Topological Insulator based Transistors for Neuromorphic Computer Systems

随着CMOS的尺寸缩放接近基本极限，需要新的信息处理设备和微架构来扩展当今基于冯诺依曼的计算机的能力。大数据、物联网（IoT）、自主系统、亿级计算等新信息时代的应用需求需要能够以超低功耗运行的人工智能（AI）芯片，以最终达到并超过人脑的效率。这种低功耗计算架构的主要候选者是尖峰神经网络（SNN），这是一种基于事件的学习，没有任何外部存储器，消耗很少的能量。然而，模拟SNN -神经元决策过程的硬件组件消耗高功率，因此目前限制了AI的性能。该项目汇集了来自格拉斯哥大学和谢菲尔德大学的专业知识，以探索三元拓扑绝缘体（Bi 2 Te 2Se）作为泄漏集成和激发神经元的低电压材料平台。通过控制器件表面状态，沟道的电导将被调谐，足以将电流尖峰从漏极传输到源极端子。这一概念的成功演示将导致人工神经网络硬件实现的里程碑式飞跃。这项工作是令人兴奋和冒险的，因为到目前为止，文献中只有两个TTI-FET（分别来自2011年和2012年），并且它们都没有被设想为按照我们在这项工作中的计划进行操作。除了既定目标之外，还将产生材料科学，自旋电子学和量子计算社区感兴趣的新知识，从该项目中获得的知识有可能解决这些研究领域的现有瓶颈。

As dimensional scaling of CMOS is approaching fundamental limits, new information processing devices and micro-architectures are required to extend the capability of present-day Von-Neumann based computers. Application pulls for the new information age such as big data, Internet of Things (IoT), autonomous systems, exascale computing require artificial intelligence (AI) chips that can operate on super ultra-low power to ultimately match and exceed the efficiency of the human brain. A leading candidate for such low power computing architecture is the spiking neural network (SNN), which is one type of event-based learning without any external memory that consumes very little energy. However, the hardware component that emulates the decision-making process of a SNN - a neuron, consumes high power and therefore presently limits the performance of AI. This project brings together expertise from University of Glasgow and Sheffield to explore ternary Topological Insulator (Bi2Te2Se) as a low voltage material platform for a leaky integrate and fire neuron. By controlling the device surface states, the conductance of the channel will be tuned, sufficient enough to transmit a current spike from the drain to the source terminals. A successful demonstration of this concept will result in a milestone leap in hardware implementations of Artificial Neural Networks. The work is exciting and adventurous because there are only two reported TTI-FETs in the literature so far (from 2011and 2012 respectively), and neither of them is being envisaged to be operated as we plan in this work. Beyond the stated goals, new knowledge will be generated of interest to materials science, spintronics and quantum computing communities, where the knowledge gained from this project offers the potential to address existing bottlenecks in these research fields.