VDAOvideo

Local LLM on Two Used GPUs: A Practical AI Machine

Lexi Volt · 292 views · 3 weeks ago
Safety check pending
Description

Used parts, turned into a practical AI machine. Two used AMD graphics cards, an RX 6900 XT and an RX 6800, run a 35-billion-parameter local LLM with a 96k context and write real code at about 90-112 tokens a second. Full guide, every command copy-paste ready, on Vibe Garage: https://vibegarage.neozonesproductions.com/learn.html?p=ZEzo6OJ6bm-dv68iTmQ-Kw Chapters: 0:00 Used parts, practical AI machine 0:25 Why this setup 0:46 Step 1: the rig 1:03 Step 2: why two cards 2:16 Step 3: the model 2:50 Step 4: the command 3:30 Is it practical? The proof 4:13 Bigger context just slows it down 4:33 The honest limits 5:01 Recap and the full guide Launch command (llama.cpp build b11871, Vulkan): llama-server -m Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf -ngl 999 -sm layer -ts 1,1 -c 98304 -fa on --spec-type dr