28nm芯片实现低功耗语义分割，能效提升超3倍。

A 28nm 0.22μJ/token memory-compute-intensity-aware CNN-Transformer accelerator with hybrid-attention-based layer-fusion and cascaded pruning for semantic-segmentation

• 混合注意力+层融合调度，降低计算冗余
• 功耗比前代低3.86至10.91倍，能效达52.90TOPS/W
• 适合边缘设备部署的轻量级视觉模型加速

本文提出一款用于语义分割的28nm 13.93mm² CNN-Transformer加速器，相较以往设计在能耗上降低3.86至10.91倍。其包含混合注意力单元、层融合调度器与级联特征图剪枝模块，峰值能效达52.90TOPS/W（INT8）。该设计通过动态计算资源分配与特征图压缩，显著提升能效比，适用于对功耗敏感的嵌入式视觉应用。

原文摘要 · Abstract (English)

This work presents a 28nm 13.93mm2 CNN-Transformer accelerator for semantic segmentation, achieving 3.86-to-10.91x energy reduction over previous designs. It features a hybrid attention unit, layer-fusion scheduler, and cascaded feature-map pruner, with peak energy efficiency of 52.90TOPS/W (INT8).

Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。