The company is located in:
With the rapid growth of artificial intelligence, the traditional von Neumann architecture is limited by the "storage wall", a huge gap between computational speed and the speed at which stored information can be delivered. In large-scale neural network computing tasks It has been unable to meet the rapid development of artificial intelligence requirements so far, a simulation of the human brain mode of operation, the calculation of the parallel computing is expected to achieve brain-like computing came into being by the human brain Synapses can enlighten memory and computation at the same time. The brain-like computing can simulate synaptic behavior with resistive memory. By changing the resistance state of resistive memory and changing its weight value in the network, According to the group's article published in Nature Newsletter in May 2017, a brain-like device can reduce the power consumption to a thousand One below
Figure 1 brain computing diagram
The company is located in:
As the most potential electronic synaptic device, resistive memory has the advantages of simple structure, low operating voltage, large-scale integration and compatibility with CMOS technology. However, the bi-directional multivalue resistive change characteristics of resistive memory required by brain- The working mechanism is not yet clear, and this problem severely limits the further performance optimization of the device, which means that the resistance of the device can change continuously from high to low or from low to high. Oxygen vacancy distribution disorder resistance model of resistive memory 'in the article for resistance of the oxide resistive memory bi-directional multi-value resistive behavior in-depth exploration using Monte Carlo simulation method to simulate oxygen vacancies dominated the resistance transition environment, proposed disorder The concept of degree quantifies the macroscopic resistance transformation caused by the distribution of microscopic oxygen vacancies.The simulation results show that the disordered distribution of oxygen vacancies is conducive to the realization of bidirectional multi-value resistance change, which makes the resistive memory for brain- Changes provide theoretical support and optimization guidance.
The company is located in:
Ordered and unordered oxygen vacancy distribution comparison chart
The company is located in:
Traditional devices and two-way continuous resistance device characteristics comparison chart
The company is located in:
Data holding capacity of resistive memory has a significant impact on device performance and working life.Although there is no lack of outstanding research results on the data retention capability of memory, the resistive memory used in brain-like computing requires the data retention and the traditional memory , The former requires a continuously varying conductance in the calculation, so more attention is paid to the data retention ability of the intermediate states (not only the high-resistance state and the low-resistance state), and little is known about this problem. For the first time, we study the statistical behavior of the intermediate state data retention in multi-valued resistive memory based on conductive filaments on a 1Kb array.This paper divides the 1024 device elements into 8 impedances, using the bidirectional modulation mode set the initial resistance state, and then baked at 125 ℃, observed a number of different resistive memory cell intermediate state under high temperature conductivity distribution over time, the results show that each resistance state The conductance distribution showed a symmetrical distribution with initial density and widening with time. The distribution at the observation point showed a normal distribution and the normal distribution The standard deviation of the cloth increases with time while the mean remains unchanged.The research results show the data failure law of the resistive memory array applied to the brain-like computing chip and also indicate the optimization of the reliability of the multivalued features of the resistive memory cells direction.
The company is located in:
Wu Huaqiang Group of Tsinghua University and Siu Yi Innovation Co., Ltd. have maintained long-term cooperation in the research of resistive memory. Dr. Chen Hongyu, a senior manager of Siu Yi Innovation participated in this research work and was co-author of the dissertation. He expressed his thanks to Professor Gao Bin And Zhao Meiran students, their research results not only provide theoretical and device basis for high-efficiency artificial intelligence chips, but also have an important significance for the design of high-density resistive memory chips.In the near future, Tsinghua University and Siu Yi Innovation will also work together to develop larger resistive memory arrays.
