As the massive usage of artificial intelligence techniques spreads in the economy, researchers are exploring new techniques to reduce the energy consumption of Neural Network (NN) applications, especially as the complexity of NNs continues to increase. Using analog Resistive RAM devices to compute matrix-vector multiplication in O(1) time complexity is a promising approach, but it is true that these implementations often fail to cover the diversity of non-linearities required for modern NN applications. In this work, we propose a novel approach where Resistive RAMs themselves can be reprogrammed to compute not only the required matrix multiplications but also the activation functions, Softmax, and pooling layers, reducing energy in complex NNs. This approach offers more versatility for researching novel NN layouts compared to custom logic. Results show that our device outperforms analog and digital field-programmable approaches by up to 8.5× in experiments on real-world human activity recognition and language modeling datasets with convolutional neural network, generative pre-trained Transformer, and long short-term memory models.

随着人工智能技术在经济中的广泛应用，研究人员正在探索降低神经网络（NN）应用能耗的新技术，特别是随着神经网络的复杂性不断增加。使用模拟电阻 RAM 器件以O (1) 时间复杂度计算矩阵向量乘法是一种很有前景的方法，但这些实现确实常常无法涵盖现代 NN 应用所需的非线性多样性。在这项工作中，我们提出了一种新颖的方法，其中电阻 RAM 本身可以重新编程，不仅可以计算所需的矩阵乘法，还可以计算激活函数、Softmax 和池化层，从而减少复杂神经网络中的能量。与自定义逻辑相比，这种方法为研究新颖的神经网络布局提供了更多的多功能性。结果表明，在使用卷积神经网络、生成式预训练 Transformer 和长短期记忆模型进行的真实人类活动识别和语言建模数据集实验中，我们的设备的性能比模拟和数字现场可编程方法高出 8.5 倍。