DGX Spark, which is somewhat similar (GB10 instead of GH200)
i did a test with just spamming date in a terminal and having a high fps video captured from my phone, it was usually under a frame (granted 60 fps so 1/60 sec)
Ah, no, that's not what I mean. It's the input devices. Mainly the mouse pointer.
I now remember there was a way to go around it (a bit cumbersome and ugly) which was to render the mouse pointer only locally. That means no mouse cursor changes for tooltips/resizing/different pointers in games, etc. But at least it gets rid of the lag.
I think the point of negative returns for gaming is going above the RTX PRO 6000 Blackwell + AMD 9800X3D CPU + latency optimized RAM + any decent NVMe drive. Seems to net ~1.1x more performance than a normal 5090 in the same setup (and both can be overclocked about equally). Aside from what the GPU is optimized for, the CPU in these servers being ARM based ends up adding more overhead for games (and breaks DRM) which still assume x86 on Windows/Linux.
> does this thing actually make games run really great
It's an interesting question, and since OP indicates he previously had a 4090, he's qualified to reply and hopefully will. However, I suspect the GH200 won't turn out to run games much faster than a 5090 because A) Games aren't designed to exploit the increased capabilities of this hardware, and B) The GH200 drivers wouldn't be tuned for game performance. One of the biggest differences of datacenter AI GPUs is the sheer memory size, and there's little reason for a game to assume there's more than 16GB of video memory available.
More broadly, this is a question that, for the past couple decades, I'd have been very interested in. For a lot of years, looking at today's most esoteric, expensive state-of-the-art was the best way to predict what tomorrow's consumer desktop might be capable of. However, these days I'm surprised to find myself no longer fascinated by this. Having been riveted by the constant march of real-time computer graphics from the 90s to 2020 (including attending many Siggraph conferences in the 90s and 00s), I think we're now nearing the end of truly significant progress in consumer gaming graphics.
I do realize that's a controversial statement, and sure there will always be a way to throw more polys, bigger textures and heavier algorithms at any game, but... each increasing increment just doesn't matter as much as it once did. For typical desktop and couch consumer gaming, the upgrade from 20fps to 60fps was a lot more meaningful to most people than 120fps to 360fps. With synthetic frame and pixel generation, increasing resolution beyond native 4K matters less. (Note: head-mounted AR/VR might one of the few places 'moar pixels' really matters in the future). Sure, it can look a bit sharper, a bit more varied and the shadows can have more perfect ray-traced fall-off, but at this point piling on even more of those technically impressive feats of CGI doesn't make the game more fun to play, whether on a 75" TV at 8 feet or a 34-inch monitor at two feet. As an old-school computer graphics guy, it's incredible to be see real-time path tracing adding subtle colors to shadows from light reflections bouncing off colored walls. It's living in the sci-fi future we dreamed of at Siggraph '92. But as a gamer looking for some fun tonight, honestly... the improved visuals don't contribute much to the overall gameplay between a 3070, 4070 and 5070.
They do still have texture units since sampling 2D and 3D grids is a useful primitive for all sorts of compute, but some other stuff is stripped back. They don't have raytracing or video encoding units for example.
That was enjoyable. I miss the days when I would buy old pieces, or find some in old dumpsters in Sao Paulo and try to use old video cards and memory modules to create little franksteins (a lot cheaper than this, but still fun).
I found interesting to learn there are businesses around converting used servers into desktops. Sounds like a good initiative to avoid some e-waste (assuming the desktops are easy to maintain).
Wow! As others have said, deal of the century!! As a side note, a few years back, I used to scrape eBay for Intel QS Xeon and quite a few times managed to snag incredible deals, but this is beyond anything anyone has ever achieved!
Wow! Kudos for thinking it was possible and making it happen. I was wondering how long it would be before big local models were possible under 10k—pretty impressive. Qwen3-235B can do mundane chat, coding, and agentic tasks pretty well.
I feel like it's going to be a long long time before we get a repeat of something like this. And David did such an incredible job on this. Custom designed frame, designed his own water-block! Wildly great effort here.
This is about more. I can run 600B+ models at home. Today I was having a discussion with my wife and we asked ChatGPT a quick question, it refused because it can't generate the result based on race. I tried to prompt it to and it absolutely refused. I used my local model and got the answer I was looking for from the latest Mistral-Large3-675B. What's the cost of that?
I'm downloading DeepSeek-V3.2-Speciale now at FP8 (reportedly Gold-medal performance in the 2025 International Mathematical Olympiad and International Olympiad in Informatics).
It will fit in system RAM, and as its mixture of experts and the experts are not too large, I can at least run it. Token/second speed will be slower, but as system memory bandwidth is somewhere around 5-600Gb/s, so it should feel OK.
Check out "--n-cpu-moe" in llama.cpp if you're not familiar. That allows you to force a certain number of experts to be kept in system memory while everything else (including context cache and the parts of the model that every token touches) is kept in VRAM. You can do something like "-c128k -ngl 99 --n-cpu-moe <tuned_amt>" where you find a number that allows you to maximize VRAM usage without OOMing.
The author was running a quantised version of GLM 4.5 _Air_, not the full fat version. API pricing for that is closer to $0.2/$1.1 at the top end from z.ai themselves, half the price from Novita/SiliconFlow.
I think there are probably Law Firms/doctors offices that would gladly pay ~3-4K euro a month to have this thing delivered and run truely "on-prem" to work with documents they can't risk leaking (patent filings, patient records etc).
For a company with 20-30 people, the legal and privacy protection is worth the small premium over using cloud providers.
Just a hunch though! This would have it paid-off in 3-4 months?
What an incredible barn-find type story. Incredible. And you are among very few buyers who could have so lovingly done such an incredible job debugging driver & motherboard issues. Please add a kitsch Serial Experiment Lain themed computing shrine around this incredible work, and all's done.
> 4x Arctic Liquid Freezer III 420 (B-Ware) - €180
Quite aside, but man: I fricking love Arctic. Seeing their fans in the new Corsi-Rosenthal boxes has been awesome. Such good value. I've been sing a Liquid Freeze II after nearly buying my last air-cooled heat-sink & seeing the LF-II onsale for <$75. Buy.
Please give us some power consumption figures! I'm so curious how it scales up and down. Do different models take similar or different power? Asking a lot, but it'd be so neat to see a somewhat high res view (>1 sample/s) of power consumption (watts) on these things, such a unique opportunity.
Huge fan of those AIOs as well! I have LFIII 420mm in my PC and I've successfully built a 10x10cm cloud chamber with another one which is really pushing it as far as it can go.
Maybe the title could be I bought an Nvidia server.....
to avoid confusion that it's something to do with Grace Hopper the person, and her servers ...or mainframes?
reply