TDK successfully develops "Spin memristor" for neuromorphic devices

     TDK Corporation announced that it has successfully developed an ultra-low energy neuromorphic device, the spin memristor. By simulating the energy-efficient operation of the human brain, the device can reduce the energy consumption of artificial intelligence (AI) applications to one-hundredth that of traditional devices. In collaboration with the French research agency, the Commission on Atomic Energy and Alternative Energy (CEA), TDK demonstrated that its "spin memristor" could serve as a basic element of a neuromorphic device. In the future, TDK will cooperate with the Innovation Integrated Electronic Systems R&D Center of Tohoku University in Japan to carry out practical development of this technology.

     In recent years, as AI technology continues to evolve and digital transformation (DX) continues to advance, it is expected that the energy consumption of utilizing big data and AI technology will increase significantly, and make certain issues more prominent - such as the complexity associated with the computational processing of massive amounts of data and the increasing electricity consumption associated with the development of AI. TDK will continue its efforts to address these social and environmental issues.

     The human brain operates at about 20W, which allows it to make more complex decisions than existing digital AI processors, but with much lower power consumption. Therefore, TDK's goal is to develop a device that can electronically mimic human brain synapses: the memristor. While traditional memory elements store data in the form of a digital 0 or 1, "spin memristors" can be used as analog memory elements, similar to the human brain. This allows the device to perform complex calculations with ultra-low energy consumption.

     Although memristors for neuromorphic devices are not new, they all face a variety of problems, such as resistance changes over time, difficulty controlling the precise writing of data, and the need to implement controls to ensure data retention. TDK's "spin memristor" successfully addresses these issues and is expected to reduce power consumption by reducing leakage current in existing devices, while providing protection from environmental changes and long-term data storage.

     To achieve this goal, TDK began working with CEA in 2020. With CEA's support, we successfully developed an AI circuit equipped with a "spin memristor" (3 components x 2 groups x4 chips), and verified its successful operation through acoustic separation demonstration, fully proving that "spin memristor" can be used as a basic component of AI circuits. During the previous demonstration, even if three types of sounds (music, speech, and noise) were mixed in any proportion, the circuit was always able to learn and separate the three sounds in real time. In a typical machine learning process, AI operations are based on data the AI model was previously trained on, but TDK's devices have the unique ability to learn in real time in a constantly changing environment.