.png){kind=icon}
.png){kind=icon}
Return
Solving linear programming in one step with RRAM-based analog matrix computing
.png){kind=icon}
OA
AI
DOI：10.3389/fnano.2026.1882288.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Linear programming (LP) is among the most fundamental optimization techniques. However; solving LP problems on conventional digital hardware is increasingly constrained by the polynomial computational complexity of matrix operations. In this work; we present an analog matrix computing (AMC) circuit built on resistive random-access memory (RRAM) crossbar arrays and the projection neural network (PNN) model that solves LP problems in one step. The proposed circuit directly maps the PNN dynamical system onto a closed-loop feedback system; where an RRAM-based projection matrix computation unit executes the projection matrix computation in one step; while analog neuron modules perform integration; nonlinear activation; and subtraction to establish a global negative feedback loop. The proposed circuit was tested on assignment problems and instances from the LP Netlib test set. Circuit-level simulations show that the solver reaches stable solutions within approximately 15 μs for problems with up to 100 variables; achieving nearly two orders of magnitude speedup over the digital PNN baseline. We further introduce an analog-digital hybrid approach where the analog circuit rapidly supplies a near-optimal seed solution to initialize a digital iterative solver; reducing subsequent iteration counts by up to 54.4%. These results demonstrate that RRAM-based AMC offers a promising route toward real-time; energy-efficient LP solving at scale.
Keywords:
linear programming
in-memory computing
RRAM
matrix computing
projection neural network
Journal
F
IF:
3.8
Papers:
508
Citations:
1.3K