Scalable In-Memory Computing with N-ary Spintronic Crossbar Arrays

The ever-growing energy and latency demands of AI workloads call for a fundamental shift away from conventional von Neumann architectures. In-memory computing (IMC), which co-locates computation and storage within analog crossbar arrays, is a compelling approach to address the von Neumann bottleneck and the memory wall. Among emerging non-volatile memory technologies, multi-state magnetic tunnel junctions (M2TJs), whose resistance is controlled by the magnetization configuration of crossed ferromagnetic ellipses, stand out for their non-volatility, CMOS compatibility, and intrinsic potential for equidistantly separated resistance levels, making them ideal candidates for n-ary neuromorphic IMC. We present a simulation framework for performing multibit neural network inference on n-ary crossbar architectures, developed for spintronic M2TJs but generalizable to any multistate memristive device. The framework retrieves the matrix-vector multiplication (MVM) result from the measured analog output with minimal assumptions, enabling the crossbar to operate as a standalone AI co-processor. Using simulated 4×4 arrays of 4-state M2TJs, we demonstrate successful inference on the XOR task and on MNIST handwritten digit classification, achieving 94.48% accuracy against a 97.56% software baseline. This gap is further reduced to 2.79% via PCA dimensionality reduction. We provide a systematic error analysis identifying weight quantization as the dominant error source. Systematic nonidealities (shared by all cells) and cell-specific random noise are also studied. Critically, cell-specific noise is less detrimental than systematic nonidealities of equal amplitude, owing to beneficial averaging across the array. The RMSE scales as √n with the number of columns while the SNR grows logarithmically, favoring wide crossbar geometries. Finally, we demonstrate the existence of an optimal number of resistance states Nopt that minimizes total MVM error by trading off quantization error against state resolution, with Nopt decreasing in noisier device environments.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-15
DOI
https://doi.org/10.5281/zenodo.24249964
Primary Topic
Advanced Memory and Neural Computing
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Scalable In-Memory Computing with N-ary Spintronic Crossbar Arrays

Flavio Abreu Araujo
Zenodo (CERN European Organization for Nuclear Research)
Advanced Memory and Neural Computing
article

Scalable In-Memory Computing with N-ary Spintronic Crossbar Arrays

Flavio Abreu Araujo
article en

Abstract

The ever-growing energy and latency demands of AI workloads call for a fundamental shift away from conventional von Neumann architectures. In-memory computing (IMC), which co-locates computation and storage within analog crossbar arrays, is a compelling approach to address the von Neumann bottleneck and the memory wall. Among emerging non-volatile memory technologies, multi-state magnetic tunnel junctions (M2TJs), whose resistance is controlled by the magnetization configuration of crossed ferromagnetic ellipses, stand out for their non-volatility, CMOS compatibility, and intrinsic potential for equidistantly separated resistance levels, making them ideal candidates for n-ary neuromorphic IMC. We present a simulation framework for performing multibit neural network inference on n-ary crossbar architectures, developed for spintronic M2TJs but generalizable to any multistate memristive device. The framework retrieves the matrix-vector multiplication (MVM) result from the measured analog output with minimal assumptions, enabling the crossbar to operate as a standalone AI co-processor. Using simulated 4×4 arrays of 4-state M2TJs, we demonstrate successful inference on the XOR task and on MNIST handwritten digit classification, achieving 94.48% accuracy against a 97.56% software baseline. This gap is further reduced to 2.79% via PCA dimensionality reduction. We provide a systematic error analysis identifying weight quantization as the dominant error source. Systematic nonidealities (shared by all cells) and cell-specific random noise are also studied. Critically, cell-specific noise is less detrimental than systematic nonidealities of equal amplitude, owing to beneficial averaging across the array. The RMSE scales as √n with the number of columns while the SNR grows logarithmically, favoring wide crossbar geometries. Finally, we demonstrate the existence of an optimal number of resistance states Nopt that minimizes total MVM error by trading off quantization error against state resolution, with Nopt decreasing in noisier device environments.

Zenodo (CERN European Organization for Nuclear Research)
UCLouvain (BE)
Advanced Memory and Neural Computing
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon