Local LLM User Weighs 4x MI50 vs Second 7900XTX for Bigger Models
Thac0-is-life · reddit · 2026-10-02
A local-LLM hobbyist asks for hardware upgrade advice. Current rig: Ryzen 5600X, 7900XTX, 64GB DDR4, on an X470 GAMING PRO motherboard that isn't great for multi-GPU. He mainly runs Qwen 3.8 27B at 3–4-bit quantization with 128k–220k context and 8-bit KV cache, getting 50 tok/s and 750 tok/s context ingestion.
His goals: run bigger models, go faster, or run multiple Qwen instances so several agents can work in parallel, plus try Qwen 3.8 Flash. He's torn between buying 2–4 MI50 16GB cards (about a third the price of 32GB cards) or a second 7900XTX (more expensive than 4 MI50s), noting he'd likely still need a new motherboard for dual-GPU use.
He adds that noise isn't an issue in his basement and solar covers summer power, and asks whether older GPUs like the MI50 make sense and what motherboard to pair them with.
More from Embodied
- Figure's data collection project Index crosses 1 million signed-up users — adcock_brett · 2026-10-02
- DATAFARM: aligning TAMP demos to VLA pretraining distribution unlocks planning plus dexterity — tomssilver · 2026-10-02
- Wearable AI lets you capture thoughts by voice anywhere you go — broodsugar · 2026-10-02
- UniWAM unifies physical reasoning, world generation, and action prediction for robots — zhenjun_zhao · 2026-10-02
- CloudAnalyzer: browser-based LiDAR tools for SLAM map fixes and Lanelet2 editing in Rust+WASM — rsasaki0109 · 2026-10-02
- Ego-Exo4D-HM: SMPL-H reconstructions for 523 hours of egocentric-exocentric video, open-sourced — geopavlakos · 2026-10-02