All-Digital Time-Domain CNN Engine Using Bidirectional Memory Delay Lines for Energy-Efficient Edge Computing

📋 论文概要

提出了一种全数字时域卷积神经网络（CNN）引擎，利用双向存储器延迟线实现低功耗乘积累加（MAC）操作，解决了边缘计算设备中CNN的能效瓶颈。

💡 主要创新点

• 使用双向存储器延迟线实现全数字时域MAC计算，避免了模拟电路和ADC的功耗开销。
• 时间域计算方式通过延迟线编码权重和输入数据，实现高能效的卷积操作。

🏷 关键词

📄 原文摘要

Convolutional Neural Networks (CNN) provide superior classification accuracy in a variety of machine learning applications, such as image/speech/sensor data processing. However, CNNs require intensive compute and memory resources making it challenging to employ in energy-constrained edge-computing devices. Specifically, Multiply-and-Accumulate (MAC) operations consume a significant portion of the total CNN energy [1]. Various analog compute techniques using charge manipulation schemes and A/D converters, as well as frequency-modulation-based approaches have been proposed to realize efficient MAC computations in a CNN accelerator (Fig. 14.4.6) [1-2, 5-6]. However, finite voltage headroom is required in analog MAC designs, whereas accurate frequency control is necessary for prior frequency-domain MAC approaches. This limits the voltage scalability of analog approaches and performance scalability of prior frequency-domain techniques degrading the MAC

👥 作者与机构

Aseem Sayal, Shirin Fathima, S. S. Teja Nibhanupudi, Jaydeep P. Kulkarni

University of Texas, Austin, TX