← All stories
● Covered by 1 source · 1 reportLow impact1 neutral

Homelab Achieves 100 Tokens/Second on Qwen 3.6 35B with Three Used Nvidia V100 GPUs

🔄 Updated 2d ago
New to BrevFeed? We gather this story from every outlet covering it into one summary — ranked by real-world impact, not just the latest headline — so you never miss what matters. What is BrevFeed? →

Key points

  • Three used Nvidia V100 GPUs achieved 100 tokens/second on Qwen 3.6 35B.
  • The homelab setup cost $1,042, using "e-waste" GPUs.
  • Qwen 3.6 35B requires multiple GPUs for Q8 quantization.
  • V100s use half the energy of P100s for the same task.

Homelab Benchmarks Decommissioned GPUs

A homelab builder conducted benchmarks using three decommissioned Nvidia V100 GPUs to run the Qwen 3.6 35B language model. The setup, which cost $1,042, achieved a performance of approximately 100 tokens per second. This project explored the potential of using older, less expensive enterprise hardware for local large language model (LLM) inference.

Testing Methodology and Hardware

The benchmark involved nine different configurations, utilizing one to three K80s, P100s, or V100-16GB cards on a Supermicro dual-Xeon board. The test platform included a Supermicro X10DRG-Q with two Intel Xeon E5-2687W 12-core CPUs and a Corsair HX1500i PSU. The testing used llama.cpp and covered over 130 LLM tests, including Qwen 2.5, Gemma 4, and Qwen 3.6, across various precisions (Q4/Q8/F16) and context sizes (4K to 128K). Energy consumption per run was also recorded.

The term "e-waste" refers to these decommissioned enterprise GPUs, which are more affordable than current alternatives. The V100s were tested as a follow-up to a previous round involving 15 "e-waste" GPUs. The builder noted that scaling decommissioned GPUs has limitations.

Performance and Model Specifics

Performance was evaluated based on a "realtime interaction" threshold of 70 tokens per second. The Qwen 3.6 35B model, a mixture-of-experts (MoE) model, requires three GPUs for Q8 (8-bit) quantization, as it fails on one or two GPUs. The model activates about 3 billion of its 35 billion parameters per token. The 100.8 tokens per second performance was achieved with a 4K context window, decreasing to 72.7 tokens per second at 12K context.

Cost and Practical Considerations

The $1,042 cost for the setup includes the GPUs and core components, though it omits RAM, storage, case, and cooling. The older cards are inexpensive partly because they lack integrated cooling, necessitating custom solutions like 3D-printed blower ducts. The V100s also demonstrated better energy efficiency, using about half the energy of an equivalent number of P100s for the same task.

✨ This summary was generated by AI from the outlets' reporting listed below. It is not independently verified and may contain errors — check the original sources. How BrevFeed works →

The daily brief

One email each morning: the day's tech stories, clustered across outlets and summarized. No account needed.

One email a day. Unsubscribe in one click, any time.

Today's brief

Spend a few minutes, get the whole day. Every topic's top stories in one hands-free rundown — listen, watch, or read the transcript.

~4 min · 3 stories · Nov 21

▶ Play today's brief Listen on Spotify

New every morning, and the back catalogue is archived by date.

Reporting from

A homelab builder benchmarked three used Nvidia V100 GPUs on Qwen 3.6 35B, reaching approximately 100 tokens per second in a custom setup costing $1,042. This setup demonstrates the viability of using decommissioned enterprise hardware for local LLM inference, particularly for models like Qwen 3.6 35B that require multiple GPUs.