Periodic Reporting for period 1 - SPIN-ION (Hybrid Spintronic Synapses for Neuromorphic Computing)
Berichtszeitraum: 2023-05-01 bis 2024-04-30
The increasing scale of deep neural networks (DNN) and their growing application space have produced demand for more energy-efficient artificial-intelligence-specific hardware. Over the past decade, artificial intelligence algorithms have achieved human-level performance on increasingly complex tasks at the cost of increased neural network size, computing resources, and energy consumption for training deep learning models.
Classical computing approaches implemented in CPUs and GPUs are still based on Von Neuman architecture where computation and memory are separated. For this reason, about 80% energy is consumed in moving the data.
Emerging non-volatile technologies such as Magnetic Random-Access Memory (MRAM), are inherently low power (thanks to non-volatility) while they can also mimic both the functionality and structure of the human brain. This approach, known as neuromorphic computing, offers the possibility of further reducing the energy consumption dramatically by bringing AI directly to embedded devices for edge computing.
At Spin-Ion Technologies we have developed a new manufacturing solution based on ion beam processes to precisely engineer magnetic properties of spintronic devices at the atomic scale. In this project, we will develop new neuromorphic chips based on low power synapses and composed of non-volatile magnetic devices customized by our ion beam process. This demonstrator will overcome both catastrophic forgetting (so-called metaplasticity) and reduce device variability, hence greatly advancing the development of highly efficient, robust hardware amenable to neural applications on the edge.
Classical computing approaches implemented in CPUs and GPUs are still based on Von Neuman architecture where computation and memory are separated. For this reason, about 80% energy is consumed in moving the data.
Emerging non-volatile technologies such as Magnetic Random-Access Memory (MRAM), are inherently low power (thanks to non-volatility) while they can also mimic both the functionality and structure of the human brain. This approach, known as neuromorphic computing, offers the possibility of further reducing the energy consumption dramatically by bringing AI directly to embedded devices for edge computing.
At Spin-Ion Technologies we have developed a new manufacturing solution based on ion beam processes to precisely engineer magnetic properties of spintronic devices at the atomic scale. In this project, we will develop new neuromorphic chips based on low power synapses and composed of non-volatile magnetic devices customized by our ion beam process. This demonstrator will overcome both catastrophic forgetting (so-called metaplasticity) and reduce device variability, hence greatly advancing the development of highly efficient, robust hardware amenable to neural applications on the edge.
The project has shown very good technical and scientific progress within this first reporting period. All scientific deliverables and milestones have been delivered. The main achievements includes :
-State of the art Magnetic Tunnel Junctions stacks with high perpendicular anisotropy and high Tunneling Magnetoresistance (WP1)
-Irradiated Magnetic Tunnel Junction stacks with modulated anisotropy (WP1)
-Irradiated memory chips with low variability and modulated anisotropy (WP1)
-Irradiated MTJ stacks and devices through a mask (WP2)
-Incorporation of metaplasticity in Binary Neural Network algorithm (WP3)
- Writing process demonstrated in 15X15 Crossbarr arrays (WP4)
-State of the art Magnetic Tunnel Junctions stacks with high perpendicular anisotropy and high Tunneling Magnetoresistance (WP1)
-Irradiated Magnetic Tunnel Junction stacks with modulated anisotropy (WP1)
-Irradiated memory chips with low variability and modulated anisotropy (WP1)
-Irradiated MTJ stacks and devices through a mask (WP2)
-Incorporation of metaplasticity in Binary Neural Network algorithm (WP3)
- Writing process demonstrated in 15X15 Crossbarr arrays (WP4)
Continuous learning is paramount in the vast majority of today's AI applications, operating in real-world settings alongside humans, as they must adapt to remain effective and efficient. By integrating our hardware neuromorphic chip with continual learning algorithms, we ensure the sustainability of edge AI solutions.
The ongoing research within the EIC project holds the potential to catalyze significant advancements in both hardware and software solutions capable of implementing metaplasticity at ultra-low power. This could unlock new functionalities and drive the development of products tailored for autonomous lifelong learning applications. Such advances are particularly pertinent in critical sectors like medical monitoring, Industry 4.0 IoT devices, autonomous vehicles, robotics, and defense, where adaptability to evolving environments is paramount without compromising existing knowledge.
While Europe excels in embedded system software, it lags behind in hardware development despite its academic excellence in emerging technologies. Our project aims to bridge this gap, transforming Europe's scientific leadership into an industrial reality and enabling Europe to regain control over its digital supply chain while securing its technological sovereignty. Our aspiration to emerge as a global leader in edge AI computing is fully aligned with the EU digital agenda and the European Chips Act.
The project has made significant strides towards achieving broad technological impact:
• We have identified multiple hardware and software outcomes that will drive the development of metaplastic Binary Neural Networks hardware at ultra-low power.
• We have validated an ion beam pilot line together with specific metrology tools to address scalability, pre-industralization and meet the requirements of our customers. Ongoing discussions with national foundries aim to establish our pilot line.
• Our technology has attracted interest from three chip makers for potential collaboration and use case development with Spin-Ion Technologies.
• We have been identified by YOLE as an important player in neuromorphic computing (Neuromorphic Computing, Memory and Sensing 2024)
• Initiatives to protect our technology through patent filings and trade secrets have started, ensuring its security and future economic viability.
• We have launched our internationalization in north America by developping partnerships with MILA and C2MI in Quebec
In term of economic impact, the neuromorphic computing market is expected to take-off from 2026 as neuromorphic-based technologies mature. This timeline aligns well with our transition project, enabling Spin-Ion to onboard initial users by 2025. Spin-Ion will be a key enabler in unlocking and accelerating the market for edge AI chips by:
• Drastically reducing power consumption and facilitating on-chip learning capabilities.
• Accelerating the integration of our manufacturing process through the deployment of our pilot line in France.
• Establishing the company as a leading IP provider, offering both hardware and software platforms.
• Supporting other European SMEs involved in developing edge AI systems.
In term of societal impact, our solution enables reliable, large-scale, and environmentally sustainable deployment of edge/IoT devices and applications which will benefit European society with activities such as monitoring of health, mobility, smart cities, smart home and entertainment. We aim to help the general public understand and appreciate our technologies & ramifications.
The ongoing research within the EIC project holds the potential to catalyze significant advancements in both hardware and software solutions capable of implementing metaplasticity at ultra-low power. This could unlock new functionalities and drive the development of products tailored for autonomous lifelong learning applications. Such advances are particularly pertinent in critical sectors like medical monitoring, Industry 4.0 IoT devices, autonomous vehicles, robotics, and defense, where adaptability to evolving environments is paramount without compromising existing knowledge.
While Europe excels in embedded system software, it lags behind in hardware development despite its academic excellence in emerging technologies. Our project aims to bridge this gap, transforming Europe's scientific leadership into an industrial reality and enabling Europe to regain control over its digital supply chain while securing its technological sovereignty. Our aspiration to emerge as a global leader in edge AI computing is fully aligned with the EU digital agenda and the European Chips Act.
The project has made significant strides towards achieving broad technological impact:
• We have identified multiple hardware and software outcomes that will drive the development of metaplastic Binary Neural Networks hardware at ultra-low power.
• We have validated an ion beam pilot line together with specific metrology tools to address scalability, pre-industralization and meet the requirements of our customers. Ongoing discussions with national foundries aim to establish our pilot line.
• Our technology has attracted interest from three chip makers for potential collaboration and use case development with Spin-Ion Technologies.
• We have been identified by YOLE as an important player in neuromorphic computing (Neuromorphic Computing, Memory and Sensing 2024)
• Initiatives to protect our technology through patent filings and trade secrets have started, ensuring its security and future economic viability.
• We have launched our internationalization in north America by developping partnerships with MILA and C2MI in Quebec
In term of economic impact, the neuromorphic computing market is expected to take-off from 2026 as neuromorphic-based technologies mature. This timeline aligns well with our transition project, enabling Spin-Ion to onboard initial users by 2025. Spin-Ion will be a key enabler in unlocking and accelerating the market for edge AI chips by:
• Drastically reducing power consumption and facilitating on-chip learning capabilities.
• Accelerating the integration of our manufacturing process through the deployment of our pilot line in France.
• Establishing the company as a leading IP provider, offering both hardware and software platforms.
• Supporting other European SMEs involved in developing edge AI systems.
In term of societal impact, our solution enables reliable, large-scale, and environmentally sustainable deployment of edge/IoT devices and applications which will benefit European society with activities such as monitoring of health, mobility, smart cities, smart home and entertainment. We aim to help the general public understand and appreciate our technologies & ramifications.