- >> Division for the Establishment of Frontier Sciences
Spintronics/CMOS Hybrid Brain-Inspired Integrated Systems
- [ Professor* ]
- [ Assistant Professor* ]
Group Web Site
The purpose of the Spintronics/CMOS Hybrid Brain-Inspired Integrated System group is to break ground for a new AI approach across from the fundamental science of material and information to the devices, circuits, architecture and software technology for information generating, storing and recognition, based on both the concept of RIEC on realization of humanity-rich communication and its past achievements on hardware technology for information and communication. This group aims to study the basic technology for high efficient semiconductor integrated circuit combing the information storage and processing of the AI system together, and to lead and contribute the innovative development of high efficient and low power hardware technology for advanced flexible information processing and recognition like human brain. This group is trying to develop the novel brain-inspired computing system realizing the precise and real-time processing for information value judgment, choice and refusal by consistently evolving the proposal, design, verification and evaluation of the next-generation AI VLSI architecture, which is able to maximize the power consumption efficiency benefiting from the high speed and high endurance of the spin-device.
Spintronics/CMOS Hybrid Brain-Inspired Integrated System (Prof. Endoh)
The Spintronics/CMOS Hybrid Brain-Inspired Integrated System group aims to concentrate the scientific principle for AI computing, brain-inspired VLSI and spintronics/CMOS hybrid device/circuit/architecture technology, to construct the new system of spintronics/CMOS hybrid VLSIs, and to realize the high functional and ultra-low-power spintronics/CMOS hybrid brain- mimicking VLSI system (Fig.1). For such occasions, the entire research group is well-organized in two separated topic of “von Neumann type” and “non von Neumann type” with organical knowledge sharing, technology transfer and feedback.
Along with advancement of AI technology for next-generation information society, not only the supergiant data generation over Yotta (1024) byte but also the AI information processing on edge devices that appends the quality of information and supports the human thought become increasingly important. On the other hand, conventional AI implementation approaches based on volatile memories are very disadvantageous in power consumption efficiency and not feasible for practical use.
In this group, on “von Neumann” approach, a test chip for verifying accuracy enhancement and a prototype chip for In this group, on “von Neumann” approach, the prototype chips for verifying accuracy enhancement and the computational speed improvement of the designed nonvolatile adaptive K-means unsupervised learning processor are evaluated with actual measurement of 316-PIN BGA packaged chips. The higher accuracy and lower computational cost for both online and offline learning of real image data are achieved by the proposed processor even for those unbalanced training data set which is extremely difficult to conventional K-means learning processor. Moreover, as shown in Fig.2, the new adaptive convolutional computing circuit module is proposed and verified with FPGA implementing the light-weight FCNN structure for nonvolatile object detection (NOD) processor. The 55nm-CMOMS/56nm-MTJ fabrication and on-wafer measurement of the FCNN-based NOD processor is also completed for design issue extraction. On “non von Neumann” approach, the prototype design of 32K-Synapse /512-Neuron multi-core nonvolatile SNN pattern recognition processor leveraging self-directive power management is performed which verified the handwritten digit recognition and evaluated the power consumption performance.