Four Tesla V100 GPUs on NVLink Beat M5 Ultra and DGX Spark by Up to 230% in AI Tests

YouTuber Jun Yamog built a system with four 9-year-old Nvidia Tesla V100 GPUs on an NVLink carrier board delivering 300GB/s. Running Qwen3.8-Flash-Next at 4-bit quantization, the setup was up to 230% faster than an M5 Ultra and up to 40% faster than two DGX Sparks. Used 32GB V100 cards sell for around $744 on eBay.

• Four Tesla V100s on NVLink deliver 300GB/s and 32GB HBM2 each
• Long-prompt processing is 230% faster than the M5 Ultra
• Token generation hits 100.6 tok/s versus 44.5 tok/s on M5 Ultra
• Used 32GB Tesla V100 cards cost about $744 on eBay