Dual MI50 64GB HBM2 local inference rig built around a PLX switch
Savantskie1 · reddit · 2026-10-06
A Redditor shared a complete local LLM inference build that uses a PLX expansion card to sidestep PCIe bifurcation headaches and fit two GPUs.
- GPUs: 2× AMD Instinct MI50 32GB (64GB HBM2 total) with aftermarket blower coolers
- Interconnect: PLX8749 expansion card (4× SFF-8654, PCIe x16) with baseplates and ribbon cables
- Host: Ryzen 5 5600G for display/system duties, MSI B550 board, 48GB DDR4, 4TB NVMe
- Power: dual PSU setup — 850W for the system, 1250W dedicated to the GPUs
- Software: Ubuntu 22.04, llama.cpp on ROCm 6.4.3, OpenWebUI frontend
A useful reference parts list for anyone building local inference on cheap datacenter cards; the author invites questions.
More from Infra
- AI tools quietly ate 44 GB of his disk, so he built free cleaner Sparewise — PossibilityKind3028 · 2026-10-06
- Nebius hikes RAM 41% and GPUs up to 21% as the chip shortage hits cloud price lists — tengyanAI · 2026-10-06
- GitHub Actions goes down again, breaking CI pipelines — generativist · 2026-10-06
- 462GB DeepSeek model runs on two desk-side DGX Sparks with experts squeezed to 2.77 bits — Teknium · 2026-10-06
- Wilderness Society calls for immediate moratorium on AI data centers on US public lands — Polymarket · 2026-10-06
- KCoral: Shared GPU Benchmark Environment Speeds Agentic Kernel Evaluation 2.58x on B200 — BeidiChen · 2026-10-06