Build a Local LLM with Tesla V100, 32GB VRAM for $266

Efforts to repurpose legacy enterprise GPUs for local large language model inference are attracting attention in the enthusiast community. According to a report by Tom’s Hardware, computing enthusiast Oscar Molnar acquired a noisy, previous-generation Nvidia Tesla V100 SXM2 for $266 (approximately ¥42,000) and added it to an existing gaming PC. This doubled the system’s total VRAM capacity to 32GB, enabling it to run a 27B parameter LLM at 32 tokens per second.

Project Overview

Molnar purchased a Tesla V100 SXM2 module equipped with 16GB of HBM2 memory. On the used market, these are currently available for under $140 on eBay listings shipped from China. However, using this GPU in a standard PC requires an SXM2-to-PCIe adapter. Molnar similarly acquired an adapter for about $66 on eBay.

The existing system was equipped with an Nvidia GeForce RTX 4080 (16GB VRAM, Ada architecture). By adding the Volta-architecture Tesla V100, a total VRAM pool of 32GB was created. Although a 32GB version of the Tesla V100 exists, it costs roughly twice as much, so the 16GB version was chosen.

Noise Issue and Cooling Modification

The standard cooler on the Tesla V100 SXM2 is known for generating extreme noise. Molnar described it as “somewhere between a garbage disposal and a lawnmower” and reported measured levels of 82 dB. To solve this noise problem, Molnar modified the cooling fan wiring by connecting it directly to a motherboard PWM fan header. Specifically, a “2.54mm male to PH2.0 female jumper cable” was used to reroute the wiring. After the modification, the fan could be run at just 10% speed while keeping GPU temperatures below 50°C even under full load.

Software Configuration

Using GPUs with different architectures simultaneously poses driver compatibility challenges. Molnar adopted NixOS and used a legacy Nvidia driver that supports both the Volta and Ada architectures, allowing the system to recognize both GPUs. This setup enabled integrated usage of the VRAM from the RTX 4080 and the Tesla V100.

Inference Performance

When testing local LLMs on the completed system, a 27B parameter model ran at 32 tokens per second. Molnar described this speed as “fast enough for interactive use” and noted it is faster than many cloud API alternatives. 32GB of VRAM is a practical capacity for current local LLM execution, opening up possibilities for quantized larger models or running multiple models simultaneously.

Cost-Effectiveness Evaluation

Total investment was $266 (Tesla V100 under $200 + adapter about $66). In the current climate of memory price inflation (RAMpocalypse), achieving a 32GB VRAM expansion at this price point is noteworthy. For existing RTX 4080 users, adding a Tesla V100 offers a realistic way to double VRAM capacity. However, the Volta architecture of the Tesla V100 lacks some of the latest AI acceleration features (e.g., differences in Tensor Core generations), so potential performance bottlenecks should be considered.

Editorial Opinion

This modification case clearly highlights the importance of VRAM capacity for local LLM execution. In the short term, the used-market price of Tesla V100s may remain high. If more enthusiasts attempt similar projects, demand could push prices even higher. However, the difficulty of the noise modification and driver compatibility issues act as barriers to entry. Over the next 3–6 months, more DIY LLM build cases leveraging similar used enterprise GPUs may be reported. In the long term, this culture of “DIY local LLM building” could influence GPU vendors’ product strategies. It might spur the introduction of consumer GPUs with larger VRAM capacities. On the other hand, the Tesla V100’s inference performance is inferior to that of modern consumer GPUs, so it has limitations as a permanent solution. In 1–3 years, dedicated AI accelerators or edge devices will likely become more realistic options. Our editorial view is that while repurposing enterprise GPUs contributes to LLM democratization, practical challenges such as noise, power consumption, and cooling should not be underestimated.

References

• ” AI enthusiast adds Nvidia Tesla V100 as loud as a lawnmower to gaming PC for $266 — 32GB of VRAM rig can run 27 billion parameter model at 32 tokens per second ”, by Mark Tyson — Tom’s Hardware, 2026-07-26T10:00:00.000Z (ARR)
• Source URL: https://www.tomshardware.com/pc-components/gpus/ai-enthusiast-adds-nvidia-tesla-v100-as-loud-as-a-lawnmower-to-gaming-pc-for-usd266-32gb-of-vram-rig-can-run-27-billion-parameter-model-at-32-tokens-per-second