Skip to content
Tech News
← Back to articles

Homelab builder runs Qwen 3.6 35B at 100 tokens/sec on three used Nvidia V100s

read original get Nvidia Tesla V100 16GB GPU → more articles
GoKawiil Brief

A homelab builder tested nine GPU configurations—using one to three decommissioned Nvidia K80, P100, or V100-16GB cards on a shared Supermicro dual-Xeon board—running over 130 LLM benchmarks including Qwen 3.6, Qwen 2.5, and Gemma 4. The three-V100 setup, costing roughly $1,042 for the GPUs alone, hit about 100 tokens per second on Qwen 3.6 35B, while the smaller Qwen 3.6 27B reached about 24 tokens per second; V100s also used roughly half the energy of an equivalent P100 setup.

Why It Matters

GoKawiil's interpretation of the reporting above, not reported fact.

The results suggest that retired enterprise GPUs, often dismissed as e-waste, can still deliver usable large-language-model performance well above the builder's 70 tokens-per-second 'realtime interaction' threshold at relatively low cost. This could offer hobbyists and small labs a cheaper entry point into local LLM hosting, though the builder notes that scaling old hardware has clear limits compared to modern GPUs.

Key Takeaways
Worth a Look

Nvidia Tesla V100 16GB GPU — This is literally the card at the center of the homelab benchmark, prized by DIY AI builders for its surprisingly strong tokens-per-second performance relative to cost and power draw. If you're looking to replicate this kind of budget LLM inference rig, grabbing a used V100 is the most direct way to start.

See Nvidia Tesla V100 16GB GPU on Amazon → Affiliate link — we may earn a commission on purchases, at no extra cost to you. Product picked by AI based on this article; it is not a tested recommendation.

Source: tomshardware.com — Shane Downing, 2026-11-19

Published there as: “Three 'e-waste' Nvidia V100s hit 100 tokens per second on Qwen 3.6 35B in a $1,042 homelab rig”

Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.
Keep reading
llama.cpp 14 briefs → homelab 4 briefs → nvidia v100 4 briefs → qwen 3.6 4 briefs →