A 125GOPS 583mW Network-on-Chip Based Parallel Processor with Bio-inspired VisualAttention Engine

📋 论文概要

本文提出了一种基于片上网络的并行处理器，集成了生物启发的视觉注意引擎。通过在目标识别前用视觉注意机制预先选择显著区域，减少了42%的计算成本，在583mW功耗下实现了125GOPS的性能。

💡 主要创新点

• 利用生物启发的视觉注意机制，在预处理阶段过滤非显著区域，大幅降低后续目标识别的计算复杂度。
• 采用片上网络(NoC)架构连接多个并行处理单元，高效支持大量数据交易，实现高吞吐量。
• 在低功耗（583mW）下达到125GOPS的峰值性能，能效比突出。

🏷 关键词

📄 原文摘要

reduce the complexity of object recognition by decreasing the amount of image data to be processed. As Fig. 16.2.1 illustrates, salient parts of a scene are roughly selected by the visual attention mechanism in advance so that next visual processing can focus on only the pre-selected objects to reduce the computational cost by 42%. For such a bio-inspired vision system, multiple parallel processing elements (PEs) and a large amount of data transactions among them are required. Recently, a massively parallel processor was presented for data-level parallelism, but it was not suitable for object-parallel processing due to limited interconnections between PEs [2]. A Network-on-Chip (NoC) is applied to achieve extensive communication bandwidth required for parallel computing. A 125GOPS NoC-based parallel processor

👥 作者与机构

Kwanho Kim, Seungjin Lee, Joo-Young Kim, Minsu Kim,

Donghyun Kim, Jeong-Ho Woo, Hoi-Jun Yoo KAIST, Daejeon, Korea The visual attention algorithm of the human visual system [1] is utilized to