Running 4x RTX 3090 on a consumer board: a Qwen3.8 local inference build
Civil_Fee_7862 · reddit · 2026-11-04
A Reddit user built a 4x RTX 3090 local inference rig on a consumer Asus ProArt B850 board with limited PCIe lanes: two 3090s run Qwen3.8-27B in tensor parallel, a third at 4x handles image tasks, and a fourth is being connected via a Gen4 1x-to-16x adapter.
- Specs: 32GB DDR5 6000MHz, severely constrained PCIe lanes
- Original NVLink plan abandoned in favor of Qwen3.8 Flash Next with experts resident in VRAM, since VRAM exceeds system RAM
- Open question: whether Strata's expert/pipeline parallelism is viable at Gen4 1x bandwidth
More from Infra
- SpecPilot: Berkeley and NVIDIA's Agent Loop Auto-Optimizes Speculative Decoding per Deployment — UCBerkeley · 2026-11-04
- AMD exec joins Fireworks AI Forge event to highlight multigenerational partnership — AnushElangovan · 2026-11-04
- Microsoft details Outlook search overhaul: 7 relevance-ranked results and offline resilience — benjamingauthey · 2026-11-04
- AMD exec gifts DHH a 192GB AMD workstation, embraces Omarchy Linux — AnushElangovan · 2026-11-04
- How LLM inference actually works: from 140GB weights to GPUs and model serving, explained as a story — _jaydeepkarale · 2026-11-04
- LlamaIndex speeds up doc-parsing model OCRia-1 10x to 2800 tok/s on one B200 — llama_index · 2026-11-04