Highlights of architecture optimizations for ultra-low power inference on larger CNNs