A 28nm 12.1TOPS/W Dual-Mode CNN Processor Using Effective-Weight-Based Convolution and Error-Compensation-Based Prediction

📋 论文概要

针对量化CNN推理中的重复权重乘法、ReLU导致的不必要MAC运算以及残差块频繁片外访问三个问题，提出了一种名为QNAP的高能效量化网络加速处理器。该处理器采用有效权重卷积（EWC）和错误补偿预测等技术，实现了12.1TOPS/W的能效。

💡 主要创新点

• 提出有效权重卷积（EWC），通过选举非零有效权重避免重复乘法运算。
• 采用错误补偿预测机制，进一步减少不必要的MAC计算或优化数据流。

🏷 关键词

📄 原文摘要

edge devices efficiently, most existing CNN processors were built on quantized CNNs to optimize the inference operations. However, three issues (Fig. 9.2.1) have not been well addressed: 1) Duplicate weights in each kernel after quantization yielding repetitive multiplications; 2) a huge number of unnecessary MACs caused by ReLU activation functions; 3) frequent off-chip memory access in residual blocks. An energy-efficient quantized network-acceleration processor (QNAP) is proposed to tackle the above problems with three key contributions: 1) Effective-Weight-based Convolution (EWC) elects no more than 6 effective weights (EWs) to stand for all quantized weights in each kernel. The activations with the same EWs can be accumulated

👥 作者与机构

Huiyu Mo1, Wenping Zhu1, Wenjing Hu1, Guangbin Wang1, Qiang Li2, Ang Li1,

Shouyi Yin1, Shaojun Wei1, Leibo Liu1 Institute of Microelectronics of Tsinghua University, Beijing, China Intel, Beijing, China 1 2 To deploy convolutional neural networks (CNNs) on