Enthusiast Adds Tesla V100 to Local LLM Rig
In a May 30 build log, Oscar Molnar described adapting a used 16GB Tesla V100 SXM2 to run beside an RTX 4080, producing 32GB of combined VRAM for about £200. The mixed-GPU system reportedly ran a 27-billion-parameter model at 32 tokens per second, though it required legacy drivers, an adapter, and custom fan control.
A hardware enthusiast has documented an unconventional route to more memory for local model inference: pairing a used 16GB Tesla V100 SXM2 with an existing 16GB RTX 4080.
Oscar Molnar published the build log on May 30, 2026. He reported spending about £150 on the V100 and about £50 on an SXM2-to-PCIe adapter, bringing the added hardware cost to roughly £200. A July 26 Tom's Hardware report, retrieved through its Yahoo syndication, converted that total to about $266.
What the build achieved
Using llama.cpp to split model layers across the two GPUs, Molnar reported 32GB of combined VRAM and 32 tokens per second on a 27-billion-parameter model. He also reported a 128k-token context window and local image-input support for the tested setup.
The result is a first-person benchmark, not an independently reproduced comparison. Combining memory across two different GPU architectures also does not provide the same performance or simplicity as a single 32GB card.
The hidden engineering cost
The V100 SXM2 was designed for servers, not a consumer PCIe slot. The build required a third-party adapter and a cooling modification after the adapter fan measured 82 dB. Molnar rewired fan control to a motherboard header and reported keeping the card below 50°C at full load with the fan at 10%.
Software compatibility imposed another constraint. The Volta-era V100 and Ada-era RTX 4080 needed an overlapping legacy Nvidia driver, an older CUDA release, and a Linux configuration tailored to that stack.
For local-AI builders, the useful takeaway is narrower than “old server GPUs are cheap.” Used accelerators can add memory capacity at low acquisition cost, but the real budget must include power, cooling, adapters, driver compatibility, reliability, and time spent integrating the system. The reported throughput is evidence about this particular rig, not a general benchmark for every V100 setup.
Key Points
- 1The May 30 build paired a used 16GB V100 SXM2 with a 16GB RTX 4080 for about £200 in added hardware.
- 2The author reported running a 27-billion-parameter model at 32 tokens per second across 32GB of combined VRAM.
- 3The setup required an SXM2-to-PCIe adapter, custom fan control, legacy drivers, and an older CUDA stack.
Scoring Rationale
A useful, reproducible first-person hardware case study with independent reporting, but its performance and economics apply to one customized mixed-GPU rig.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
