Periodic Reporting for period 3 - RadioSpin (DEEP OSCILLATORY NEURAL NETWORKS COMPUTING AND LEARNING THROUGH THE DYNAMICS OF RF NEURONS INTERCONNECTED BY RF SPINTRONIC SYNAPSES)
Période du rapport: 2024-03-01 au 2025-08-31
The overall goal of RadioSpin is to build a hardware neural network that computes using neural dynamics as in the brain, has a deep layered architecture as in the neocortex, but runs and learns faster, by seven orders of magnitude. For this purpose, we will use ultrafast radio-frequency (RF) oscillators to imitate the rich, reconfigurable dynamics of biological neurons. Within the RadioSpin project, we will develop a new breed of nanosynapses, based on spintronics technology, that directly process the RF signals sent by neurons and interconnects them layer-wise. We will demonstrate and benchmark our concept by building a lab-scale prototype that co-integrates for the first time CMOS RF neurons with spintronic RF synapses. We will develop brain-inspired algorithms harnessing oscillations, synchrony and edge-of-chaos for computing and show that they can run on RadioSpin deep network RF technology. Finally, we will benchmark RadioSpin technology for biomedical and RF fingerprinting applications where fast and low energy consumption classification of RF signals are key.
The overall goal of RadioSpin is to build a hardware neural network that computes using neural dynamics as in the brain, has a deep layered architecture as in the neocortex, but runs and learns faster, by seven orders of magnitude. For this purpose, we will use ultrafast radio-frequency (RF) oscillators to imitate the rich, reconfigurable dynamics of biological neurons. Within the RadioSpin project, we will develop a new breed of nanosynapses, based on spintronics technology, that directly process the RF signals sent by neurons and interconnects them layer-wise. We will demonstrate and benchmark our concept by building a lab-scale prototype that co-integrates for the first time CMOS RF neurons with spintronic RF synapses. We will develop brain-inspired algorithms harnessing oscillations, synchrony and edge-of-chaos for computing and show that they can run on RadioSpin deep network RF technology. Finally, we will benchmark RadioSpin technology for biomedical and RF fingerprinting applications where fast and low energy consumption classification of RF signals are key.
Progress made towards the RadioSpin overall goal during RP3:
During this reporting period, we have made significant progress towards each objective, as detailed in the section below. The main achievements of the project in RP3 are:
- Successful optimization of the spintronic devices (improved efficiency, lower variability, addition of non-volatile frequency control).
- First experimental demonstration of a fully spintronic multilayer neural network and classification of RF signals.
- Completion of the CMOS circuits design and layout for demonstrator run #1, sent to the foundry.
- Post-processing of the CMOS wafers to deposit the spintronic nanodevices
- Electrical testing of the CMOS circuits and of the co-integrated systems
- Development of recurrent neural network algorithms and validation on time series benchmark tasks.
- Successful completion of a clinical trial in several hospital, resulting in a dataset for the RF mammography use-case.
- Classification of the drone dataset (RF fingerprinting use-case) with the improved spintronic neural network simulator, with over 99 % accuracy.
The key expected achievement during the rest of the project are:
- Achieving the co-integration of the spintronic synapses on the CMOS circuits.
- Assembly, validation and use of the demonstration.
- RF classification with the demonstrator.
- Using the simulator to classify the data from the RF mammography clinical trail
- Benchmarking the performance of RadioSpin technology.
Progress made towards the RadioSpin overall goal during RP3:
During this reporting period, we have made significant progress towards each objective, as detailed in the section below. The main achievements of the project in RP3 are:
- Successful optimization of the spintronic devices (improved efficiency, lower variability, addition of non-volatile frequency control).
- First experimental demonstration of a fully spintronic multilayer neural network and classification of RF signals.
- Completion of the CMOS circuits design and layout for demonstrator run #1, sent to the foundry.
- Post-processing of the CMOS wafers to deposit the spintronic nanodevices
- Electrical testing of the CMOS circuits and of the co-integrated systems
- Development of recurrent neural network algorithms and validation on time series benchmark tasks.
- Successful completion of a clinical trial in several hospital, resulting in a dataset for the RF mammography use-case.
- Classification of the drone dataset (RF fingerprinting use-case) with the improved spintronic neural network simulator, with over 99 % accuracy.
The key expected achievement during the rest of the project are:
- Achieving the co-integration of the spintronic synapses on the CMOS circuits.
- Assembly, validation and use of the demonstration.
- RF classification with the demonstrator.
- Using the simulator to classify the data from the RF mammography clinical trail
- Benchmarking the performance of RadioSpin technology.
RadioSpin proposes solutions to achieve AI acceleration at all levels: from novel nanodevices to system computations, from physics to microelectronics to algorithms. RadioSpin is already based on two patents by CNRS and Thales and further IP will be developed and protected during the project and in the years to follow. RadioSpin proposes a ground-breaking concept for building hardware neural networks: spin-RF neuromorphic edge intelligence. Its objectives are to accelerate AI and decrease its energy consumption. In this way, RadioSpin will contribute to minimizing the environmental impact linked to the increasing amount of electricity consumed by ICT. Furthermore, RadioSpin will unlock novel applications that cannot be supported by current hardware such as on-the-fly classification of RF signals, for biomedical applications RF fingerprinting, but also fully-autonomous vehicles and robots, which are important in applications from Automotive to Space. RadioSpin will therefore enable new products to be manufactured that can then potentially lead to entirely new services and business where AI can help in spanning extremely diverse aspects of our everyday life such as health (X-ray free radiology), telecommunications, transportation (air traffic management, autonomous cars), aerospace (aircrafts and satellites), education (teacher robots) and personal assistants (smartphones, portable automatic translation devices).
RadioSpin envisages the following impacts:
1. RadioSpin will bring neuromorphic engineering at a level where it can be benchmarked in terms of performance, power consumption, size, latency or other relevant metric (e.g. learning capacity, speed or plasticity) on real-life applications.
2. RadioSpin paves the way to market take-up for existing and new neuromorphic application areas Health and biomedical applications.
3. RadioSpin will help stimulate the emergence of a European innovation ecosystem around neuromorphic engineering, well beyond the world of research alone.
4. Economic Impact: RadioSpin with the two selected use cases in the fields of health and RF fingerprinting for device classification targets highly promising sectors with clear potential for economic impact. Progress achieved during RP1 and first results in technology benchmarking for the two use cases have shown the high potential of RadioSpin for future industrialisation resulting in economic impact.
5. Impact on European policies and strategies: RadioSpin will contribute to major strategic European initiatives via an ambitious concept of RF edge intelligence merging Spintronics and Neuromorphics, both “beyond CMOS” technologies. Key roadmaps identified to be of highest relevance also for the strategic direction of RadioSpin are: 1) the European Strategy for Micro-and Nanoelectronic component and systems, 2) the Nanoelectronics roadmap for Europe (NEREID), 3) the International Roadmap for Devices and Systems (IRDS™) and 4) The Electronic Components & Systems (ECS) Strategic Research Agenda (SRA).
6. Impact on low energy consumption and environmentally friendly technology: RadioSpin technology offers a low energy path to neural network deployment, both in data centres and edge devices.
RadioSpin envisages the following impacts:
1. RadioSpin will bring neuromorphic engineering at a level where it can be benchmarked in terms of performance, power consumption, size, latency or other relevant metric (e.g. learning capacity, speed or plasticity) on real-life applications.
2. RadioSpin paves the way to market take-up for existing and new neuromorphic application areas Health and biomedical applications.
3. RadioSpin will help stimulate the emergence of a European innovation ecosystem around neuromorphic engineering, well beyond the world of research alone.
4. Economic Impact: RadioSpin with the two selected use cases in the fields of health and RF fingerprinting for device classification targets highly promising sectors with clear potential for economic impact. Progress achieved during RP1 and first results in technology benchmarking for the two use cases have shown the high potential of RadioSpin for future industrialisation resulting in economic impact.
5. Impact on European policies and strategies: RadioSpin will contribute to major strategic European initiatives via an ambitious concept of RF edge intelligence merging Spintronics and Neuromorphics, both “beyond CMOS” technologies. Key roadmaps identified to be of highest relevance also for the strategic direction of RadioSpin are: 1) the European Strategy for Micro-and Nanoelectronic component and systems, 2) the Nanoelectronics roadmap for Europe (NEREID), 3) the International Roadmap for Devices and Systems (IRDS™) and 4) The Electronic Components & Systems (ECS) Strategic Research Agenda (SRA).
6. Impact on low energy consumption and environmentally friendly technology: RadioSpin technology offers a low energy path to neural network deployment, both in data centres and edge devices.