Mixed signal computing-in-memory solver for integer constrained least square problems

Abstract

A mixed signal computing-in-memory (CIM) system solves integer constrained least square (ICLS) optimization problems through parallel analog computation within memory arrays. The system transforms ICLS problems into energy minimization problems solved using Hopfield neural network dynamics. A first CIM component performs iterative optimization using quantized weights stored in memory arrays, with analog current accumulation providing parallel computation and stochastic feedback enabling escape from local minima. A second CIM component evaluates solution quality through distance metric computation. The system converts integer optimization variables to binary representations suitable for neural network processing. Binary neuron states are generated through current comparison and modified with programmable noise injection for global optimization. The system supports constant time operation for predictable latency and dynamic time operation for faster convergence. The CIM approach eliminates data movement bottlenecks while achieving significant speedup and energy efficiency compared to conventional digital processing methods for solving nondeterministic polynomial time-complete ICLS problems.