Multibit neural inference in a N-ary crossbar architecture

arXiv cs.AI · 6 hours ago

Researchers developed a simulation framework for in-memory computing using N-ary crossbar architectures with magnetic tunnel junctions to perform neural network inference directly in memory arrays. A 4x4 crossbar array with 4-state MTJs achieved 93.56% accuracy on MNIST classification compared to a 97.56% software baseline. Weight quantization emerged as the primary error source, and the study identified an optimal number of states per cell that minimizes total matrix-vector multiplication error while balancing quantization and resistance resolution tradeoffs.