May 31, 2026
Server chip, main character energy
I Put a Datacenter GPU in My Gaming PC for £200
He turned a bargain server part into a monster PC — and the comments instantly split
TLDR: A hobbyist used a cheap old server graphics chip and an unofficial adapter to boost his home PC for AI work at a tiny fraction of the usual price. Commenters were torn between calling it an incredible bargain and mocking it as a loud, impractical science project for people who enjoy suffering.
A PC tinkerer bought an old data-center graphics chip for about £150, added a weird adapter, and crammed it into a home gaming computer to get 32GB of video memory total alongside an existing card. In plain English: he found a scrappy way to run much bigger local AI models without paying £2,000+ for a shiny top-end card. The crowd’s reaction? Equal parts "genius bargain hack" and "absolutely not in my house." The biggest gasp wasn’t even the price — it was the noise. This thing reportedly screamed at 82 decibels, which commenters basically treated as the sound of financial savings physically taking revenge.
The hottest praise came from the value hunters. One commenter called the capability-per-pound absurd, which is basically the nerd equivalent of saying this is a clearance-bin heist. But the skeptics showed up fast. Some rolled their eyes at the dramatic writing style, with one snarking, "Because humans write exactly like this", while others said the whole project felt like "local copium" — fun, sure, but not exactly normal-person-friendly. Another commenter jumped in with context that sounded like a reality check: yes, this old chip used to cost around $10,000, but there’s a reason it’s cheap now — it’s awkward, weird, and meant for giant server racks, not bedrooms.
And then came the true comments-section energy: escalation. If this Frankenstein setup impressed you, one person casually raised the stakes by suggesting even bigger ex-server chips with 128GB of memory, only to admit they’re even harder to plug in. So the vibe is deliciously chaotic: brilliant hack, cursed machine, terrible roommate. The article brought the hardware, but the comments brought the popcorn.
Key Points
- •The author added a used NVIDIA Tesla V100 SXM2 16GB datacenter GPU to a gaming PC with an unofficial SXM2-to-PCIe adapter for about £200 total.
- •The resulting system combined the V100 with an existing RTX 4080 to provide 32GB of total VRAM across two GPUs.
- •The article reports running a 27 billion parameter model locally at 32 tokens per second with this setup.
- •The V100's 16GB HBM2 memory and 900 GB/s memory bandwidth are presented as key advantages for LLM inference despite the GPU's age.
- •A major practical issue was cooling noise: the adapter fan measured 82 decibels and could not be controlled through standard software tools.