NVIDIA Tesla V100 SXM2 Modified for Desktop PCs Runs Local LLMs

Hardware Haven demonstrated a method to repurpose NVIDIA Tesla V100 SXM2 accelerators for standard desktop computers. The project involves adapting a second-hand server-grade GPU to fit into a conventional PCIe slot. This modification allows users to run local large language models on consumer hardware that lacks specialized data center infrastructure.

Active cooling replaces passive solution with Noctua fan

The modification requires replacing the card's passive cooling with an active solution. Builders attach an 80mm Noctua fan and a custom 3D printed shroud to the unit. The card contains 16GB of HBM2 memory and connects via an adapter that bridges the SXM2 interface to standard PCIe.

Spec comparison

Performance testing shows the modified V100 significantly outperforms the GeForce RTX 3060 12GB in AI inference tasks. The V100 achieves 108 tokens per second when running Gemma 4 E4B, compared to 76 tokens per second for the RTX 3060. Even when power is capped at 100W, the V100 maintains 95 tokens per second, while the RTX 3060 drops to 68 tokens per second under the same limit.

The total cost for the used V100, adapter, and cooling components is approximately 1700 CNY, which equals about $249. A separate estimate for the GPU and adapter combination alone places the price at roughly 200 USD. This pricing makes the setup a budget alternative for local AI experimentation.

NVIDIA has not confirmed the viability of this specific modification or its long-term stability. The DIY cooling solution and driver debugging process present technical barriers for average users. Hardware Haven notes that system recognition requires specific driver adjustments that may not be stable for sustained heavy loads.