arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2007.12357physics.app-phcond-mat.mes-hall
收藏

用于自旋神经形态计算的超低功耗畴壁器件

Ultra-low Power Domain Wall Device for Spin-based Neuromorphic Computing

Durgesh Kumar, Chung Hong Jing, Chan JianPeng, Tianli Jin, Lim Sze Ter, Rachid Sbiaa, S. N. Piramanayagam

摘要:

神经形态计算(NC)作为一种实现低功耗智能设备的潜在技术正在获得广泛的认可。为了实现NC,研究人员研究了各种类型的合成神经元和突触器件,如忆阻器和自旋电子畴壁(DW)器件。相比之下,基于DW的神经元和突触具有潜在的更高的耐力。然而,为了实现低功率器件,需要低能量(通常低于pJ/bit)的DW运动。在这里,我们证明了畴壁运动在电流密度低至1 E6 A/m2通过定制的β-W自旋霍尔材料。通过我们的设计,我们实现了超低的钉扎场和电流密度降低了10000倍。将畴壁移动约20微米的距离所需的能量为0.4fJ,其转化为对于20 nm的位长度的0.4aJ/位的能量消耗。利用曲折畴壁器件结构,我们已经建立了一个突触应用的受控DW运动,并显示了超低能量自旋为基础的神经形态元素的方向。

英文摘要:

Neuromorphic computing (NC) is gaining wide acceptance as a potential technology to achieve low-power intelligent devices. To realize NC, researchers investigate various types of synthetic neurons and synaptic devices such as memristors and spintronic domain wall (DW) devices. In comparison, DW-based neurons and synapses have potentially higher endurance. However, for realizing low-power devices, DW motion at low energies - typically below pJ/bit - are needed. Here, we demonstrate domain wall motion at current densities as low as 1E6 A/m2 by tailoring the beta-W spin Hall material. With our design, we achieve ultra-low pinning fields and current density reduction by a factor of 10000. The energy required to move the domain wall by a distance of about 20 micrometers is 0.4 fJ, which translates into energy consumption of 0.4 aJ/bit for a bit-length of 20 nm. With a meander domain wall device configuration, we have established a controlled DW motion for synapse applications and have shown the direction to make ultra-low energy spin-based neuromorphic elements.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑