A homelab enthusiast has demonstrated that decommissioned Nvidia V100 GPUs can effectively run large language models like Qwen 3.6 35B at impressive speeds. By using three used V100s in a custom $1,042 rig, the builder achieved approximately 100 tokens per second, surpassing the performance of a $5,399 M5 Max Mac Studio at context lengths of 64K and beyond. This setup utilizes 'e-waste' GPUs, which are older enterprise cards, and highlights their cost-effectiveness and surprisingly competitive performance for certain LLM tasks, especially when compared to modern, high-priced alternatives. AI
IMPACT Demonstrates cost-effective hardware solutions for running LLMs, potentially lowering the barrier to entry for advanced AI experimentation.
RANK_REASON Article details a specific hardware configuration for running LLMs, not a new model release or frontier research.
- Corsair HX1500i
- eBay
- Gemma 4
- Intel Xeon E5-2687W
- M5 Max Mac Studio
- Nvidia K80s
- Nvidia P100s
- Nvidia V100s
- Qwen 2.5
- Qwen 3.6 27B
- Qwen 3.6:35B
- Supermicro
- Tom's Hardware
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →