Frontier development of spintronics-based synapse and neuron for artificial neural networks

本研究的目的是为人工神经网络开发基于自旋电子学的器件。在前一年，我发表了使用随机磁隧道结（MTJ）的概率计算的实验结果，使用8个概率比特（p-bits）来解决整数分解问题。在最近的一年里，我探索了p-computing的巨大应用空间，以确定它的潜力。首先，我们与普渡大学合作，展示了基于mtj的p位可以代表玻尔兹曼机器并执行机器学习的实验结果。这些结果在人工智能机器学习领域具有重大意义，特别是在需要区域高效机器学习芯片的领域，如边缘计算。其次，在与加州大学圣巴巴拉分校（美国）的合作中，我们从理论上证明了MTJ结构可以改变以产生纳秒波动和独立性。这些结果将为构建高度集成的、大规模的基于mtj的p位网络奠定基础。最后，MTJ波动速度越快，求解时间越快。与实验室的其他成员一起，我们实验展示了一个平面内易轴的MTJ，其波动速度高达8ns，这是世界上最快的波动MTJ。

The objective of this research was to developed spintronics-based devices for artificial neural networks. In the previous year I published experimental results of a probabilistic computing using stochastic magnetic tunnel junctions (MTJ) to solve integer factorization using 8 probabilistic bits (p-bits).In this recent year, I have explored p-computing’s vast application space to determine it’s potential. First, in collaboration with Purdue Univ., we showed experimental results that MTJ-based p-bits can represent a Boltzmann machine and perform machine learning. These results have large implications in the realm of AI machine learning, especially in areas where area-efficient machine learning chips are desired, such as edge computing.Second, in a work with the Univ. of Calif. Santa Barbara (USA), we showed theoretically that the MTJ structure can be changed to produce nanosecond fluctuations and independence. These results are will lead to the foundation of building highly-integrated, large-scale MTJ-based p-bit networks.Finally, faster MTJ fluctuation speed leads to faster solution times. Along with other members in the lab, we experimentally demonstrated an MTJ with an in-plane easy axis with fluctuation speeds as fast as 8 ns, the world’s fastest fluctuating MTJ.