Skyrmionics for Neuromorphic Technologies

Lead Research Organisation: The University of Manchester
Department Name: Computer Science

Abstract

In today's world of pervasive Information Technology (IT), there is a pressing need to develop novel computing paradigms to move beyond current architectures with the goal of achieving intelligent computing with superior efficiency. Modern, conventional, computers operate in a very different manner to that of the human brain. In stark contrast, the main building blocks of the human brain are neurons (the computing elements) and synapses (the adaptive memory elements) and neurons are massively interconnected with synapses. Since learning is intricately connected to synaptic behavior, this project seeks to build next-generation artificial synapses In particular, we will explore the potential of non-volatile artificial synapses, based on nanoscale magnets, for energy-efficient brain-inspired operations, also known as neuromorphic computing. In a market with products requiring an abundance of sensors at the edge (e.g. mobiles or wearables like smart watches), there is a recognised need for ultra-low power and always-on sensory data processing. Neuromorphic hardware is one of the most promising routes for Artificial Intelligence (AI) applications. We propose to demonstrate that nanoscale skyrmionics synapses (that use nanoscale whirling vortex-like magnetic states called skyrmions as information carriers) are ideal for energy-efficient smart edge-computing devices.
 
Description Horizon Europe Framework Programme
Amount € 3,991,356 (EUR)
Funding ID 101135729 
Organisation European Commission 
Sector Public
Country Belgium
Start 01/2024 
End 01/2028
 
Title Power profiling feature for SNNTorch 
Description We developed a new Power Profiling feature for SNNTorch, the popular open-source Python package for spiking neural networks (SNNs), extending PyTorch capabilities. We chose SNNTorch due to its popularity & uptake by the community (> 100,000 downloads & integration with Graphcore's IPUs, enabling a considerable reduction in energy consumption for SNNs users). Our feature estimates energy consumption for neural network inference & adapts to different hardware types, supporting both spiking and non-spiking networks. The feature is tailored to both network topology & hardware specifics. M. Lewandowski (team member from summer internship 2023 to February 2024) contributed it to the official repository as an experimental feature and it is expected that it will be merged into the master branch. 
Type Of Material Improvements to research infrastructure 
Year Produced 2023 
Provided To Others? Yes  
Impact The power profiling code allows researchers to benchmark the overall power consumption of spiking neural networks against competing hardware and/or network topologies. It therefore will inform hardware- and network-level design decisions, both in our work, and in the research of the wider community. 
URL https://github.com/jeshraghian/snntorch/tree/Power-Profiling
 
Description CNRS - Thales 
Organisation Unité Mixte de Physique CNRS/Thales
Country France 
Sector Public 
PI Contribution We started a new collaboration with CNRS and Thales, that was inspired by our work on neuromorphic skyrmionic synapses and skyrmionic interconnects in the "Skyrmionics for Neuromorphic Technologies" project. We will work with CNRS/Thales on magnetic multilayers that support skyrmion transport and skyrmion detection, all crucial parts of a new skyrmionic synapse design.
Collaborator Contribution CNRS will provide magnetic multilayers that are optimised for Skyrmion transport to complement our Manchester grown magnetic multilayers. Thales will provide know-how on skyrmion detection in order for our skyrmionic synapses to be implemented in a fully electrical way, not only relying on imaging.
Impact As this a new collaboration we expect outputs and outcomes over the next 1 year.
Start Year 2023