YouTuber Jun Yamog recently showcased a custom quad-GPU workstation utilizing four NVIDIA Tesla V100 units on an NVLink carrier board. The setup features modified A100 heatsinks to manage 300W per chip, all connected via a PCIe x8 link with passthrough adapters.
The 2017-era Volta architecture remains potent, with each V100 offering 32GB of HBM2 memory and 900GB/s bandwidth. Benchmarks using Qwen3.8-Flash-Next show the cluster hitting 100.6 tok/s, outperforming the M5 Ultra and dual DGX Spark systems in prompt tasks.
This project highlights the value of legacy hardware, as a 32GB V100 currently costs $744 on eBay. The configuration proved 230 percent faster than an M5 Ultra for long prompts, proving that high-bandwidth NVLink setups still rival modern unified memory chips.