Local LLM on 3090 + 64GB RAM hits 70-80 t/s with QFN quants; eyeing a $3-3.5k Strix Halo box
makkalot · reddit · 2026-11-12
A Reddit user reports running Strata/QFN (iq3xss) quantized models on an RTX 3090 with 64GB RAM at 70-80 tokens/s as their daily driver — though the setup saturates GPU and RAM, leaving the machine unusable for anything else.
They're now considering a dedicated inference box, specifically an AMD Strix Halo mini PC at around €3-3.5k in the EU, and are asking whether other Strix Halo + QFN users see similar speeds or something much slower.
More from Infra
- DGX Spark resale beat NVDA stock: same $15,996 returned +73.8% vs -3.4% in 13 months — OwariDa · 2026-11-12
- Musk reportedly betting on Terafab: one plant producing a terawatt of silicon chips per year — XFreeze · 2026-11-12
- Physics-informed learning by IITM professor slashes AI simulation compute costs — ravi_iitm · 2026-11-12
- Precision on Demand: Escalating Weight Bits at Uncertain Decoding Steps — Mohammed Nayeem Khan · 2026-11-12
- MiniMax-H3-Image FP8 fits a single 32GB AMD card: 4K images in 33s — evp-cloud · 2026-11-12
- Energy Tech Draws Nearly $200B in 2025 as Data-Center Loads Test Grid Interconnection Queues — ingliguori · 2026-11-12