Budget 70B rigs revive old Xeons

This is the most practical side of local AI right now — not model hype, but brutal cost engineering around VRAM, lanes, power, and memory generation.

• –The core bet is sound: three 24GB 3090s are still one of the cheapest ways to get enough combined VRAM for serious 70B experimentation
• –X79 and X99 both come from Intel’s 40-lane HEDT/server era, so the question is less “will it boot” and more “how much pain do you accept in bandwidth, thermals, and platform age”
• –X99 is the cleaner choice because newer Xeons and DDR4 reduce platform drag, even if the GPUs still do most of the real inference work
• –Recent llama.cpp multi-GPU discussions suggest performance bottlenecks often come from orchestration and CPU/RAM offload overhead, not just raw PCIe bandwidth
• –The post is notable because it captures the real local-LLM market: old enterprise gear stays relevant as long as consumer GPUs keep winning on VRAM-per-dollar