Duplicated with Why inference in jetson nano with fp16 is slower than fp32 - Jetson & Embedded Systems / Jetson Nano - NVIDIA Developer Forums