notesum.ai

Published at October 21

Enabling Energy-Efficient Deployment of Large Language Models on Memristor Crossbar: A Synergy of Large and Small

cs.AR

cs.DC

q-bio.QM

Released Date: October 21, 2024

Authors: Zhehui Wang, Tao Luo, Cheng Liu, Weichen Liu, Rick Siow Mong Goh, Weng-Fai Wong

Arxiv: https://arxiv.org/abs/2410.15977v1

Refer to caption

No.	$F(\mathbb{X}\cdot\mathbb{Y})=F(\mathbb{Z})$	Type	$\mathbb{X}\cdot\textrm{col}_{t}(\mathbb{Y})$	$\textrm{col}_{t}(\mathbb{Z})$	$F$	$F(\textrm{col}_{t}(\mathbb{Z}))$	$t\in$
1	$E(u_{t\ast})=\$ row ${}_{t}(\textbf{U})\cdot\textbf{1}$	NW	$\begin{bmatrix}u_{t1}&u_{t2}&\cdots&u_{tm}\\ \end{bmatrix}$ $\begin{bmatrix}1\\ \vdots\\ 1\\ \end{bmatrix}$	${\sum}_{i=1}^{m}u_{ti}=u_{t1}+u_{t2}+\cdots+u_{tm}$	/m	${\sum}_{i=1}^{m}u_{ti}/m$	$[1,n]$
2	$E(u^{2}_{t\ast})=\$ row ${}_{t}(\textbf{U})\cdot$ row ${}_{t}(\textbf{U})^{T}$	NW	$\begin{bmatrix}u_{t1}&u_{t2}&\cdots&u_{tm}\\ \end{bmatrix}$ $\begin{bmatrix}u_{t1}\\ \vdots\\ u_{tm}\\ \end{bmatrix}$	${\sum}_{i=1}^{m}u^{2}_{ti}=u^{2}_{t1}+u^{2}_{t2}+\cdots+u^{2}_{tm}$	/m	${\sum}_{i=1}^{m}u^{2}_{ti}/m$	$[1,n]$