Building a Local LLM Rig from Used Parts
molbal · reddit · 2026-07-17
This is a local AI machine "pieced together from old parts": a Supermicro X9DR3-F motherboard, dual Xeon E5-2660 v2 CPUs, 8×32GB DDR3 RAM, an RTX 3090, and a 1000W PSU. Based on local secondhand prices in the Netherlands, the author estimates the total cost at a few hundred euros.
The author details the various failures during the build process: eGPU dock incompatibility, a server destroyed in transit, a damaged AliExpress motherboard, and a PSU failure, before finally getting a bootable configuration. They admit it cost them a lot of "sanity and patience."
Benchmarks & Configuration
- Used the Q4KXL quantized version of DeepSeek V4 Flash.
- Set ctx-size to 32768, but actually only achieved 8k context.
- Speed results: pp 37.47 t/s, tg 6.62 t/s.
- Also shared the llama-server configuration, emphasizing auto offload, CPU MoE, and NUMA distribution for local inference tuning.
More from Infra
- Why a 1GW Chinese AI data center may be plausible after all — teortaxesTex · 2026-07-22
- China’s AI arms race is increasingly defined by chips, data centers, and open models — BenBajarin · 2026-07-22
- Agent search bottlenecks are now about variance, not raw latency — rohanpaul_ai · 2026-07-22
- Gavin Baker argues Nvidia may be one of open source AI’s biggest supporters — GavinSBaker · 2026-07-22
- AI Power Demand Exposes US Energy Gap, Urging Shift from Scarcity to Abundance — bradneuberg · 2026-07-22
- Gavin Baker says Nvidia’s $630B figure would be system revenue, not all Nvidia’s — GavinSBaker · 2026-07-22