MAV-NPU: A 22nm Sparse Neural Network Processor for Real-Time, Resolution-Scalable Vision of Autonomous Micro Aerial Vehicles

Real-time environment perception is essential for enabling autonomous navigation of Micro Aerial Vehicles (MAVs). Yet, power and bandwidth constraints limit efficient deployment of NN-based algorithms. MAV-NPU is a sparse neural network processor in 22 nm optimized for MAV vision. It processes a visual odometry network at 52.4 FPS, consuming 27.4 mW and 2.36 MB/frame. MAV-NPU features three key innovations: 1) a model-level Tiled Dataflow adapting to scaling resolution of vision tasks; 2) grouped hierarchical accumulation reducing write-back stalling; 3) an AvgLeakyReLU activation function maintaining hardware utilization. MAV-NPU enables practical deployment of learning-based MAV vision and achieves peak 28.97 TOPs/W.