A crossbar array of magnetoresistive memory devices for in-memory computing
A crossbar array of magnetoresistive memory devices for in-memory computing
复制标题
DOI：
10.1038/s41586-021-04196-6
复制
发表时间：
2022-01-13
期刊：
影响因子：
64.8
通讯作者：
Kim, Sang Joon
中科院分区：
文献类型：
--
作者：
Jung, Seungchul;Lee, Hyungwoo;Kim, Sang Joon
借用模拟技术的人工神经网络的实现可能为全数字方法提供低功耗替代方案（1-3）。一个值得注意的例子是基于非易失性存储器（4-7）的交叉阵列的存储器内计算，其以模拟方式执行人工神经网络中流行的乘法-累加运算。各种非易失性存储器-包括电阻式存储器（8-13）、相变存储器（14，15）和闪存（16-19）-已经用于这种方法。然而，尽管该技术具有诸如耐久性和大规模商业化（5）的实际优点，但开发自旋转移矩磁阻随机存取存储器（MRAM）的交叉阵列（20-22）仍然具有挑战性。困难源于MRAM的低电阻，这将导致在使用电流求和用于模拟乘法-累加操作的常规交叉开关阵列中的大功率消耗。在这里，我们报告了一个64 × 64的交叉阵列的MRAM单元，克服了低电阻的问题与架构，使用模拟乘法累加运算的电阻求和。该阵列与28纳米互补金属氧化物半导体技术的读出电子集成在一起。使用这个数组，一个两层感知器实现分类10,000修改的国家标准和技术研究所的数字的准确率为93.23%（软件基线：95.24%）。在模拟更深层次的八层Visual Geometry Group-8神经网络时，分类准确率提高到98.86%（软件基线：99.28%）。我们还使用该阵列实现了十层神经网络中的单层，以实现人脸检测，准确率为93.4%。
Implementations of artificial neural networks that borrow analogue techniques could potentially offer low-power alternatives to fully digital approaches(1-3). One notable example is in-memory computing based on crossbar arrays of non-volatile memories(4-7) that execute, in an analogue manner, multiply-accumulate operations prevalent in artificial neural networks. Various non-volatile memories-including resistive memory(8-13), phase-change memory(14,15) and flash memory(16-19)-have been used for such approaches. However, it remains challenging to develop a crossbar array of spin-transfer-torque magnetoresistive random-access memory (MRAM)(20-22), despite the technology's practical advantages such as endurance and large-scale commercialization(5). The difficulty stems from the low resistance of MRAM, which would result in large power consumption in a conventional crossbar array that uses current summation for analogue multiply-accumulate operations. Here we report a 64 x 64 crossbar array based on MRAM cells that overcomes the low-resistance issue with an architecture that uses resistance summation for analogue multiply-accumulate operations. The array is integrated with readout electronics in 28-nanometre complementary metal-oxide-semiconductor technology. Using this array, a two-layer perceptron is implemented to classify 10,000 Modified National Institute of Standards and Technology digits with an accuracy of 93.23 per cent (software baseline: 95.24 per cent). In an emulation of a deeper, eight-layer Visual Geometry Group-8 neural network with measured errors, the classification accuracy improves to 98.86 per cent (software baseline: 99.28 per cent). We also use the array to implement a single layer in a ten-layer neural network to realize face detection with an accuracy of 93.4 per cent.