Three 'e-waste' Nvidia V100s hit 100 tokens per second on Qwen 3.6 35B in a $1,042 homelab rig
Reported by Tom's Hardware and Mezha2 sources ↓
A homelab builder tested three used Nvidia Tesla V100 GPUs running Qwen 3.6 35B, reaching 100.8 tokens per second at a 4K context in a build priced at $1,042.
Key points
- The V100 setup recorded 72.7 tokens per second at a 128K context
- The $1,042 total excludes RAM, storage, the case and cooling
- Three P100s reached 55.2 tokens per second at 4K on the same model
- The builder tested nine GPU configurations with llama.cpp across more than 130 LLM tests
- The V100s used about half the GPU energy of the same number of P100s for a 128K run
Sources
Details
Tom's Hardware reports that the benchmark used a Supermicro X10DRG-Q board with two Intel Xeon E5-2687W processors, keeping the board and CPUs constant across nine builds. The $1,042 V100 total counts one CPU and omits RAM, storage, the case and cooling; the builder said the listed prices may not match current eBay prices.
For Qwen 3.6 35B at Q8 quantization, the three V100s recorded 100.8 tokens per second at 4K context, declining to 72.7 at 128K. Three P100s reached 55.2 at 4K and 40.2 at 128K. The 35B model is a mixture of experts that activates about 3 billion parameters per token; the dense Qwen 3.6 27B ran at 24.3 tokens per second at 4K on the V100 setup.
The benchmark used llama.cpp and measured more than 130 tests across models, quantizations and context sizes. The builder also reported about 0.02 kWh of GPU energy for a 128K Qwen 3.6 35B run on the V100s, versus roughly twice that for the P100s; those energy figures cover the GPUs, not the full system.
Written from coverage as of
Images
- Tom's Hardware
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh