It has been collaborating on memristors with French research lab CEA since 2020, and has now announced a collection of memristors packaged in an IC (described as ‘three elements by two sets’, photo).
Semiconductor process behind this IC were developed with CIES (Center for Innovative Integrated Electronic Systems) at Tohoku University, which has a 300mm-compatible spintronics semiconductor prototyping line.
The resulting integrated memristors are controllable, have low current leakage, and are robust against resistance drift over time (the output of a memristor is a variable resistance) allowing long-term data storage, said TDK, adding that they are also “expected to offer immunity to environmental influences”.
Together, TDK and CEA have built an AI circuit with four of the chips, and implemented an analogue algorithm which learned to separate three types of sound: music, speech and noise.
“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 uniquely capable of learning in a changing environment in real-time,” said the company. “In the demonstration, even when three types of sound were mixed with arbitrary ratios, the circuit was able to learn and separate the three types of sound in real-time.”
Analogue algorithms are of interest because their power consumption can be dramatically lower than conventional digital neural network implementations – TDK puts the figure at 100x lower.
“The human brain requires around 20W, which enables it to make more complex decisions than current digital AI processors, but with far lower power consumption,” it said.
Following the demonstration, TDK and Tohoku University will attempt to combine semiconductor and spintronic manufacturing processes.
“This integration has been achieved in the manufacturing of MRAM, a product similar to memristors, [and] Tohoku University [is] a leading academic institution in MRAM research and development,” said TDK.
