I Put a Datacenter GPU in My Gaming PC for £200

Author: Oscar Molnar (blog: Tymscar) Published: 2026-05-30 Read time: 13 min (2,710 words)

Summary

Molnar documents installing a used NVIDIA Tesla V100 SXM2 (16GB HBM2) datacenter GPU alongside his existing RTX 4080 to achieve 32GB total VRAM for local LLM inference, all for roughly £200.

Key Hardware Details

The GPU: Tesla V100 SXM2 16GB — a Volta architecture card with 5,120 CUDA cores, 900 GB/s memory bandwidth over a 4096-bit HBM2 bus. Purchased on eBay for ~£150.

Bandwidth comparison (article's data): - V100: 900 GB/s (2017) - RTX 4080: 736 GB/s - M3 Max: 400 GB/s | M4 Max: 546 GB/s | M5 Max: 614 GB/s - RX 7900 XTX: 960 GB/s (~£700+) - RTX 5090: 1,792 GB/s (~£2,000+)

The Adapter: An SXM2-to-PCIe adapter PCB (~£50) that allows the proprietary-connector card to slot into a standard motherboard.

The Fan Issue

The card's fan measured 82 dB — described as between "a garbage disposal and a lawnmower." It runs at 100% uncontrollably via software. The author traced the pinout using jumper wires and a 9V battery, confirmed it's a standard PWM fan on a JST PH2.0 connector, then wired it to a motherboard fan header via a £2 adapter cable. At 10% PWM, "it never goes above 50C even at full load."

Software Configuration (NixOS)

Critical constraints: NVIDIA dropped Volta support in driver branch 560, so the author uses branch 550.x (legacy_535), which requires kernel 6.6 and only supports CUDA 12.2 (backported from nixpkgs 24.05). A quirk: services.xserver.enable = true is required even for a headless setup, or NVIDIA kernel modules won't load.

Driver/kernel config

boot.kernelPackages = pkgs.linuxPackages_6_6;
hardware.nvidia.package = config.boot.kernelPackages.nvidiaPackages.legacy_535;
services.xserver.enable = true;
services.xserver.videoDrivers = [ "nvidia" ];

CUDA 12.2 overlay

nixpkgs.overlays = [
  (final: prev: {
    cudaPackages_12_2 = nixpkgs-cuda.legacyPackages.${prev.system}.cudaPackages_12_2;
  })
];

NFS mount for models

fileSystems."/mnt/nas" = {
  device = "truenas-nfs.tymscar.com:/mnt/oasis/services";
  fsType = "nfs";
  options = [ "nfsvers=4" "_netdev" "auto" "nofail" ];
};

Llama.cpp vision flags

--mmproj /mnt/nas/llamacpp/mmproj-F16.gguf --mmproj-offload

The full machine definition is linked: GitHub dotfiles commit.

Running the Model

With Multi-Token Prediction (MTP), speed can reach 50–60 tok/s on predictable output like code. The author had to compile llama.cpp from source at a specific commit for MTP support.

Model Quality Claim

Molnar asserts the Qwen3.6-27B "ties with Claude Sonnet 4.6 on Artificial Analysis's Agentic Index" and beats Sonnet 4.6 on MMMU-Pro and Terminal-Bench 2.0, while trailing on GPQA and SWE-Bench Verified. His conclusion: "the model you run in your bedroom is in the same conversation as the ones that charge you per token."

Vision Support

The Qwen3.6-27B model accepts image input via a ~928MB multimodal projector (mmproj). A vision encoder "compresses the image into a sequence of vectors that live in the same mathematical space as text tokens."

System Architecture

• Dual-boot via USB: NixOS runs from a Corsair MP600 MINI in a USB-C NVMe enclosure. To game, the author unplugs the drive and boots Windows (4080 only). To use LLMs, plugs it back in and reboots.
• Models stored on TrueNAS mounted over NFS; llama.cpp waits for the mount before starting.
• Remote access via OpenCode over Tailscale.

Annoyance

The V100 occasionally disappears from lspci/nvidia-smi after a warm reboot — an ACPI enumeration issue. A cold power cycle always fixes it.

Cost Breakdown & Alternatives

Alternatives mentioned: P40 (24GB, slower, no tensor cores) and V100 32GB variant (more expensive but still undercuts consumer cards). The author notes: "the 7900 XTX costs £700+ and ROCm support for LLM inference is still rough compared to CUDA."