Analog computing with crossbars of memristors is a promising approach to build compact energy-efficient vector-by-matrix multiplier (VMM), a key block in many data-intensive algorithms. However, device non-linearity, process variations, interconnect parasitics, noise, and memory state drift limit the computing precision of such systems. In this article, we investigate the impact of such non-idealities in analog current-mode memristive VMMs through simulations and experiments on the most prospective passive crossbars. We show that there is an optimal tuning voltage to minimize the computation error. Furthermore, error balancing and bootstrapping are introduced as two techniques for improving the precision. It is also shown that when size of N × N crossbar is scaled up, the optimum interconnect wire conductance should increase quadratically with N to preserve the computing precision when using naive error balancing approach, and that the differential scheme is imperative for temperature insensitive operation and also to reduce the IR-drop effect.

中文翻译：

---

基于忆阻器的逐矩阵乘法器计算精度的内在界限

使用忆阻器交叉开关的模拟计算是构建紧凑型节能矢量矩阵乘法器 (VMM) 的一种很有前景的方法，VMM 是许多数据密集型算法中的关键模块。然而，器件非线性、工艺变化、互连寄生、噪声和存储器状态漂移限制了此类系统的计算精度。在本文中，我们通过对最有前景的无源交叉开关的模拟和实验来研究此类非理想性对模拟电流模式忆阻 VMM 的影响。我们表明有一个最佳的调谐电压来最小化计算误差。此外，引入了误差平衡和自举作为提高精度的两种技术。还表明，当 N × N 横杆的尺寸按比例放大时，