Computing‐in‐memory (CIM) architecture inspired by the hierarchy of human brain is proposed to resolve the von Neumann bottleneck and boost acceleration of artificial intelligence. Whereas remarkable progress has been achieved for CIM, making further improvements in CIM performance is becoming increasingly challenging, which is mainly caused by the disparity between rapid evolution of synaptic arrays and relatively slow progress in building efficient neuronal devices. Specifically, dedicated efforts are required toward developments of more advanced activation units in terms of both optimized algorithms and innovative hardware implementations. Here a novel bio‐inspired dendrite function‐like neuron based on negative‐differential‐resistance (NDR) behavior is reported and experimentally demonstrates this design as a more efficient neuron. By integrating electrochemical random‐access memory (ECRAM) with ionic regulation, the tunable NDR neuron can be trained to enhance neural network performances. Furthermore, based on a high‐density RRAM chip, fully hardware implementation of CIM is experimentally demonstrated by integrating NDR neuron devices with only a 1.03% accuracy loss. This work provides 516 × and 1.3 × 105 × improvements on LAE (Latency‐Area‐Energy) property, compared to the digital and analog CMOS activation circuits, respectively. With device‐algorithm co‐optimization, this work proposes a compact and energy‐efficient solution that pushes CIM‐based neuromorphic computing into a new paradigm.

中文翻译：

---

通过集成可训练树突神经元和高密度 RRAM 芯片实现全硬件忆阻神经形态计算

受人脑层次结构启发，提出了内存计算（CIM）架构，以解决冯诺依曼瓶颈并促进人工智能的加速。尽管 CIM 已经取得了显着的进展，但进一步提高 CIM 性能变得越来越具有挑战性，这主要是由于突触阵列的快速进化与构建高效神经元设备的相对缓慢进展之间的差距造成的。具体来说，需要在优化算法和创新硬件实现方面致力于开发更先进的激活单元。这里报道了一种基于负微分电阻（NDR）行为的新型仿生树突功能神经元，并通过实验证明了这种设计是一种更有效的神经元。通过将电化学随机存取存储器 (ECRAM) 与离子调节相结合，可训练可调谐 NDR 神经元以增强神经网络性能。此外，基于高密度RRAM芯片，通过集成NDR神经元器件，实验证明了CIM的完全硬件实现，精度损失仅为1.03%。这项工作提供了 516 × 和 1.3 × 105× 分别与数字和模拟 CMOS 激活电路相比，LAE（延迟区域能量）属性有所改进。通过设备算法协同优化，这项工作提出了一种紧凑且节能的解决方案，将基于 CIM 的神经拟态计算推向了新的范式。