BACK_TO_FEEDAICRIER_2
AIfred Intelligence turns AOOSTAR GEM10 into 120GB rig
OPEN_SOURCE ↗
REDDIT · REDDIT// 14d agoINFRASTRUCTURE

AIfred Intelligence turns AOOSTAR GEM10 into 120GB rig

The AIfred Intelligence repo now reads like the software half of a serious self-hosted AI stack: the author documents how an AOOSTAR GEM10 Pro Max, three Tesla P40s, and a Quadro RTX 8000 became a 120 GB VRAM box for running 235B-class models locally. The post traces the OCuLink, USB4, and M.2 expansion path, then shows how ReBarUEFI and custom cooling made the whole setup usable.

// ANALYSIS

This is peak hobbyist infra: it looks excessive until you compare it with buying a ready-made 128 GB machine, and then the mix of used cards, consumer mini-PCs, and firmware hacks starts to make weird sense.

  • AOOSTAR's GEM10 is unusually well suited to this kind of abuse because it gives you OCuLink, USB4, and extra M.2 slots without forcing a tower build.
  • ReBarUEFI is the real unlock, and the repo/docs line up with the post's experience: large BAR support and 4G decoding are what make awkward GPUs behave on consumer UEFI.
  • The software stack matters as much as the hardware. AIfred's docs already steer Tesla P40-class cards toward llama.cpp/Ollama, so this rig is matched to the backend rather than fighting it.
  • The payoff is density, not elegance: 120 GB of VRAM and low idle power give you desk-sized access to models that normally assume a rack, while AIfred's multi-agent, voice, and RAG features make that headroom useful.
  • The tradeoff is obvious: sawed fan grilles, per-card cooling, no vLLM, and a lot more maintenance than a unified-memory box.
// TAGS
aifred-intelligencellminferencegpuself-hostedagentautomationopen-source

DISCOVERED

14d ago

2026-03-28

PUBLISHED

14d ago

2026-03-28

RELEVANCE

8/ 10

AUTHOR

Peuqui