pull down to refresh

I do not run any LLMs locally. But a guy can dream, can't he? Curious if this matches up with what anyone else who is running things locally has found.

ContentsContents

SectionTL;DR
How much are you willing to spend?$2k gets you Qwen and good STT (pretty far!); $40k gets you almost-Opus
Base systemLast-gen EPYC + eBay DDR4 for $5.6k
GPUs4× RTX PRO 6000, 384GB VRAM, where the money went
c-payne switch sub-BOMIndie PCIe switching so GPUs talk peer-to-peer
GPU enclosureA day of carpentry
Making the switch behaveBIOS bifurcation, link speed, ASPM
Kernel / GRUB paramsiommu=off or NCCL hangs
ACS disableKeep P2P traffic inside the switch fabric
GPU power limitingRunning $46k of silicon on a 110V circuit
ResultGen4 line rate: 27.5/50.4 GB/s, sub-µs latency
runners/Ready-to-run serving configs: https://github.com/jamesob/local-llm/blob/master/runners/GLM-5.2-594B: vLLM docker-compose, DCP4+MTP5, ~80 t/s @ 240k ctx
tools/https://github.com/jamesob/local-llm/blob/master/tools/measure-gpu-speed.sh: P2P bandwidth/latency benchmark
Resourcesrtx6kpro repo, c-payne

I like the enclosure he built: