Reproduce “Tri-Stack-PIM: A Heterogeneous Multi-Die 3D-Integrated Processing-in-Memory Architecture for Efficient LLM Inference” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shuao Jia, Zichao Ling, Chen Bai, Kang Zhao, Jianwang Zhai
Tri-Stack-PIM divides attention work across three dies: DRAM stores weights and the key-value cache, a middle die handles interfaces, control, and softmax, and a compute die performs arithmetic. Its hardware scheduler issues ready operations to reduce delays from blocked work. In modeled evaluations, it achieves geometric-mean per-token attention speedups of 4.95× over GPU-only, 2.57× over AttAcc, and 1.68× over H²-LLM; these are not end-to-end generation results.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.