Hardware neural networks (HNNs) which use synapse device (SD) arrays show promise as an approach to energy efficient parallel computation of massive vector-matrix multiplication. To maximize the inference accuracy of application-specific HNNs, we propose a highly reliable 2-terminal SD with fixed resistance based on WOx films. First, we investigate the device requirements of an array-based HNN through MATLAB and SPICE simulations taking into account the parasitic resistance effects in the array. On top of that, to fabricate the SD we utilize the intrinsic properties of the WOx film, which exhibits substantial changes in conductivity from 10−8 to 104 Ω−1cm−1 by varying the oxygen vacancy concentration. After the process optimization of oxide stoichiometry and electrode materials, we can form nanoscale WO×-based SDs with excellent conductance uniformity and I-V linearity. Our results show that inference accuracy is significantly improved by using WOx-based SD arrays even in advanced node scaling. Through experimental hardware implementation, 16 × 16 pixel images are correctly classified and we show the potential of WOx-based SD for future large-scale HNN applications.

中文翻译：

---

基于 WOx 的突触设备具有出色的电导均匀性，适用于硬件神经网络

使用突触设备 (SD) 阵列的硬件神经网络 (HNN) 显示出作为大规模向量矩阵乘法的节能并行计算方法的前景。为了最大限度地提高特定应用 HNN 的推理精度，我们提出了一种基于 WOx 薄膜的具有固定电阻的高度可靠的 2 端 SD。首先，考虑到阵列中的寄生电阻效应，我们通过 MATLAB 和 SPICE 仿真研究了基于阵列的 HNN 的设备要求。最重要的是，为了制造 SD，我们利用 WOx 薄膜的固有特性，通过改变氧空位浓度，其电导率从 10-8 到 104 Ω-1cm-1 显着变化。经过氧化物化学计量和电极材料的工艺优化，我们可以形成具有优异电导均匀性和 IV 线性的纳米级 WOx 基 SD。我们的结果表明，即使在高级节点缩放中，通过使用基于 WOx 的 SD 阵列，推理精度也显着提高。通过实验硬件实现，16 × 16 像素图像被正确分类，我们展示了基于 WOx 的 SD 在未来大规模 HNN 应用中的潜力。