A 28-nm 25.1 TOPS/W Sparsity-Aware CNN-GCN Deep Learning SoC for Mobile Augmented Reality

This work presents the first CNN-GCN SoC for diverse AI vision computations on mobile augmented reality (AR). A CNN engine utilizes the channel-wise feature sparsity with a specialized processing element to achieve an up to 8× higher throughput and 6.1× energy efficiency. A GCN engine is implemented for graph-based action recognition. The computational complexity and memory usage are minimized by lever-aging matrix and graph properties. The proposed SoC achieves 25.1 TOPS/W energy efficiency for CNN inference, outperforming prior designs by 2.0×. It delivers 72 action/s on action recognition, exceeding prior art by 18× in latency.