Low-Barrier Magnet Design for Efficient Hardware Binary Stochastic Neurons
Low-Barrier Magnet Design for Efficient Hardware Binary Stochastic Neurons
复制标题
DOI：
10.1109/lmag.2019.2910787
复制
发表时间：
2019-01-01
影响因子：
1.2
通讯作者：
Datta, Supriyo
中科院分区：
文献类型：
--
作者：
Hassan, Orchi;Faria, Rafatul;Datta, Supriyo
二进制随机神经元（BSN）构成了许多机器学习算法不可或缺的一部分，激发了该复杂功能的硬件加速器的开发。人们已经认识到，可以通过最小修改当今的磁磁性随机存储器（MRAM）设备来使用低屏障磁铁（LBM）来实现硬件BSN。确定这些基于LBM的BSN设计响应的关键参数是磁化tau（C）的相关时间。在这封信中，我们表明，对于具有低能屏障的磁铁（三角洲近似于k（b）t及以下），具有平面磁各向异性（IMA）的圆盘磁铁导致Tau（c）值为两个，这是两个值磁体的磁体小（c）小的数量级，具有垂直磁各向异性（PMA）。分析描述表明，tau（c）中的这种显着差异是由于ima磁体中的大型消灭磁场启用了一种凹陷般的波动机制。我们提供了基于旋转轨道扭矩MRAM和自旋转移扭矩MRAM的详细的能量延迟性能评估，该设计通过香料模拟使用低演形磁盘磁铁。这些设计表现出次纳秒响应时间，导致大约几个FEMTOJOULE的能量需求评估BSN功能，比具有更大表面积的数字CMOS实现的数量级低。尽管现代MRAM技术是基于PMA磁铁的，但在这封信中结果表明，低屏障圆形IMA磁铁可能更适合此应用。
Binary stochastic neurons (BSNs) form an integral part of many machine learning algorithms, motivating the development of hardware accelerators for this complex function. It has been recognized that hardware BSNs can be implemented using low-barrier magnets (LBMs) by minimally modifying present-day magnetoresistive random-access memory (MRAM) devices. A crucial parameter that determines the response of these LBM-based BSN designs is the correlation time of magnetization tau(c). In this letter, we show that, for magnets with low-energy barriers (Delta approximate to k(B)T and below), circular disk magnets with in-plane magnetic anisotropy (IMA) lead to tau(c) values that are two orders of magnitude smaller than tau(c) of magnets with perpendicular magnetic anisotropy (PMA). Analytical descriptions demonstrate that this striking difference in tau(c) is due to a precessionlike fluctuation mechanism that is enabled by the large demagnetization field in IMA magnets. We provide a detailed energy-delay performance evaluation of previously proposed BSN designs based on spin-orbit torque MRAM and spin-transfer torque MRAM employing low-barrier circular IMA magnets by SPICE simulations. The designs exhibit subnanosecond response times leading to energy requirements of approximately a few femtojoules to evaluate the BSN function, orders of magnitude lower than digital CMOS implementations with a much larger surface area. While modern MRAM technology is based on PMA magnets, results in this letter suggest that low-barrier circular IMA magnets may be more suitable for this application.