Scottcjn/ram-coffers

LLM infrastructure cost reduction via NUMA-aware weight banking: 147 t/s (8.8x stock llama.cpp) on refurbished enterprise POWER8. Self-hosted inference, no cloud APIs. Part of the Proof of Physical AI stack.

2 best points · 1 comments · 1 Hacker News post · 173 GitHub stars · View on GitHub

Hacker News discussions

• RAM Coffers– I built conditional memory for LLMs 27 days before DeepSeek'sEngram (2 points, 1 comments, by AutoJanitor, 2026-01-19)