Updated

Updated · Wccftech · Jul 26

Modder Adds 32GB Tesla V100 to Gaming PC, Running 27B AI Model at 32 Tokens/Second

Updated

Updated · Wccftech · Jul 26

Modder Adds 32GB Tesla V100 to Gaming PC, Running 27B AI Model at 32 Tokens/Second

Q: Can a $100 datacenter GPU really turn a gaming PC into a 32GB local AI box—or do cooling and driver hacks ruin the bargain?

Yes—if you already own a capable main GPU and are comfortable tinkering, a used $100 Tesla V100 can meaningfully turn a desktop into a 32GB-class local AI machine. But the bargain holds only for hobbyists willing to accept real complexity. The economics are compelling. Tymscar reportedly spent about £200 total on the SXM2 V100 plus the SXM2-to-PCIe adapter, far below the price of modern high-VRAM consumer cards. The V100 still offers strong inference-relevant specs: 16GB HBM2, roughly 900GB/s memory bandwidth, and CUDA support that remains useful for llama.cpp and similar local LLM tools. In Tymscar’s setup, paired with an RTX 4080, it delivered 32GB of usable VRAM and ran a 19GB Qwen3.6-27B-MTP quant at about 32 tokens per second with a 128K context. The catch is that this is not plug-and-play. SXM2 cards need an adapter, custom mounting, and custom airflow. Cooling is the biggest practical risk: these datacenter parts were designed for server airflow, not quiet desktop use. Reports of 82dB fan noise, voltage tweaks, PWM workarounds, and even 3D-printed ducts show that noise and thermals can easily erase the “cheap and easy” appeal. Drivers are the other caveat. Mixed-generation setups such as a Volta V100 plus an Ada RTX 4080 can require older NVIDIA driver branches, and Windows support can be less predictable than Linux. So the bargain is real for offline AI inference, but only if you price in noise, power draw, driver friction, and modding time.

1 articles · Updated · Wccftech · Jul 26

Tymscar fitted an NVIDIA Tesla V100 into an RTX 4080-based gaming PC, giving the system 32GB of usable VRAM for local AI workloads rather than gaming.
A £200 setup—about $266 for two 16GB V100s and an SXM2-to-PCIe adapter—solved the VRAM bottleneck that had limited larger language models on the existing GPU.
The modified system ran Qwen3.6-27B-MTP quantized at 19GB with a 128K-token context at 32 tokens per second, while prompt processing reached 133 to 160 tokens per second.
Getting there required custom cooling and power work because the Tesla V100 lacks a standard PCIe slot, display outputs and PCIe power connectors; its stock cooler also hit 82 dB before fan-speed tweaks.
The build highlights a low-cost path to offline home AI inference, using older data-center GPUs that can still deliver 900GB/s bandwidth for small to mid-sized models.