Can 9 year old GPU beats today's latest GPUs?

A 9-year-old NVIDIA Tesla V100 rig beat two brand-new DGX Sparks at local LLM token generation: 155 tok/s against about 60 on Qwen 3.8 Flash Next (NVFP4, vLLM). Here's why old HBM2 bandwidth wins, and why the speed really comes from community software. The build: four used V100 SXM2 32GB cards on an NVLink carrier board, 128GB of VRAM for about $3,500. The other side is a DGX Spark Founders Edition plus a Gigabyte AI TOP Atom, two GB10 Grace Blackwell boxes clustered over ConnectX-7 (the same GB10 Superchip behind the new RTX Spark laptops). We break down prompt processing vs token generation