Recently, the research findings of the team of Professors Wang Xingsheng and Miao Xiangshui on the disruptive technology of memristor-based memory-computing integration, entitled "Reconfigurable and Efficient Implementation of 16 Boolean Logics and Full-Adder Functions with Memristor Crossbar for Beyond von Neumann In-Memory Computing", are published in the high-impact academic journal Advanced Science. Professor Wang Xingsheng and his Ph.D. student Song Yujie are the corresponding author and the first author of the paper, respectively. Advanced Science is an interdisciplinary premium open access journal founded by Wiley in 2014, which is dedicated to publishing high-level papers in various disciplines, with an impact factor of 16.806 in 2021.
(See https://onlinelibrary.wiley.com/doi/10.1002/advs.202200036 for details).
Current Research Status
Latency, energy consumption, and area are the main performance indicators of data processors. With the popularity of IoT, a large number of edge computing devices must compute and store massive amounts of data, which makes low-power become a key requirement. The separation of storage and computation units inherent in the current mainstream von Neumann computing architecture leads to frequent data transfer between units. The statistics on the usage frequency of the processor's main instructions show that data migration is the most frequently used instruction, and data moves back and forth between the high-speed CPU and low-speed memory frequently. When compared to the computing process, the load and store processes result in several orders of magnitude higher delays. Furthermore, data migration consumes one to two orders of magnitude more energy than computing, resulting in massive energy usage.
Unlike the existing transistor-based volatile voltage logic, the non-volatile computing in-memory technology is fundamentally different in principle. By introducing non-volatile physical variables (e.g., resistance) to participate in the operation and performing calculations based on physical laws such as Ohm's law and Kirchhoff's law in devices and arrays, such technologies are predictable energy-efficient computing solutions at this stage. In the current literature reported solutions for memristor-based logic operations, arithmetic operations, and matrix operations, most of the work can only support the physical implementation of a single computational method, and there are limitations in the application, which severely restricts the development of memristor-based in-memory computing. In addition, previous technical solutions have low computing efficiency and have no mature, systematic, and recognized computing theories and methods. Therefore, this research focus on developing the energy-efficient and functionally reconfigurable memristor in-memory computing theories and circuits.
The research team of Prof. Wang has developed a high-performance Ti/HfO2/TiN binary memristor, which has been integrated on a large scale on 8-inch wafers. Based on the resistive transformation behavior of the memristor, the research team has proposed a new V/R-R type for logic calculation based on the basic structure of "two parallel memristors connected in series with one resistor (2M1R)". It implements complete Boolean logic functions within 2 steps by configuring the port voltages. The working circuit unit structure is simple and uniform, with few calculation steps, cascadable, simple addressable, with low requirement for the number of voltage sources and reduced maximum operating voltage. The feasibility of the reconfigurable logic scheme is experimentally confirmed in the integrated Ti/HfO2/TiN memristor array.
This work provided an alternative scheme for future memory-and-computing integration technology. It achieved an important breakthrough in the field of in-memory logic computing, providing a logic algorithm support for the construction of non-von computational systems. It is expected to make new contributions to complex logic splitting, efficient hardware mapping, complex architecture and design methods for energy-efficient memristor_based memory arithmetic logic units (MALUs) in the future.
This work is partly supported by the National Key Research and Development Program of China, Hubei Yangtze Memory Laboratories, and Hubei Key Laboratory of Advanced Memories.
Source: School of Optical and Electronic Information
Edited by: Liu Tianxin, Jiang Jing