Horizon Europe Consortium: MultiSpin.AI

Spintronic Neuromorphic Hardware for Energy-Efficient Artificial Intelligence

Since February 2024, Prof. Klein has been coordinating MultiSpin.AI, an EIC Pathfinder Open project funded under the Horizon Europe programme. The project brings together leading academic and industrial partners to develop neuromorphic edge-computing hardware based on multilevel spintronic devices, with the goal of enabling faster and significantly more energy-efficient AI inference.

MultiSpin.AI addresses a central challenge facing artificial intelligence today: while AI applications continue to expand rapidly—spanning language models, medical diagnostics, industrial automation, and autonomous systems—the underlying computing hardware is approaching fundamental physical and architectural limits. Data volumes reached approximately 97 zettabytes in 2022 and are doubling every two to three years, placing unsustainable demands on existing digital infrastructures. At the same time, information and communication technologies are projected to consume up to 21% of global electricity by 2030, underscoring the urgent need for new, energy-efficient computing paradigms.

The Challenge: Hardware Limits of Contemporary AI

Current AI hardware faces several structural bottlenecks:

• Von Neumann bottleneck: The separation of memory and processing units severely limits data throughput and energy efficiency, particularly for AI workloads.

• End of Moore’s and Dennard’s scaling: As transistor dimensions approach the 2 nm regime, further performance gains through conventional scaling are becoming increasingly marginal.

• Escalating energy consumption: The power requirements of large-scale AI systems threaten sustainability and restrict deployment at the edge.

These limitations motivate a transition away from purely CMOS-based architectures toward fundamentally new computing concepts.

The MultiSpin.AI Approach

MultiSpin.AI proposes a new class of spintronics-based AI edge co-processors built on n-ary (multilevel) magnetic tunnel junctions. Unlike conventional binary devices, these structures can reliably access multiple magnetic states, enabling higher information density, reduced data movement, and intrinsic compatibility with neuromorphic computing concepts.

By processing data locally at the edge—closer to sensors and data sources—MultiSpin.AI aims to reduce latency, bandwidth usage, and energy consumption, while maintaining high inference accuracy. This approach is particularly relevant for distributed and autonomous systems, including next-generation unmanned airborne platforms and IoT applications.

Scientific and Technical Objectives

The primary objective of MultiSpin.AI is to design, fabricate, and experimentally validate multistate spintronic devices for AI inference. Key goals include:

• Design and realization of magnetic structures supporting 8 to 16 stable states, enabling n-ary computation beyond binary logic.

• Fabrication and characterization of multilevel magnetic tunnel junctions using advanced nanofabrication techniques.

• Demonstration of AI inference using a prototype co-processor array, initially based on devices with four stable states, serving as a proof of concept.

• Co-design of hardware, electronics, and AI algorithms to fully exploit the capabilities of multilevel spintronic devices.

Consortium and Expertise

The MultiSpin.AI consortium combines complementary expertise across materials science, spintronics, electronics, artificial intelligence, and innovation:

• Materials science and spintronics: Researchers from Bar-Ilan University (BIU), INESC-MN, UCLouvain, and SpinEdge focus on device modeling, simulation, and co-design. BIU and INESC-MN lead nanofabrication and patterning.

• Nanomagnetism and characterization: Experimental and computational characterization is carried out by teams at BIU, UCLouvain, and INESC-MN.

• Electronics and system integration: Engineers and physicists from INESC-MN, SpinEdge, UCLouvain, and I-FEVS develop the PCB-level co-processor prototype, including driver electronics and control software.

• AI, ethics, and human-centred design: AI algorithm developers, IoT experts, ethicists, and human-centred designers collaborate to ensure responsible, unbiased, and application-relevant AI demonstrations.

Project Information

• Programme: Horizon Europe – EIC Pathfinder Open

• Grant Agreement: 101130046

• Duration: 2024–2027

• Coordinator: Prof. Klein

For further information and project updates, visit https://multispinai.eu and follow the consortium on LinkedIn:
https://www.linkedin.com/company/multispin-ai