Power loss (“dissipation”) has become the major bottleneck for computing and, more generally, information and communication technologies (ICT), with strong knock-on effects in terms of e.g. rising energy costs for end consumers, challenging on-chip waste heat management for device architects, limited computing power for scientific and medical modellers, and increasing carbon footprint for wider society.
In MANNGA, we aim to alleviate this energy efficiency bottleneck by taking advantage of the wave motion of spins (“spin waves” or “magnons”) in a magnetic material (“magnonic medium”). Such spin waves can be used to transport data without incurring power dissipation in electrical leads. So, the associated research field, known as magnonics, is now commonly accepted as a green path forward.
MANNGA goes further yet by developing magnonic versions of artificial neural networks, which are the cornerstone of another energy-efficient technology paradigm – neuromorphic computing. The basic idea of the latter is to trade in some accuracy in order to save energy, akin what biological brains do.
The basic building block of MANNGA’s magnonic neural networks is a so called “magnonic resonator” – an island of a magnetic material whose spins can interact with and thereby control spin waves propagating in a magnonic medium nearby. In MANNGA, such magnonic resonators are used as artificial neurons – magnonic neurons, arrays of which then form our magnonic networks. Spin waves in the medium transport information among the magnonic neurons, as synapses do in biological brains.
The experimental implementation of such magnonic neurons and neural networks is MANNGA’s primary research objective. Once achieved, we will proceed to development of methods of using the networks to implement computing devices: (i) magnonic field programmable gate arrays (mFPGAs), (ii) magnonic reservoir computers (mRCs), and (iii) magnonic recurrent neural networks (mRNNs).
The mFPGAs operate with spin waves of constant amplitude and phase. They have multiple inputs and outputs. The amplitude and / or phase of spin waves at the inputs encode data. The data undergoes a complex transformation inside the mFPGA which results in certain values of the amplitude and / or phase of spin waves that reach the different outputs. The mFPGA relates different sets of output values to different sets of input values, i.e. performs a logical operation. The magnetic states of the different magnonic resonators within the neural network acting as the mFPGA can then allow one to programme which specific logic function performed and then to switch between different functions. The process of identifying the resonators’ states enabling a particular logic function is called “training” of the hidden layers of the neural network.
The mRCs operate on dynamic data (encoded into the spin wave amplitude and / or phase, which are therefore no longer constant in time) and may consist of just a few magnonic resonators. The resonators are not “trained”, but the result of the “computation” is extracted instead by manipulating the modulated spin wave that is read out from the mRC’s output (also called “output training”).
The mRNNs are most complex and powerful of MANNGA’s devices. Firstly, they combine the relatively large arrays of resonators as in the mFPGAs and the use dynamical data encoding and output training as in the mRCs. In addition, the networks feature dynamic feedback to enhance their computing power.
Our demonstration and comprehensive characterisation of the devices and key performance indicators (KPIs) will constitute the main scientific and technological impacts of the project. Our next steps will be determined by the results of this characterisation and KPIs achieved. In particular, the latter will be compared with those achieved by state-of-the-art approaches in a comparable operation class.
