The element mimics the energy-efficient operation of the human brain and could cut the power consumption of AI applications to around 1/100th of traditional devices, according to TDK.
Collaborating with the French research organisation CEA (Alternative Energies and Atomic Energy Commission), TDK has been able to prove that its “spin-memristor” can serve as the basic element of a neuromorphic device. Going forward, TDK will collaborate with the Center for Innovative Integrated Electronic Systems at Tohoku University on the practical development of the technology.
Further developments in AI are expected to highlight the complexity around the computational processing of vast amounts of data, as well as the increased power consumption associated with the development of AI.
The human brain requires around 20 W of power, which enables it to make more complex decisions than current digital AI processors, but with far lower power consumption. Therefore, TDK’s aim was to develop a device that electrically simulates the synapses of the human brain: the memristor.
While conventional memory elements store data in the form of either 0 or 1, a “spin-memristor” can store data in analogue form, just as the brain does. This makes it possible to perform complex computations with ultra-low power consumption.
Although memristors for neuromorphic devices already exist, they face issues such as changes in resistance over time, difficulties in controlling the precise writing of data, and the need for control to ensure that data is retained.
TDK’s “spin-memristor” looks to address these issues and is expected to offer immunity to environmental influences and long-term data storage while reducing power consumption by cutting leakage current in existing devices.
TDK started collaborating with the CEA in 2020 and has developed an AI circuit equipped with a “spin-memristor” (3 elements × 2 sets × 4 chips) and confirmed its successful operation through a sound separation demonstration, proving that “spin-memristors” can serve as basic elements in AI circuits.
In the demonstration, even when three types of sound (music, speech, and noise) were mixed with arbitrary ratios, the circuit was able to learn and separate the three types of sound in real-time. In general machine learning, AI operations are performed based on data that the AI model has previously been trained on, but TDK’s device is capable of learning in a changing environment in real-time.
Now that it has been confirmed that the “spin-memristor” can serve as the basic element of a neuromorphic device, TDK said that it will look to further advance this project from basic development to the next stage, practical application.
The manufacturing of these products requires the integration of semiconductor and spintronic manufacturing processes, and this has been achieved in the manufacturing of MRAM, a product similar to memristors, and TDK has decided to pursue integrated technological development on a joint basis with Tohoku University, a leading academic institution in MRAM research and development.
Commenting Dr. Marc Duranton, Senior Fellow of CEA, said, “This research partnership is breaking new ground to develop more sustainable, reliable, highly efficient solutions that will meet the growing demands of modern AI applications.”
“AI semiconductors are extremely important for the information-oriented society of the future, but the societal issues are improving AI processing power and reducing power consumption,” explained Dr. Tetsuo Endoh, the director of CIES Tohoku University. “In light of this social demand, TDK's AI semiconductor development program, which fuses memristor and spintronics technology, is extremely important. We will do our best to contribute to this project with the academic knowledge held by Tohoku University and the manufacturing technology of the 12-inch prototype line.”
