A 1mW Always-on Computer Vision Deep Learning Neural Decision Processor

📋 论文概要

该论文介绍了一款名为NDP200的超低功耗边缘AI处理器，专为计算机视觉应用设计，能够在1mW功率下持续运行。该处理器解决了门铃摄像头等功耗受限设备对低延迟推理的需求。

💡 主要创新点

• 采用Syntiant Core 2架构，专门针对计算机视觉深度学习网络进行优化，实现超低功耗高性能。
• 集成8位直接视频端口（DVP）和I2C接口，适配多种摄像头传感器，支持always-on工作模式。

🏷 关键词

📄 原文摘要

Majid Shaghaghi, Vinay Parthasarathy, Stephen Gibellini, Stephen Bailey, Mallik Moturi, Pieter Vorenkamp, Kurt Busch, Jeremy Holleman, Behrooz Javid, Alireza Yousefi, Mohsen Judy, Atul Gupta Syntiant, Irvine, CA Syntiant®’s NDP200 is a specially architected edge AI processor, bringing the ultra-lowpower, high performance, deep neural network processing capabilities of Syntiant Core 2™ (SC2) to computer vision (CV) applications. Ideally suited for low-latency inferences on power-constrained devices such as doorbell cameras, the NDP200 is equipped with an 8-bit direct video port (DVP), I2C, and SPI interfaces to interface directly with lowpower image sensors, as well as Pulse Density Modulation (PDM) and Inter-IC Sound interface (I2S) to facilitate use cases requiring both video and audio processing. As seen in Fig. 9.2.1, the largest component in the device is the SC2 neural decision processor. The rest of the device is dedicated to communicating and extracting

👥 作者与机构

David Garrett, Youn Sung Park, Seongjong Kim, Jay Sharma, Wenbin Huang,