Scottcjn/ram-coffers
View on GitHub ↗ ★ 174 Python
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.
3 mentions on Hacker News by 1 person: 2 stories, 1 comment · first 2026-01-19 · latest 2026-01-21
20082026
-
▲ 1
Subject-based weight routing for LLMs (27 days before DeepSeek Engram)story by AutoJanitor · 2026-01-21 · 0 comments
-
…via O(1) lookup. Same idea. 27 days apart. No connection. DOI: 10.6084/m9.figshare.31093429 GitHub: https://github.com/Scottcjn/ram-coffers DeepSeek paper: https://arxiv.org/abs/2601.07372 Running on "obsolete" POWER8 hardware, I hit 147…
comment by AutoJanitor · 2026-01-19 · in RAM Coffers– I built conditional memory for LLMs 27 days before DeepSeek'sEngram -
▲ 2
RAM Coffers– I built conditional memory for LLMs 27 days before DeepSeek'sEngramstory by AutoJanitor · 2026-01-19 · 1 comment