Tesla V100 for local AI: a £200 server GPU in a gaming PC, and whether it's worth copying

A hobbyist got to 32GB of VRAM for about £200 by bolting a datacenter GPU onto a gaming PC. The card is a used Tesla V100 SXM2 16GB, wired through a bare adapter board and paired with an RTX 4080 that keeps doing display duty. The rig runs a 27-billion-parameter model at about 32 tokens per second. The adapter's fan, as shipped, measured 82 decibels. We read the build log and NVIDIA's documents on 4 October 2026; we did not build this ourselves.

Is a Tesla V100 worth it for local AI?

For most people, no. For a tinkerer who already has a modern GPU driving the monitor and wants more VRAM cheaply, it can be. The V100's memory bandwidth is excellent for LLM generation. What you give up is support: an unofficial adapter, no video output, an aging driver stack and loud cooling.

The V100 launched in 2017 as NVIDIA's flagship datacenter accelerator. NVIDIA's datasheet lists 16GB or 32GB of HBM2, 900 GB/s of memory bandwidth, 5,120 CUDA cores and 640 Tensor Cores. The SXM2 version plugs into NVIDIA's server board socket, not a PCIe slot, and is rated at 300W maximum.

Bandwidth matters because generating each token means reading the model's weights. At 900 GB/s, the V100 sits close to an RTX 3090's 936 GB/s. That is why a nine-year-old card is still useful for inference.

What the build actually took

The builder bought the V100 SXM2 16GB for about £150 on eBay and an SXM2-to-PCIe adapter for about £50. In his words, the adapter "is not made by NVIDIA. It is not officially supported by anyone." It is a bare board with the SXM2 socket on one side and a PCIe edge on the other.

The V100 ran next to his RTX 4080, which drove the display. He split Qwen3.6-27B across both cards evenly with llama.cpp. The model file at Q5_K_M was about 19GB, and he ran it with a 128k-token context.

How fast is a V100 for LLMs?

In this build, about 32 tokens per second generation and 133–160 tokens per second prompt processing on Qwen3.6-27B at Q5_K_M, split across the V100 and an RTX 4080. That is one person's two-GPU setup, not a V100-only benchmark, so treat it as a real-world data point rather than a spec.

The builder reported the V100 drawing about 150W at most, half of its 300W rating. He also called the model "genuinely competitive" with cloud models, citing a benchmark ranking. That is a claim about the model, not the card, so keep the two apart.

Specs from NVIDIA's V100 datasheet and RTX 3090 spec page, and CUDA support from the CUDA 13.4 release notes, all read 4 Oct 2026. The 3090's 936 GB/s comes from our GPU spec database; NVIDIA's page lists its 24GB GDDR6X on a 384-bit bus.

The catches nobody puts in the headline

• No display output. The V100 is compute-only, so a second GPU has to drive your monitor.
• CUDA 13 dropped Volta. NVIDIA's CUDA 13 release notes say support for "Maxwell, Pascal, and Volta GPUs" was removed. The current toolkit is 13.4, so a V100 keeps you on CUDA 12.
• The driver is a balancing act. The builder writes that NVIDIA dropped Volta from driver branch 560, so he used the 550 branch (NixOS calls the package legacy_535) with CUDA 12.2 and Linux kernel 6.6. That combination had to support his RTX 4080 too.
• Noise. The adapter's fan measured 82 dB on his Apple Watch. He fixed it by wiring the fan to a motherboard header for speed control.
• Quirks. The V100 sometimes vanished after a warm reboot and needed a full power-off to reappear, per the log.
• No warranty. It is decommissioned server hardware with unknown hours.

Should you buy a V100 or a used RTX 3090?

Buy the RTX 3090 if you want one card that just works: 24GB, a display output, current drivers and CUDA 13 support. Buy the V100 only if you already have a modern GPU, want the cheapest extra VRAM, and are happy running an older CUDA stack.

There are other cheap server cards with similar trade-offs. Our pages on the Tesla P40 and the AMD Instinct MI50 32GB cover them. For the mainstream route, start with the used RTX 3090 buying guide or the best used GPU for local AI.

V100s, P40s and other old server cards keep turning up in home rigs as GPU prices climb, and the memory shortage behind those prices is part of why. Before you buy any of them, check how much VRAM you actually need and whether local AI is worth it for your use.