Periodic Reporting for period 1 - SkyANN (Skyrmionic Artificial Neural Networks)
Période du rapport: 2024-01-01 au 2025-06-30
As energy demands soar, traditional computing struggles with inefficiency. In addition, existing neural networks hardware faces limitations in energy consumption and reliability. In this context, the EU-funded SkyANN project (Skyrmionic Artificial Neural Networks) will develop interconnected magneto-electric devices that target double bandwidth and an energy consumption that is four orders of magnitude lower than state-of-the-art, thereby challenging conventional computing paradigms. SkyANN will achieve these objectives by mirroring brain functions at the synapse level. This magneto-electric neural network hardware can offer superior energy-efficiency and reliability. With ambitious plans to pioneer magneto-electric neural networks, SkyANN aims to reshape European microelectronics. Its multi-actor consortium will ensure rapid innovation transfer, bolstering Europe's semiconductor sector in alignment with the vision of the EU Green Deal.
The overarching objective of the SkyANN project is to realise a hardware-based deep Skyrmionic Artificial Neural Network (SkyANN). This goal is structured into four specific objectives. Objectives 1–3 develop the core building blocks, ultra-low-power synapses, skyrmionic neurons, and multiplexed dendritic connections, that together enable this demonstrator in Objective 4.
Specific Objective (SO) 1: Ultra-low power skyrmionic synapses
For SO1 we have so far established synaptic weighting and integration with magnetic skyrmions as information carriers, including a published demonstration of field-controlled weights and near-linear summation. Future work is shifting toward ultra-low-power, voltage-controlled synapses with non-volatile weight setting and benchmarking.
Objective 2 - Skyrmionic multiplexer/demultiplexer dendritic connections
For SO2, we have established the experimental platform for co-hosting multiple skyrmionic species (magnetic skyrmions and magnetic cocoons) and shown current-driven motion of cocoons and selective demultiplexing of the species in simulation. The next step is to demonstrate current-driven motion of cocoons and demultiplexing experimentally.
Objective 3 - Fully integrated skyrmionic neuron with high fan-out
For SO3, we have demonstrated synaptic integration at the device level, corresponding to the neuromorphic weighted sum operation, using AHE readout of magnetic skyrmions. This includes both linear addition of signals from two tracks under magnetic-field sweeps and summation of skyrmion inputs generated by current pulses, with results published. The next milestones are to realise non-linear activation and achieve high fan-out through CMOS integration.
Objective 4 - A Deep Skyrmionic Artificial Neural Network
This objective is the project’s ultimate goal: a hardware demonstrator of a feed-forward skyrmionic neural network. Building on Objectives 1–3, this objective integrates all SkyANN innovations as a proof of concept for future scaled-up deep networks. Work towards this demonstrator began in WP4 (M10) and continues through to project completion.
Specific Objective (SO) 1: Ultra-low power skyrmionic synapses
For SO1 we have so far established synaptic weighting and integration with magnetic skyrmions as information carriers, including a published demonstration of field-controlled weights and near-linear summation. Future work is shifting toward ultra-low-power, voltage-controlled synapses with non-volatile weight setting and benchmarking.
Objective 2 - Skyrmionic multiplexer/demultiplexer dendritic connections
For SO2, we have established the experimental platform for co-hosting multiple skyrmionic species (magnetic skyrmions and magnetic cocoons) and shown current-driven motion of cocoons and selective demultiplexing of the species in simulation. The next step is to demonstrate current-driven motion of cocoons and demultiplexing experimentally.
Objective 3 - Fully integrated skyrmionic neuron with high fan-out
For SO3, we have demonstrated synaptic integration at the device level, corresponding to the neuromorphic weighted sum operation, using AHE readout of magnetic skyrmions. This includes both linear addition of signals from two tracks under magnetic-field sweeps and summation of skyrmion inputs generated by current pulses, with results published. The next milestones are to realise non-linear activation and achieve high fan-out through CMOS integration.
Objective 4 - A Deep Skyrmionic Artificial Neural Network
This objective is the project’s ultimate goal: a hardware demonstrator of a feed-forward skyrmionic neural network. Building on Objectives 1–3, this objective integrates all SkyANN innovations as a proof of concept for future scaled-up deep networks. Work towards this demonstrator began in WP4 (M10) and continues through to project completion.
The European neuromorphic computing ecosystem presently explores a gamut of devices and related emerging technologies (such as FeFET, MTJs, memristors, resistive memories) as means to place Europe at the forefront of AI evolution amidst a highly competitive landscape. Just recently the advantages and possible opportunities that skyrmionic devices offer have been demonstrated by partners of the consortium. SkyANN attempts the next step showing how to integrate these devices with CMOS, a stride for both neuromorphic computing and semiconductors industry in Europe.
More importantly, SkyANN will generate significant impact through two major innovations: i) appropriate use of material processes and lithography to passively multiplex/demultiplex particles (affecting readily the energy demand for such operations across the layers of the neural networks) and ii) exploit different skyrmionic quasiparticles as information carriers while sharing the same physical medium (i.e. interconnect and possibly readout circuitry). The latter will reduce system footprint leveraging miniaturization and cost reduction.
More importantly, SkyANN will generate significant impact through two major innovations: i) appropriate use of material processes and lithography to passively multiplex/demultiplex particles (affecting readily the energy demand for such operations across the layers of the neural networks) and ii) exploit different skyrmionic quasiparticles as information carriers while sharing the same physical medium (i.e. interconnect and possibly readout circuitry). The latter will reduce system footprint leveraging miniaturization and cost reduction.