VIDEO DOCUMENTEDSOURCE CHECKED
Can a $200 Tesla V100 beat newer consumer GPUs for local AI?
Hardware Haven adapted a 16 GB Tesla V100 SXM2 accelerator for a normal PCIe system, added custom cooling, and compared local AI speed and power use with an RTX 3060 and Radeon RX 7800 XT.
Video by Hardware Haven · Reviewed August 22, 2026The video remains on YouTube. It belongs to its creator, and LocalLens does not host a copy.
What this evidence supports
The video is useful evidence that an inexpensive used V100 can deliver strong local inference speed in this particular Ollama setup. It is not a general recommendation to buy one. The adapter, cooling, display, power, driver, and software-support limits make it an expert project rather than a simple upgrade.
Medium to high confidenceThis is not a LocalLens reproduction.
Computer setup
- NVIDIA Tesla V100 SXM2, 16 GB HBM2
- SXM2 to PCIe adapter
- Custom 3D-printed cooling duct and 80 mm fan
- RTX 3060 12 GB comparison
- Radeon RX 7800 XT 16 GB comparison
Model and engine
- Ollama
- gpt-oss-20b
- Gemma 4 E4B
- Exact model tags and quantization were not available in the reviewed material
Results recorded in the video
V100 with gpt-oss-20bAbout 130 tokens per secondReported in video
RX 7800 XT with gpt-oss-20bAbout 90 tokens per secondReported in video
V100 with Gemma 4 E4BAbout 108 tokens per secondReported in video
RTX 3060 with Gemma 4 E4BAbout 76 tokens per secondReported in video
V100 at 100 W GPU limitAbout 95 tokens per second, 170 W at the wallReported in video
RTX 3060 at 100 W GPU limitAbout 68 tokens per second, 171 W at the wallReported in video
What the evidence supports
- NVIDIA lists the V100 SXM2 with 16 or 32 GB HBM2, 900 GB/s memory bandwidth, and a 300 W maximum power rating.
- Tom's Hardware independently reported the same video setup and the same approximate performance and power measurements.
- The V100 has no normal display output in this adapted configuration and needs a separate display solution.
What we cannot claim
- LocalLens did not reproduce the measured token speeds.
- The exact CPU, RAM, operating system, Ollama version, model tags, quantization, and context settings were not confirmed.
- Performance may change with drivers, model revisions, cooling, power settings, and Ollama versions.