TURN A 10,000 DOLLAR DATACENTER GPU INTO A HOME AI SERVER
A 100-dollar eBay V100 beats a modern desktop at running local AI, once you solve the cooling and the socket.
by Hardware Haven
Built with3D printing
- difficulty
- ●●●●○
- time
- a weekend
- license
- license not specified
- repo
- repo
●●●●○ · a weekend · license not specified · repo
WHAT YOU’LL NEED
Jump to section
COMPAREE VERDICT
This is a clever way to get serious AI inference power for about 235 dollars, but it only works if you are comfortable with legacy drivers, higher idle power draw than modern cards, and hunting for the right eBay listing. Hardware Haven bought the 16GB SXM2 version for 100 dollars, added a 100-dollar adapter board to convert the datacenter socket to PCIe x16, then designed and printed a cooling shroud for an 80 mm Noctua fan. The card ran a 20 billion parameter model at roughly 130 tokens per second in his tests, faster than his desktop, and also handles Frigate object detection for home security cameras. The single thing most likely to go wrong: buying the wrong V100 variant. The SXM2 module has no built-in cooling or power connectors, so you need the adapter and the shroud. The PCIe version of the V100 has both, but costs more used. If you want plug-and-play, this is not it. If you want the cheapest path to 16GB of VRAM for local LLMs and you are fine troubleshooting driver quirks, this is a weekend well spent.
IN THE REPO
GOOD TO KNOW
- —The 3D printed cooling shroud STL is in the video description.
- —There is no written BOM — you watch the video and note the parts as he goes.
- —No firmware: this is a hardware build that uses standard Nvidia drivers.
- —Driver support for the V100 is legacy only, so newer CUDA features may not work.
- —No repository — this is a YouTube build guide.
- —The build uses a commercial adapter board (SXM2 to PCIe x16, roughly 100 dollars) that you buy, not make.
Parts to buy
4 itemsFrom our check of the build. Exact quantities and part numbers are in the creator’s BOM.
Can I build this?
Build at your own risk. Projects involve tools, electronics and sometimes mains voltage — follow the creator’s safety notes.
Videos
This Ridiculous $200 AI GPU Shouldn't Be This Good
Hardware Haven walks through the entire build: buying the card, the adapter, designing the shroud, and running benchmarks. The STL is in the description.
More builds like this
All projectsGallery
Start here
Navigation into the creator’s own docs — we don’t rewrite the guide, we route you to the source.
- 1.Watch the video to understand what you are buying (The first 5 minutes explain the V100 variants and why the SXM2 version needs an adapter.)
- 2.Source a V100 SXM2 16GB on eBay(Prices vary. Hardware Haven paid 100 dollars, but check that it is the SXM2 module, not the PCIe card.)
- 3.Buy an SXM2 to PCIe x16 adapter(Roughly 100 dollars. Search for 'SXM2 adapter' on AliExpress or eBay. Verify compatibility before ordering.)
- 4.Print the cooling shroud (STL is in the video description. Sized for an 80 mm fan.)
- 5.Install and test with legacy Nvidia drivers(Driver support is legacy only. Expect some trial and error with CUDA versions.)
KNOWN ISSUES
- Buying the wrong V100 variant. The PCIe version has built-in cooling and power; the SXM2 module does not. If you buy SXM2, you need the adapter and the printed shroud.
- Underestimating power draw. The V100 uses more power at idle than a modern RTX 3060, so your PSU and cooling need headroom.
- Assuming driver support is current. The V100 is on legacy-only Nvidia drivers, so some newer CUDA features and frameworks may not work.
- Overpaying for the adapter. Shop around — prices for SXM2 adapters vary widely, and some sellers charge twice what Hardware Haven paid.
- Expecting plug-and-play performance. The benchmarks in the video are small tests. Larger models or multi-GPU setups will expose other bottlenecks.
- Ignoring case airflow. The 80 mm fan cools the GPU, but if your case has poor airflow the whole system will throttle.
Can I use the 32GB version of the V100?
Yes, if you can find one at a reasonable price. The build is the same, but 32GB modules are rarer and cost more used.
Will this work with any PCIe x16 slot?
Yes, as long as your motherboard and PSU support it. The adapter converts the SXM2 socket to standard PCIe x16.
How does this compare to a modern GPU for AI?
The V100 has more VRAM than most consumer cards at this price, which helps with large models. But it uses more power, runs hotter, and has legacy driver support. A modern card like an RTX 3060 is more efficient and easier to set up.
Do I need the exact fan Hardware Haven used?
No, any 80 mm fan that fits the shroud will work. Noctua fans are quiet and reliable, but not required.
Can I run this in a laptop?
No. This is a full-size PCIe card with significant power and cooling requirements. Desktop only.
Community builds
No community builds yet — be the first, we feature the best ones.
Discussion1
FROM THE COMPAREE TEAM
Hardware Haven got his V100 for 100 dollars and the adapter for another 100 — but eBay prices move fast. What are you seeing for SXM2 modules right now?
Hardware Haven
Hardware Haven is a YouTube channel that builds practical, budget-focused hardware projects. This V100 build came from wanting serious AI inference power without spending thousands on a new GPU.
DISCLAIMER
- Comparee is not the author of the projects featured here. All rights to each project belong to its creator — every page links to the original source, and we never host creators’ files.
- Information is provided without warranty and may become outdated as projects evolve. Prices are indicative bands only — always check the creator’s parts list for current costs.
- Building and operating any project is at your own responsibility. Protective equipment, safe workshop practice and compliance with local regulations are the builder’s responsibility.
CompareeTEAM4d ago
Practical notes from our verification: the video is the entire source — there is no repository, no written BOM, and no official project page. The STL for the cooling shroud is in the YouTube description. The biggest decision is whether to buy the SXM2 module (cheaper, needs adapter and printed cooling) or the PCIe version (more expensive, plug-and-play). The 16GB version limits which models you can run locally, and legacy driver support means some newer frameworks will complain. If you have never built a PC before, start with something simpler; if you have, this is a clever way to get 16GB of VRAM for inference without spending four figures.