Conv-RAM: An Energy-Efficient SRAM with Embedded Convolution Computation for Low-Power CNN-Based Machine Learning Applications

📋 论文概要

提出了一种名为Conv-RAM的能效型SRAM，将卷积计算嵌入到存储阵列中，以降低基于CNN的机器学习应用的功耗。通过电荷分享机制和本地位线电容复制来实现模拟域的计算，并采用输入翻转技术减少敏感放大器偏移的影响。

💡 主要创新点

• 在SRAM存储阵列内直接执行卷积操作，大幅减少数据搬运带来的能耗。
• 利用电荷分享和本地位线电容复制实现高效的低功耗模拟计算。
• 通过周期性地翻转敏感放大器的输入来降低偏移误差，提高计算精度。

🏷 关键词

📄 原文摘要

local column that replicates the local bit-line capacitance. This process continues until the voltage of the rail being integrated exceeds the other one, at which point the SA output flips. This signals conversion completion and no further SA_EN pulses are generated for the SA. Figure 31.1.4 shows the waveforms for a typical operation cycle. To reduce the effect of SA offset on YOUT value, a multiplexer is used at the input of the SA to flip the inputs on alternate cycles. Massachusetts Institute of Technology, Cambridge, MA Convolutional neural networks (CNN) provide state-of-the-art results in a wide variety of machine learning (ML) applications, ranging from image classification to speech recognition. However, they are very computationally intensive and require huge amounts of storage. Recent work strived towards reducing the size of the CNNs: [1] proposes a binary-weight-network (BWN), where the filter

👥 作者与机构

Avishek Biswas, Anantha P. Chandrakasan, charge-sharing is used to integrate the lower of the 2 voltage rails with a reference